Revenge of the Liar
This page intentionally left blank
Revenge of the Liar New Essays on the Paradox
edited by
JC...

Author:
JC Beall

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Revenge of the Liar

This page intentionally left blank

Revenge of the Liar New Essays on the Paradox

edited by

JC Beall

1

1

Great Clarendon Street, Oxford ox2 6dp Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With ofﬁces in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York the several contributors 2007 The moral rights of the authors have been asserted Database right Oxford University Press (maker) First published 2007 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose this same condition on any acquirer British Library Cataloguing in Publication Data Data available Library of Congress Cataloging in Publication Data Data available Typeset by Laserwords Private Limited, Chennai, India Printed in Great Britain on acid-free paper by Biddles Ltd, King’s Lynn, Norfolk ISBN 978–0–19–923390–8 ISBN 978–0–19–923391–5 10 9 8 7 6 5 4 3 2 1

CONTENTS

Notes on Contributors 1 Prolegomenon to Future Revenge JC Beall

vii 1

2 Embracing Revenge: On the Indeﬁnite Extendibility of Language Roy T. Cook

31

3 The Liar Paradox, Expressibility, Possible Languages Matti Eklund

53

4 Solving the Paradoxes, Escaping Revenge Hartry Field

78

5 Validity, Paradox, and the Ideal of Deductive Logic Thomas Hofweber

145

6 On the Metatheory of Field’s ‘Solving the Paradoxes, Escaping Revenge’ Hannes Leitgeb

159

7 Reducing Revenge to Discomfort Tim Maudlin

184

8 Understanding the Liar Douglas Patterson

197

9 Revenge, Field, and ZF Graham Priest

225

10 Field on Revenge Agust´ın Rayo and P. D. Welch

234

vi / Contents 11 Bradwardine’s Revenge Stephen Read 12 Curry’s Revenge: The Costs of Non-classical Solutions to the Paradoxes of Self-reference Greg Restall

250

262

13 Aletheic Vengeance Kevin Scharp

272

14 Burali–Forti’s Revenge Stewart Shapiro

320

15 Revenge and Context Keith Simmons

345

Index

369

NOTES ON CONTRIBUTORS

JC Beall, Professor of Philosophy, University of Connecticut, and Arch´e Associate Research Fellow, University of St Andrews. In addition to articles and edited volumes on truth, paradox, and related issues, Beall is the author of Logical Pluralism with Greg Restall and the textbook Possibilities and Paradox: An Introduction to Modal and Many-Valued Logic with Bas C. van Fraassen. He is currently ﬁnishing a monograph on transparent truth and paradox (forthcoming with Oxford University Press). When not doing philosophy, Beall enjoys walking in the woods (and trying not to do philosophy), listening to music, and home schooling his cats. Roy T. Cook, Visiting Assistant Professor, Villanova University, and Arch´e Associate Research Fellow, University of St Andrews. Cook’s main research interests are philosophy of mathematics, philosophical logic, mathematical logic, and the philosophy of language. He has published papers on these topics in Mind, The Monist, Journal of Philosophical Logic, Journal of Symbolic Logic, Notre Dame Journal of Formal Logic, Dialectica, and Analysis, among other places. He is also editor of The Arch´e Papers on the Mathematics of Abstraction. Cook is currently working on a comprehensive dictionary of philosophical logic. When not thinking about logic and mathematics, Cook enjoys building sculptures and mosaics out of LEGO bricks, and attempting to keep his three cats from disassembling them. Matti Eklund, Assistant Professor, Sage School of Philosophy, Cornell University. Research interests: metaphysics, philosophy of language, and philosophy of logic. His publications include ‘Inconsistent Languages’ (Philosophy and Phenomenological Research 2002), ‘What vagueness consists in’ (Philosophical Studies 2005) and ‘Carnap and ontological pluralism’ (forthcoming in D. Chalmers, D. Manley, and R. Wasserman (eds.), Metametaphysics, Oxford University Press). His current research focuses on the implications of the liar and sorites paradoxes, and on various issues in metaontology. Hartry Field, Silver Professor of Philosophy, New York University. Field is the author of Science Without Numbers (Blackwell 1980), which won the Lakatos Prize, of Realism, Mathematics and Modality (Blackwell 1989), and of Truth and the Absence of Fact (Oxford 2001). His current research interests include objectivity and indeterminacy, a priori knowledge, causation, and the semantic and set-theoretic paradoxes; he is currently ﬁnishing a monograph on truth and paradox (forthcoming with Oxford University

viii / Notes on Contributors Press). Field hates writing either bibliographical or biographical blurbs, and so leaves that task to the editor. Thomas Hofweber, Associate Professor of Philosophy, University of North Carolina at Chapel Hill. Hofweber mainly works in metaphysics, the philosophy of language and the philosophy of mathematics. Sample publications: Inexpressible Properties and Propositions (Oxford Studies in Metaphysics, vol. 2, 2006), ‘Number determiners, numbers and arithmetic’ (Philosophical Review 2005), ‘Supervenience and object-dependent properties’ (Journal of Philosophy 2005), ‘A puzzle about ontology’ (Nous 2005). At present he is working on a book on the domain and methods of metaphysics, in particular ontology. Hannes Leitgeb, Reader in Mathematical Logic and Philosophy of Mathematics, University of Bristol. Leitgeb’s chief research interests are in philosophical logic, epistemology, cognitive science, and philosophy of mathematics. He has published, among others, in Journal of Philosophical Logic, Synthese, Analysis, Philosophia Mathematica, Erkenntnis, Journal of Logic, Language and Information, Studia Logica, Notre Dame Journal of Formal Logic, Topoi, Logique et Analyse, Artiﬁcial Intelligence. In 2004 he published Inference on the Low Level: An Investigation into Deduction, Nonmonotonic Reasoning, and the Philosophy of Cognition in the Kluwer/Springer Applied Logic series. At present, Leitgeb is trying to resurrect Carnap’s Logical Structure of the World. Apart from philosophy he likes music and marathons. Tim Maudlin, Professor of Philosophy, Rutgers University. Professor Maudlin’s primary interest is in the nature of reality. Recent books include Truth and Paradox and The Metaphysics Within Physics. He is inordinately fond of Belgian wafﬂes with whipped cream. Douglas Patterson, Associate Professor of Philosophy, Kansas State University. Patterson works in the philosophy of language, philosophical logic, and related areas in metaphysics, epistemology, and the philosophy of mind. Patterson is the author of a number of articles on truth and related topics, including ‘Theories of truth and convention T’, Philosophers’ Imprint 2:5 and ‘Tarski, the liar and inconsistent languages’, The Monist 89:1. He is the editor of a collection of essays on Tarski, forthcoming with Oxford University Press, and of Inquiry 50:6 on inconsistency theories of semantic paradox. In his spare time Patterson enjoys hiking and travel, and he was the 1986 Minnesota Junior Men’s Biathlon champion. Graham Priest, Boyce Gibson Professor of Philosophy, University of Melbourne, and Arch´e Professorial Fellow, University of St Andrews. Priest’s chief research interests are logic and related areas, including metaphysics and the history of philosophy (East and West). He has published in nearly all the major philosophy journals. Recent books include, Towards Non-Being, Doubt Truth to be a Liar, and the second edition of In Contradiction (all with Oxford University Press). He is currently ﬁnishing a second

Notes on Contributors / ix volume of his Introduction to Non-Classical Logic. When not doing philosophy, Priest enjoys doing philosophy. Agust´ın Rayo, Associate Professor of Philosophy, MIT, and Arch´e Associate Research Fellow, University of St Andrews. Rayo’s chief research interests are in philosophical logic and philosophy of language. Rayo has published in such areas in various journals, including Nous and the Journal of Symbolic Logic. He recently edited Absolute Generality with Gabriel Uzquiano. He is currently engaged in a research project on content. When he isn’t doing philosophy, Rayo enjoys scuba diving and opera. Stephen Read, Reader in History and Philosophy of Logic, University of St Andrews. His chief research interests are in philosophical logic, medieval logic, and metaphysics, in particular, the notion of logical consequence; and extend from medieval theories in the philosophy of language, mind, and logic, to the more modern concerns of relevance logic and the philosophy of logic. He has published in such areas in History and Philosophy of Logic, Journal of Philosophical Logic, Mind, Philosophy, Vivarium, and others. His books include Relevant Logic (1987) and Thinking about Logic (1995). He is currently working on a critical edition and English translation of Thomas Bradwardine’s Insolubilia. When not doing philosophy, he enjoys opera, cycling, and playing the clavichord. Greg Restall, Associate Professor of Philosophy, University of Melbourne. Restall’s research interests are in logic, metaphysics, and related ﬁelds. Recent books include Logic (a textbook) and Logical Pluralism, with co-author JC Beall, and he is currently working on the connections between proof theory and meaning. Restall enjoys looking after his young son. Kevin Scharp, Assistant Professor of Philosophy, The Ohio State University. Scharp’s main areas of research are philosophy of language and philosophical logic. He has published papers in British Journal for the History of Philosophy, International Journal of Philosophical Studies, and Inquiry. Scharp is editor (along with Robert Brandom) of the forthcoming volume of Wilfrid Sellars’ essays entitled In the Space of Reasons from Harvard University Press. Scharp is currently working on a book on truth and the liar paradox. When not doing philosophy, he loves backpacking, cooking, and listening to music. Stewart Shapiro, O’Donnell Professor of Philosophy, The Ohio State University, and Arch´e Professorial Fellow, University of St Andrews. He specializes in philosophy of mathematics, philosophical logic, and philosophy of language, notably vagueness. His most recent book is Vagueness in Context (Oxford: Oxford University Press, 2007) and he is the editor of the Oxford Handbook of the Philosophy of Mathematics and Logic. He still jogs (if you can still call it that), and likes the Grateful Dead and Incredible String Band. Keith Simmons, Professor of Philosophy, University of North Carolina at Chapel Hill. He has research interests in logic, philosophy of logic, and philosophy of language. He

x / Notes on Contributors is the author of Universality and the Liar (Cambridge University Press) and, with Simon Blackburn, the editor of Truth (in the Oxford Readings in Philosophy series). He is currently at work on a monograph about the paradoxes, and, with Dorit Bar-On, a monograph about truth. P. D. Welch, Professor of Pure Mathematics, University of Bristol. He has research interests in set theory, models of computation, theories of truth, and possible worlds semantics. His work on these and related topics have appeared in many of the leading logic and mathematics journals. Welch is currently working with the The Luxemburger Zirkel on a major research project entitled Logical Methods in Epistemology, Semantics, and Philosophy of Mathematics.

1 Prolegomenon to Future Revenge JC Beall

This chapter attempts to lay out some background to the target phenomenon: the Liar and its revenge. The phenomenon is too big, and the literature (much) too vast, to give anything like a historical summary, or even an uncontroversial sketch of the geography. Accordingly, my aim is simply to lay out a few background ideas, in addition to brieﬂy summarizing the contributed essays. I also try to avoid overlap with the individual chapters’ rehearsals of revenge (including standard references to historical theses, like ‘semantic self-sufﬁciency’), as the chapters do a nice job covering such material. Finally, because some of the ideas are presupposed by many of the chapters in this volume, a gentle sketch of Kripke’s ‘ﬁxed point’ approach to truth is given in an appendix.

1.1 Truth Whatever else it may do, truth is often thought to play Capture and Release. Where Tr(x) is our truth predicate, α a sentence, and α a name of α, Capture and Release are as follows. Capture: α ⇒ Tr(α) Release: Tr(α) ⇒ α When ⇒ is a conditional, we have the Conditional Form of Capture and Release: namely, when conjoined, the familiar T-biconditionals. When ⇒ is a turnstile, we have the Rule Form of Capture and Release, which indicates ‘valid inference’ (in some sense).

2 / JC Beall The names ‘Capture’ and ‘Release’ arise from the fact that Tr(x) captures the information in x, fully storing it for its eventual release. In practice, a familiar—if not the—role of Tr(x) is its release function: an assertion of Tr(α) releases all of the information in α. This is useful for ‘long generalizations’ or ‘blind generalizations’ or the like, many of which generalizations would be practically impossible if we didn’t enjoy a truth predicate that played (at least the rule form of) Capture and Release.¹ That truth plays Capture and Release in Conditional Form is plausible but controversial. That truth plays Capture and Release in at least Rule Form is less controversial, and will henceforth be assumed.²

1.2 The Liar The Liar phenomenon involves sentences that imply their own falsity or, more generally, untruth. By way of example, consider the ticked sentence in §1.2 of this chapter. √ The ticked sentence in §1.2 of ‘Prolegomenon to future revenge’ is not true. Assume that the ticked sentence is true. Release, in turn, delivers that the ticked sentence is not true. Hence, the ticked sentence, given Release (an essential feature of truth), implies its own untruth. Is the ticked sentence untrue? Capture gives reason for pause: that the ticked sentence is not true implies, via Capture, that the ticked sentence is true! The question is: what shall we say about the ‘semantic status’ of the ticked sentence? Answering this question invites the Liar’s revenge.

1.3 The Liar’s Revenge On one hand, the revenge phenomenon—the Liar’s revenge—is not so much a distinct phenomenon from the Liar as it is a witness to both the difﬁculty and ubiquity of Liars. ¹ Depending on the language, Rule Capture and Release is insufﬁcient for a fully transparent truth predicate, one such that Tr(x) and x are intersubstitutable in all (non-opaque) contexts, for all sentences x in the language. By my lights, the Liar phenomenon is at its most difﬁcult incarnation when truth is fully transparent, since any distinction between, for example, ‘Excluded Middle’ and ‘Bivalence’ collapses. But I will set this aside here. See Appendix for one approach to transparent truth, and see Field’s chapter (Chapter 4) for another, stronger approach, as well as relevant discussion. ² A variety of theories reject even Rule Form of Capture and Release, but it will be assumed throughout this ‘introduction’. One of the better known examples of rejecting Rule Capture is so-called Kripke–Feferman [3]. (See too [16].)

Prolegomenon to Future Revenge / 3 On the other hand, ‘revenge’ is often launched as an objection to an account of truth (or a response to Liars). Without intending a stark distinction, I will discuss revenge qua Liar phenomenon and revenge qua objection separately, with most of the discussion on the latter but all of the discussion brief.

1.3.1 The revenge phenomenon The revenge phenomenon arises at the point of classifying Liars. Consider, again, the ticked sentence. As above, classifying the ticked sentence as true results in inconsistency; Release delivers that the ticked sentence is also not true. Likewise, classifying the ticked sentence as not true results in inconsistency; Capture delivers that the ticked sentence is also true. How, then, shall we classify the ticked sentence? A natural suggestion is that the ticked sentence is neither true nor false. The trouble with this suggestion—even apart from logical issues involving negation—is the apparent connection between being neither true nor false and being not true.³ In particular, presumably, we have ntf–nt. ¬Tr(α) ∧ ¬Tr(¬α) ⇒ ¬Tr(α) Again, ⇒ may be a conditional or a turnstile. Either way, the problem at hand is plain. Assume, as per the current suggestion, that the ticked sentence is ‘neither true nor false’. By ntf–nt, we immediately get that the ticked sentence is not true. But, now, we’re back to inconsistency, as Capture, in turn, delivers that the ticked sentence is (also) true. So, while natural, the suggestion that the ticked sentence is neither true nor false is not a promising proposal.⁴ Quick reﬂection leads to a general lesson: whatever category one devises for the ticked sentence, it had better not imply untruth. For example, suppose that one introduces the category bugger for Liars. On this proposal, the ticked sentence is a bugger. Whatever else being a bugger might involve, we had better not have bug if we’re to avoid inconsistency. bug. Bugger(α) ⇒ ¬Tr(α) The trouble with bug is exactly the trouble with ntf–nt. On the current proposal, the ticked sentence is a bugger, in which case, via bug, it is not true. Capture, as before, delivers that the ticked sentence is (also) true, and inconsistency remains. ³ Throughout, I will assume that falsity is truth of negation—i.e., that α is false just if ¬α is true. (This is a standard line, but it might be challenged. Fortunately, in the present context, nothing substantive turns on the issue.) ⁴ I should note that my presentation simpliﬁes matters a great deal. One might postulate a different negation at work in (wide-scope positions in) ntf–nt, thereby complicating matters. Moreover, one might—perhaps with some philosophical motivation—reject ntf–nt altogether. And there are other options, as will be evident in various chapters of this volume.

4 / JC Beall One lesson, then, is that our classiﬁcation of the ticked sentence cannot consistently deliver its untruth. With the lesson in mind, suppose that we classify the ticked sentence as a bugger but, whatever else ‘bugger’ might mean, we reject bug (in both Rule and Conditional Forms). Notwithstanding further details on ‘buggerhood’, this course yields the promise of consistently classifying the ticked sentence (and its negation): it is a bugger. The revenge phenomenon re-emerges. Having, as we’re assuming, consistently classiﬁed the ticked sentence as a ‘bugger’, other Liars emerge to thwart our aims at consistently (and completely) classifying Liars. By way of example, consider the starred sentence. The starred sentence in §1.3.1 of ‘Prolegomenon to future revenge’ is either not true or a bugger. Assume that the starred sentence is true. Release delivers that the starred sentence is not true or a bugger, and hence true and either not true or a bugger. Similarly, that the starred sentence is either not true or a bugger implies, via Capture, that it is true—and, hence, true and either not true or a bugger. Accordingly, given normal conjunction and disjunction behavior, if we have it that the starred sentence is either true, not true, or a bugger, we have either inconsistency (viz., true and not true) or some true buggers. While the latter option, without bug (or similar principles), might afford a consistent theory, it is prima facie objectionable if, whatever else ‘bugger’ might mean, the buggers are thought to be somehow ‘defective’, sentences that ought to be rejected.⁵ The revenge phenomenon, at least in one relevant sense, is as above: it is not so much a separate phenomenon from the Liar as it is what makes the Liar phenomenon challenging. The Liar’s revenge is reﬂected in the apparent hyrdra-like appearance of Liars: once you’ve dealt with one Liar, another one emerges. In short, if one manages to consistently classify a Liar as a such-n-so, another Liar emerges—e.g. a sentence that says of itself only that it’s not true or a such-n-so. Dramatically and very generally put, Liars attempt to wreak inconsistency in one’s language. If the Liar can’t have what she wants, she’ll enlist ‘strengthened’ relatives to frustrate your wants, in particular, your expressive wants. As it is sometimes put, Liars force—or try to force—you to choose between either inconsistently expressing what you want to express or not expressing what you want to express. A quietist advises that we give up on our aim to classify Liars; there are Liars in the language, but there is no ‘semantic category’ in which the ticked sentence may truly be said to reside. Accordingly, whereof one cannot truly classify, thereof one must—or, in any event, might as well—be silent. The virtue of such an approach ⁵ Whether ‘buggers’ should be conceived as defective (in some sense) is an open issue. Field’s chapter (see Chapter 4) is relevant to this issue. See too [1].

Prolegomenon to Future Revenge / 5 is that it avoids revenge (since it doesn’t engage); the salient defect is that it offers no clear account of truth or the paradoxes at all. Against such a ‘proposal’, little can be said, and so won’t. Another—so-called dialetheic—option is to accept the apparent inconsistency engendered by Liars. Provided that our logic tolerates such inconsistency—and part of theproposed lessonoftheLiaristhatourlogicdoestoleratesuchinconsistency—there’s no obvious problem. What Liars teach us, on the dialetheic view, is that truth is inconsistent—that some true sentences have true negations. Whether such a position avoids the Liar’s revenge is an open question.⁶ And there are (many) other options, as subsequent chapters reﬂect. What is uncontroversial is that the revenge phenomenon has fueled, and continues to fuel, work on the Liar phenomenon. This is not surprising, at least if, as suggested, the revenge phenomenon just is the Liar phenomenon—indeed, as above, a witness to the Liar’s ubiquity.

1.3.2 Towards revenge qua objection The literature on truth and paradox exhibits a familiar and ubiquitous pattern: each proposed ‘account of truth’ is followed by a charge of revenge, that the account can’t accommodate such and so a notion (e.g. ‘untruth’, ‘exclusively false’, or whathaveyou) and, in that respect, is thereby inadequate. Indeed, were it not for alleged ‘revenge’ problems, many proposed theories of truth might be objection-free—or, at least, the number of known or cited objections would be greatly diminished. Such ‘revenge’ charges, as said, are often launched as inadequacy objections against proposed accounts of truth. Unfortunately, there is some unclarity about the relevance of such charges, and, more to the point, unclarity with respect to the burden involved in successfully establishing the intended inadequacy result. Without aiming to resolve them, §1.4 brieﬂy discusses some of the given issues involved in revenge qua objection. Before turning to §1.4, two background issues need to be brieﬂy covered.⁷ 1.3.2.1 Incoherent operators By way of background, it is important to see that there are operators that cannot coherently exist if our language enjoys various features. Tarski’s Theorem gives one ⁶ That some sentences are true and false is one thing; however, the dialetheic position is rational only if at least some sentences are just true. The worry is whether the dialetheist can give an adequate account of ‘just true’ without the position exploding into triviality. Some of the chapters have discussion of this point. For a general discussion (and defense) of dialetheism, see [14, 15]. ⁷ I should warn that, from this point forward, my presentation may border on controversial.

6 / JC Beall concrete example of such a result,⁸ but another example might be useful. In particular, suppose, as is plausible, that our language has features F1 and F2. F1. There’s a predicate Tr(x) that ‘obeys’ (unrestricted) Release and Capture in at least Rule Form. F2. ‘Reasoning by Cases’ is valid: if α implies γ , and β implies γ then α ∨ β implies γ , for all α, β, γ . As such, the language, on pain of triviality, has no operator such that both E1 and E2 hold.⁹ E1. α ∨ α E2. α, α ⊥ Suppose that we do have such an operator. Consider a familiar construction, which will be guaranteed via diagonalization, self-reference or the like: a sentence λ that ‘says’ Tr(λ). From E1, we have Tr(λ) ∨ Tr(λ) which yields two cases. 1. Case one: (a) Tr(λ) (b) Release yields:¹⁰ Tr(λ). (c) E2 yields: ⊥ 2. Case two: (a) Tr(λ) (b) Capture yields: Tr(λ) (c) E2 yields: ⊥ The point, for present purposes, is modest but important: there are incoherent notions, notions that cannot coherently exist if our language enjoys various features. While modest, the point is something on which all parties can agree. ⁸ Tarski’s Theorem, in effect, is that (classical) arithmetical truth is not deﬁnable in (classical) arithmetic. For a user-friendly discussion of the theorem and its broader implications, see [20] and, more in-depth, [18]. For a user-friendly discussion of what Tarski’s Theorem does not teach us, see [23], which is also highly relevant to ‘revenge’ issues, in general, and particularly relevant to Field’s proposal (see Chapter 4). ⁹ E1 might be thought of as an exhaustion principle, and E2 as exclusion or explosion principle. Throughout, ⊥ is an ‘explosive’ sentence, one that implies all sentences. ¹⁰ Intersubstitutability of Identicals is also involved here (and at the same place in Case two). This is usually assumed to be valid, but it, like so much in the area, has been challenged. See [17].

Prolegomenon to Future Revenge / 7 A principal question, at the heart of Liar studies, is this: what is our language like, given that it enjoys such and so features? More to the point: assuming that our language has a truth predicate that plays Capture and Release (in at least rule form), what are its other features? One might say that it fails to contain a fully exhaustive device, something that would yield E1, or fails to have any fully explosive device, something that would yield E2. One might, with various theorists, say that F2, in its given unrestricted form, fails for our language. One might say other things. Whatever one says, one aims to give a clear, precise account of the matter—a clear, precise account of what our language is like, given that it has such and so features. This is normally done by way of a ‘formal modeling’. 1.3.2.2 Models and reality Like much in philosophical logic, constructing a formal account of truth is ‘model building’ in the ordinary ‘paradigm’ sense of ‘model’. The point of such a model is to indicate how ‘real truth’ in our ‘real language’ can have the target (logical) features we take it to have—e.g. consistency (or, perhaps, inconsistency but non-triviality), Release and Capture features, perhaps full intersubstitutability of Tr(α) and α. In that respect, formal accounts of truth are idealized models to be evaluated by their adequacy with respect to the ‘real phenomena’ they purport to model.¹¹ Formal accounts (or theories) of truth aim only indirectly at being accounts of truth. What we’re doing in giving such an account is two-fold. 1. We construct an artiﬁcial model language—one that’s intended to serve as a heuristic, albeit idealized, model of our own ‘real’ language—and, in turn, give an account of how ‘true’ behaves in that language by constructing a precise account of truth-in-that-language. 2. We then claim that the behavior of ‘true’ in our language, at least in relevant, target respects, is like the behavior of the truth predicate in our model language. By far the most dominant approach towards the ﬁrst task—viz. constructing one’s model language—employs a classical set theory. One reason for doing so is that classical set theory is familiar, well-understood, and generally taken to be consistent. A related reason is that, in using a classical set theory, one’s formal account of truth ¹¹ Theories, like McGee’s [13], that purport not to be ‘descriptive’ but, rather, ‘revisionary’ or ‘normative’, are not typically subject to ‘revenge’-charges to the same extent that ‘descriptive’ theories are, and so are not the chief concern here. On the other hand, McGee aims to give a revisionary theory (not to be confused with revision theory) that aims to stay as close to the phenomena—our ‘real language’—as possible. In that respect, ‘revenge’ objections might well arise.

8 / JC Beall can be more than merely a heuristic picture; it can also serve as a ‘model’ in the technical sense of establishing consistency.¹² That a classical set theory is used in constructing our artiﬁcial language serves to emphasize the heuristic, idealized nature of the construction. We know that, due to paradoxical sentences, there’s no truth predicate in (and for) our ‘real language’ if our real language is (fully) classical.¹³ But the project, as above, is to show how we can have a truth predicate in our ‘real language’, despite such paradoxical sentences. And the project, as above, is usually—if not always—carried out in a classical set theory. Does this mean that the project, as typically carried out, is inexorably doomed? Not at all. Just as in physics, where idealization is highly illuminating despite its distance from the real mess, so too in philosophical logic: the classical construction is illuminating and useful, despite its notable idealization. But it is idealized, and, pending argument, on the surface only heuristic. That’s the upshot of using classical set theory.

1.4 Comments on Revengers’ Revenge A quick glance at the Liar literature will indicate that ‘revenge’ is often invoked as a problem for a given theory of truth and paradox. For present purposes, a revenger is one who charges ‘revenge’ against some proposed account of truth. The principal issue of this section is the burden of revenge—the burden that revengers carry. The chapters in this volume will tell their own (and not necessarily compatible) story on this issue.

1.4.1 Too easy revenge As above, in giving a formal theory of truth, one does not directly give a theory of truth; rather, one gives a theory of Lm -truth, an account, for some formal ‘model language’ Lm , of how Lm ’s truth predicate behaves, in particular, its logical behavior. By endorsing a formal theory of truth, one is endorsing that one’s own truth predicate is relevantly like that, like the truth predicate in Lm , at least with respect to various phenomena in question—for example, logical behavior. ¹² In paraconsistent contexts, the aim is basically the same, except that the target result is non-triviality despite negation-inconsistency. In the more dominant non-paraconsistent cases, the aim is also non-triviality, but that’s ensured by consistency. ¹³ The same applies, of course, if the truth predicate has an extension: the extension isn’t really a classical set. Every classical set S is such that x ∈ S ∨ x ∈ / S, which, given paradoxical sentences, results in inconsistency. (The point is independent of ‘size’ issues. Classical proper classes are likewise such that x ∈ C ∨ x ∈ / C.) If T is the extension of Tr(x) and T is a set, a sentence λ that ‘says’ λ ∈ /T makes the point—assuming, as is plausible, suitable ‘extension’ versions of Capture and Release (e.g., α ⇒ α ∈ T, etc.).

Prolegomenon to Future Revenge / 9 Revenge qua objection—revenger’s revenge—is an adequacy objection. Typically, the revenger charges that a given ‘model language’ is inadequate due to expressive limitation. Let L be our ‘real language’, English or some such natural language, and let Lm be our heuristic model language. Let ‘Lm -truth’ abbreviate ‘the behavior of Lm ’s truth predicate’. In broadest terms, the situation is this: we want our (heuristic) Lm , and in particular Lm -truth, to illuminate relevant features of our own truth predicate, to explain how, despite paradoxical sentences, our truth predicate achieves the features we take it to have. Revenge purports to show that Lm achieves its target features in virtue of lacking expressive features that L itself (our real language) appears to enjoy. But if Lm enjoys the target features only in virtue of lacking relevant features that our real L enjoys, then Lm is an inadequate model: it fails to show how L itself achieves its target features (e.g. consistency). That, in a nutshell, is one common shape of revenge. Consider a familiar and typical example, namely, Kripke’s partial languages.¹⁴ Let Lm , our heuristic model language, be such a (ﬁxed point) language constructed via the Strong Kleenescheme.¹⁵ Inconstructing Lm ,weuse—inourmetalanguage—classical set theory, and we deﬁne truth-in-Lm (and similarly, false-in-Lm ), which notions are used to discuss Lm -truth (the behavior of Lm ’s truth predicate). Moreover, we can prove—in our metalanguage—that, despite paradoxical sentences, a sentence Tr(α) is true-in-Lm exactly if α is true-in-Lm . The familiar revenge charge is that Lm , so understood, is not an adequate model; it fails to illuminate how our own truth predicate, despite paradoxical sentences, achieves consistency. In particular, the revenger’s charge is that Lm -truth achieves its consistency in virtue of Lm ’s expressive poverty: Lm cannot, on pain of inconsistency, express certain notions that our real language can express. Example: suppose that Lm contains a predicate ϕ(x) that deﬁnes {β : β is not true-in-Lm }. And now, where λ says ϕ(λ), we can immediately prove—in the metalanguage—that λ is true-in-Lm iff ϕ(λ) is true-in-Lm iff λ is not true-in-Lm . Because—and only because—we have it (in our classical metalanguage) that λ is true-in-Lm or not, we thereby have a contradiction: that λ is both true-in-Lm and not. But since we have it that truth-in-Lm is consistent (given consistency of classical set theory in which Lm is constructed), we conclude that Lm cannot express ‘is not true-in-Lm ’. The revenger’s charge, then, amounts to this: that the Kripkean model language fails to be enough like our real language to explain at least one of the target phenomena, namely, truth’s consistency. Our metalanguage is part of our ‘real language’, and we can deﬁne {β : β is not true-in-Lm } in our metalanguage. As the Kripkean language ¹⁴ See Appendix for a sketch of the Kripkean ‘partial predicates’ approach. ¹⁵ The point applies to any of the given languages, but the K3 -construction (Strong Kleene) is probably most familiar.

10 / JC Beall cannot similarly deﬁne {β : β is not true-in-Lm }, the Kripkean model language is inadequate: it fails to illuminate truth’s target features. A revenger engages in ‘too easy revenge’ if the revenger only points to such a result without establishing its relevance.¹⁶ The relevance of such a result is not obvious. After all, the given notion is a classically constructed notion; it is a ‘model-dependent’ notion—a notion that makes no sense apart from the given (classically constructed) models—deﬁned entirely in a classical metalanguage. As such, the given notion, presumably, is not one of the target (model-independent, or ‘absolute’) notions in L that Lm is intended to model. The question, then, isn’t whether there’s some notion X (e.g., ‘not true-in-Lm ’) that is inexpressible—or, at least, not consistently expressible—in Lm . The question is the relevance of such a result. One might think that the relevance is plain. One might, for example, think that the semantics for Lm is intended to reﬂect the semantics of L, our real language. Since the semantics of the former essentially involves, for example, not true-in-Lm , the semantics of our real language must involve something similar—at least if Lm is an adequate model of our real language. But, now, since not true-in-Lm is (provably) inexpressible in Lm , we should conclude that Lm is an inadequate model of our real language L, since our real language can express its own semantic notions—i.e. the notions required for giving the semantics of our language. Such an argument might serve to turn otherwise ‘too easy revenge’ into a plainly relevant and powerful objection; however, the argument itself relies on various assumptions that involve quite complex issues. For example, one conspicuous assumption is that the ‘semantics’ of Lm is intended to reﬂect the semantics of our real language L. This needn’t be the case. For example, suppose that one rejects that semantics—the semantics of our real language—is a matter of giving ‘truth conditions’ or otherwise involves some explanatory notion of truth. In the face of Liars (or other paradoxes), one still faces questions about one’s truth predicate, and in particular its logical behavior. By way of answering such questions, one might proceed as above: construct a model language that purports to illuminate how one’s real truth predicate enjoys its relevant features (e.g. Capture and Release) without collapsing from paradox. In constructing and, in turn, describing one’s ‘model language’, one might give ‘truth-conditional-like semantics’ for the model language by giving ‘truth-in-a-model conditions’ for the language. If so, it is plain that the ‘semantics’ of the model language are not intended to reﬂect the ‘real semantics’ of one’s real language; they may, in the end, be only tools used for illuminating the logic of our real language, versus illuminating the ‘real semantics’ of our real language. So, a critical assumption in the argument above—the argument towards the relevance of the given inexpressibility results—requires argument. Likewise, the assumption that our ¹⁶ Thanks to Lionel Shapiro for very useful discussion on ‘too easy revenge’.

Prolegomenon to Future Revenge / 11 real language can ‘express its own semantic notions’, the notions involved in ‘giving the semantics’ of our language, requires argument, argument that may turn, as with the ﬁrst assumption, on difﬁcult issues concerning the very ‘nature of semantics’.¹⁷ A would-be revenger, involved in too easy revenge, would have it easy but too easy. What is (generally) easy is showing that some classically constructed notion is inexpressible—or, at least, not consistently expressible—in a (classically constructed) non-classical ‘model language’. What is too easy is the thought that showing as much is sufﬁcient to undermine the adequacy of the given model language. The hard part is clearly establishing the relevance of such inexpressibility results, that is, clearly substantiating the alleged inadequacy. The difﬁculty, as above, is that the alleged inadequacy often relies on very complicated issues—the ‘nature of semantics’, the role of given model-dependent notions, and more.

1.4.2 Revenger’s recipes, in general Towards clarifying the burden involved in launching revenger’s revenge, it might be useful to lay out a few common recipes for revenge qua objection. For simplicity, let Lm be a given formal model language for L, where L is our target, real language—the language features of which Lm is intended to illuminate. Let M(Lm ) be the metalanguage for Lm , and assume, as is typical, that M(Lm ) is a fragment of L. Then various (related) recipes for revenge run roughly as follows.¹⁸ Rv1. Recipe One. • Find some semantic notion X that is used in M(Lm ) to classify various Lm sentences (usually, paradoxical sentences). • Show, in M(Lm ), that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). Rv2. Recipe Two. • Find some semantic notion X that, irrespective of whether it is explicitly used to classify Lm -sentences, is expressible in M(Lm ). • Show, in M(Lm ), that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). ¹⁷ Some of the chapters in this volume discuss this assumption, an assumption that often goes under the heading ‘semantic self-sufﬁciency’. For arguments against such an assumption, see [7, 8]. ¹⁸ This is not in any way an exhaustive list of recipes!

12 / JC Beall Rv3. Recipe Three. • Find some semantic notion X that is (allegedly) in L. (Argue that X is in L.) • Argue that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). As above, a revenger is one who charges ‘revenge’ against a formal theory of truth, usually along one of the recipes above. The charge is that the model language fails to achieve its explanatory goals. In general, the revenger aims to show that there’s some sentence in our real language L that ought to be expressible in Lm if Lm is to achieve explanatory adequacy. The question is: how ought one reply to revengers? The answer, of course, depends on the details of the given theories and the given charge of revenge. For present purposes, without going into such details, a few general remarks can be made. The weight of Rv1 or Rv2 depends on the sort of X at issue. As in §1.3.2.2 and §1.4.1, if X is a classical, model-dependent notion constructed in a proper fragment of L, then the charge of inadequacy is not easy to substantiate, even if the inexpressibility of X in Lm is easy to substantiate. In particular, if classical logic extends that of Lm , then there is a clear sense in which you may ‘properly’ rely on a classical metalanguage in constructing Lm and, in particular, truth-in-Lm . In familiar non-classical proposals, for example, you endorse that L, the real, target language, is non-classical but enjoys classical logic as a (proper) extension, in which case, notwithstanding particular details, there is nothing prima-facie suspect about relying on an entirely classical fragment of L to construct your model language and, in particular, classical model-dependent Xs. But, then, in such a context, it is hardly surprising that X, being an entirely classical notion, would bring about inconsistency or, worse, triviality, in the (classically constructed) non-classical Lm .¹⁹ Because classical logic is typically an extension of the logic of Lm , the point above is often sufﬁcient to blunt, if not undermine, a revenger’s charge, at least if the given recipe is Rv1 and Rv2. As in §1.4.1, the revenger must establish more than the unsurprising result that a (usually classically constructed) model-dependent X is expressible in M(Lm ) but not in Lm ; she must show the relevance of such a result, which might well involve showing that some non-model-dependent notion—some relevant ‘absolute’ notion—is expressible in L but, on pain of inconsistency (or non-triviality), inexpressible in Lm . And this task brings us to Rv3. Recipe Rv3 is perhaps what most revengers are following. In this case, the idea is to locate a relevant non-model-dependent notion in L and show that Lm cannot, on pain of inconsistency (or triviality), express such a notion. The dialectic along these lines is delicate. ¹⁹ For closely related discussion, see Field’s chapter (Chapter 4) and also [4].

Prolegomenon to Future Revenge / 13 Suppose that Theorist proposes some formal theory of truth, and Revenger, following Rv3, adverts to some ‘absolute’ notion X that, allegedly, is expressible in L. If, as I’m now assuming, Theorist neither explicitly nor implicitly invokes X for purposes of classifying sentences, then Revenger has a formidable task in front of her. In particular, without begging questions, Revenger must show that X really is an intelligible notion of L. For example, recall, from §1.3.2.1, the discussion of ‘incoherent operators’, and assume that Theorist proposes a theory that has features F1 and F2 (and, for simplicity, is otherwise normal with respect to extensional connectives). Let any operator that satisﬁes E1 and E2 be an EE device (for ‘exclusive and exhaustive’). Against a typical paracomplete (or paraconsistent) proposal,²⁰ an Rv3-type revenger might maintain that L, our real language, enjoys an EE device. If the revenger is correct, then standard paracomplete and paraconsistent proposals are inadequate, to say the least. But the issue is: why think that the revenger is correct? Argument is required, but the situation is delicate. What makes the matter delicate is that many arguments are likely to beg the question at hand. After all, according to (for example) paracomplete and paraconsistent theorists, what the Liar teaches us is that, in short, there is no EE device in our language! Accordingly, the given revenger cannot simply point to normal evidence for such a device and take that to be sufﬁcient, since such ‘evidence’ itself might beg questions against such proposals. On the other hand, if the given theorist cannot otherwise explain—or, perhaps, explain away—normal evidence for the (alleged) device, then the revenger may make progress. But the situation, as said, is delicate. The difﬁculty in successfully launching Rv3 might be put, in short, as follows. Theorist advances Lm as a model of (relevant features of) L, our real language. Rv3 Revenger alleges that X exists in L, and shows that, on pain of triviality, X is inexpressible in Lm . The difﬁculty in adjudicating the matter is that, as in §1.3.2.1, Theorist may reasonably conclude that X is incoherent (given the features of our language that Theorist advances). Of course, if Revenger could establish that we need to recognize X, perhaps for some theoretical work or otherwise, then the debate might be settled; however, such arguments are not easy to come by. The burden, of course, lies not only on the Rv3 Revenger; it also lies with the given theorist. For example, typical paracomplete and paraconsistent theorists must reject the intelligibility of any EE device in our language. Inasmuch as such a notion is independently plausible—or, at least, independently intelligible—such theorists carry the burden of explaining why such a notion appears to be intelligible, despite its ultimate unintelligibility. Along these lines, the theorist might argue that we are making a common, reasonable, but ultimately fallacious generalization from ‘normal ²⁰ A paracomplete proposal rejects LEM, and a paraconsistent proposal rejects ‘Explosion’ (i.e., α, ¬α ⇒ β, in both Rule and Conditional form). (See Appendix for the former type of approach.)

14 / JC Beall cases’ to all cases, or some such mistake. (E.g. some connective, if restricted to a proper fragment of our language, behaves in the EE way.) Alternatively, such theorists might argue that, contrary to initial appearances, the allegedly intelligible notion only appears to be a clear notion but, in fact, is rather unclear; once clariﬁed, the alleged EE device (or whatever) is clearly not such a device. (E.g. one might argue that the alleged notion is a conﬂation of various notions, each one of which is intelligible but not one of which behaves in the alleged, problematic way.) Whatever the response, theorists do owe something to Rv3 revengers: an explanation as to why the given (and otherwise problematic) notion is unintelligible.

1.5 Some Closing Remarks I have hardly scratched the surface of revenge in the foregoing remarks. The phenomenon (or, perhaps more accurately, family of phenomena) has in many respects been the fuel behind formal theories of truth, at least in the contemporary period. Despite such a role, a clear understanding of revenge is a pressing and open matter. What, exactly, is revenge? How, if at all, is it a serious problem? Is the problem logical? Is the problem philosophical? And relative to what end, exactly, is the alleged problem a problem? Answers to some of the given questions, I hope, are clear enough in foregoing remarks, but answers—clear answers—to many of the questions remain to be found. Until then, full evaluation of current theories of truth remains out of reach. The hope, however, is that the papers in this volume move matters forward.

Chapter Summaries What follows are brief synopses of the chapters, ordered alphabetically in terms of author(s). The synopses are intended to help the reader ﬁnd chapters of particular interest, rather than serve as discussion of the chapters. Cook. Call a concept C indeﬁnitely extensible just if there’s a rule r such that, when applied to any ‘deﬁnite collection’ of objects falling under C, r yields a new object falling under C. In his ‘Embracing revenge: on the indeﬁnite extendibility of language’, Roy Cook argues that the revenge phenomenon is reason to think that our concept of language, and the associated concept of truth value (or semantic value), is indeﬁnitely extensible. In the end, the revenge phenomenon is a witness to the indeﬁnite extensibility of our language, and, in particular, its ‘semantic values’.

Prolegomenon to Future Revenge / 15 Eklund. In his ‘The liar paradox, expressibility, possible languages’, Matti Eklund focuses on general theses that are standardly tied to the Liar phenomenon. On one hand, there are two related lessons that are sometimes drawn from the Liar’s revenge: namely, radical inexpressibility and (the weaker) inexpressibility. On the other hand, there are two related principles that often make for frustration in the face of revenge: namely, semantic self-sufﬁciency and (what Eklund calls) weak universality. Eklund elucidates the given theses, and focuses attention on inexpressibility and weak universality. Eklund argues that common approaches to such theses may confront difﬁculties from facts governing the space of possible languages, an issue at the heart of Eklund’s essay. Field. In his ‘Solving the paradoxes, escaping revenge’, Hartry Field advances a (paracomplete) theory of truth that, he argues, undermines the ‘received wisdom’ about revenge, where such wisdom, as Field puts it, maintains that ‘any intuitively natural and consistent resolution of a class of semantic paradoxes immediately leads to other paradoxes just as bad as the ﬁrst’. After presenting his own theory of truth (which extends the Kripke approach with a suitable conditional), Field argues that, pace ‘received wisdom’, it is revenge-free. The overall theory and arguments for its revenge-free status have provoked discussion in other chapters (see especially Leitgeb, Priest, Rayo–Welch).²¹ Hofweber. Validity is often thought to be truth-preserving: an inference rule is valid just if truth-preserving.²² Thomas Hofweber, in his ‘Validity, paradox, and the ideal of deductive logic’, argues that two senses of ‘an inference rule is valid just if truthpreserving’ are important to distinguish. One sense is the ‘strict reading’, according to which each and every instance of the given rule is truth-preserving. The other reading is the ‘generic reading’, which, in some sense, is analogous to the claim that bears are dangerous, a claim that is true even though not true of all bears. This distinction, which Hofweber discusses, holds the key to resolving the revenge phenomenon. In ²¹ One issue not discussed is the ideal of ‘exhaustive characterization’, according to which we can truly say (something equivalent to) that all sentences are either True, False, or Whathaveyou (where ‘Whathaveyou’ is a stand in for the predicates used to classify Liars or the like), and do as much in our own language. One might wonder whether the ‘received wisdom’ counts as ‘natural’ only those theories that afford exhaustive characterization, in which case, Field’s argument against ‘received wisdom’ might miss the mark. (Without further clariﬁcation of ‘exhaustive characterization’, I do not intend these remarks as a serious objection, but rather only something for the reader to consider.) ²² I should note that if, as is usual, ‘truth-preserving’ is understood via a conditional, so that α, β is ‘truth-preserving’ just if α → β is true (for some suitable conditional in the language), then many standard theories of transparent truth (i.e., fully intersubstitutable truth) will not have it that valid arguments are truth-preserving. See [2] for some discussion, but also [5] for broader, philosophical issues. This issue, regrettably, is not discussed much in this volume, but it is highly important. Restall’s chapter (Chapter 12) has some direct relevance for the issue, as does Field’s (Chapter 4). Hofweber brieﬂy mentions the issue as it arises for Field’s theory.

16 / JC Beall particular, the Liar’s revenge teaches us that we should abandon the traditional ideal of deductive logic, which requires that our theories be underwritten by rules that are valid in the ‘strict sense’. On the positive side, the Liar’s revenge teaches us that we should embrace the ‘generic’ ideal of deductive logic, which requires only that our rules be ‘generically valid’. Leitgeb. In his ‘On the metatheory of Field’s Solving the Paradoxes, Escaping Revenge’, Hannes Leitgeb argues that whether, in the end, Field’s proposed theory escapes revenge turns on the details of its metatheory. Leitgeb argues that without a clear, explicitly formulated metatheory, the intended interpretation of Field’s proposed truth theory—and, hence, the proposed resolution of paradox—remains unclear. What is ultimately required, Leitgeb argues, is a metatheory that includes a nonclassical set theory for which the logic is the logic of Field’s truth theory.²³ Towards moving matters forward, Leitgeb sketches two target metatheories, a classical and a non-classical one. Leitgeb conjectures that, for reasons he discusses, revenge may emerge for Field’s proposal once a full metatheory is in place. Maudlin. In his ‘Reducing revenge to discomfort’, Tim Maudlin argues that the revenge phenomenon ultimately teaches us something about our normative principles of assertion. As in §1.3 (above), invoking a new category for Liars—say, bugger—seems inevitably to lead to new Liars (e.g. the starred sentence in §1.3 above). While Maudlin maintains that we do need three semantic categories (viz., truth, falsity, and ungroundedness), he argues that we need no more than three. In particular, we may—and should—assert that the ticked sentence in §1.2 above is not true; it’s just that we’ll be bucking the traditional principle according to which only truths are properly assertible. The problem, Maudlin argues, is not with principles of truth (e.g. Release and Capture); the problem is with the traditional principle of assertion.²⁴ On the other hand, Maudlin admits that the revenge phenomenon returns even for his revised principle of assertion (e.g. ‘I am not properly assertible according to Maudlin’s revised principles’). Maudlin argues that this is revenge, but that it is at most a discomfort; it is far from threatening the coherence of truth. Patterson. In his ‘Understanding the liar’, Douglas Patterson advances an ‘inconsistency view’ of the semantic paradoxes in English; however, his view is not a dialetheic ²³ Actually, Leitgeb’s claim needn’t be that the metatheory include a non-classical set theory, but rather that it include a non-classical theory of objects that play the relevant role that sets typically play—e.g. serving as a ‘model’ or etc. ²⁴ I should ﬂag one potential confusion here. Maudlin claims, at least in his fuller work (see references in Chapter 7), that Rule Capture is valid, in the sense that, necessarily, if α is true, then so too is Tr(α). At the same time, the logic governing assertibility is closer to KF, where α ⇒ Tr(α) fails in Rule form (and Conditional form).

Prolegomenon to Future Revenge / 17 view (according to which English is inconsistent, in the sense that some true English sentence has a true negation). Patterson argues that such a view is not that natural languages are inconsistent, but rather that competent speakers of natural languages process such languages in accord with an inconsistent theory. One of Patterson’s principal aims is to show that, perhaps contrary to common thinking, understanding a language can be—and, in the case of English, is—a relation to a false theory. Patterson argues that such an ‘inconsistency view’ is the most promising lesson to draw from the revenge phenomenon. Priest. In his ‘Revenge, Field, and ZF’, Graham Priest does three things. First, Priest characterizes the Liar’s revenge, and carves up three options for dealing with it. Second, Priest directs the discussion towards Field’s chapter (see Chapter 4), and argues that Field’s proposal is not revenge-free, contrary to Field; in particular, it faces an expected problem with the notion of having value 1. (Priest anticipates the immediate thought that, as sketched in §1.4.1 above, he is merely launching a form of ‘too-easy revenge’, conﬂating model-dependent and ‘real’ notions. Priest argues that unless ‘having value 1’ is a real notion, Field has given no reason to think that Field’s proposed logic has anything to do with real validity—i.e. validity in our real language.) Third, Priest argues that the (alleged) troubles facing Field’s proposal are a symptom of deeper revenge in the background theory of ZF, which theory, Priest argues, itself faces a serious revenge-like situation involving V (the cumulative hierarchy): the logic deﬁned by the theory (in terms of models) does not apply to the theory itself, thereby leaving us ‘bereft of a justiﬁcation for reasoning about sets’, as Priest puts it. Rayo and Welch. In their ‘Field on revenge’, Agust´ın Rayo and Philip Welch argue that Field’s allegedly revenge-free truth theory (see Chapter 4) is not really revengefree—or, at least, that its prospects for being revenge-free crucially depend on the outcome of current debates over higher-order languages. Rayo and Welch argue that, just as ‘received wisdom’ maintains, Field’s proposed theory enjoys consistency only in virtue of expressive limitations. In particular, by invoking the appropriate higher-order language, we can explicitly characterize a key semantic notion involved in Field’s proposal: viz., an intended interpretation of L+ , where L+ is the language of Field’s theory (a language enjoying transparent truth and a suitable conditional). Such a notion, as Rayo and Welch argue, plays the Liar’s revenge role: it would generate inconsistency were it expressible in Field’s proposed language.²⁵ Read. In his ‘Bradwardine’s revenge’, Stephen Read discusses a theory of truth proposed by Thomas Bradwardine (who was principally a physicist and theologian in ²⁵ I should be slightly more precise and note that Field (Chapter 4) considers a class of languages (or theories) that enjoy the desiderata of transparent truth and a suitable conditional, and Rayo and Welch direct their remarks against the relevant class.

18 / JC Beall the 1300s). Read shows that Bradwardine’s theory, according to which Liars are not true (because they’d have to be true and not true, which is impossible), is a subtler theory than the later Buridan-like theories that, in effect, reject unrestricted Capture for truth (see §1.1 above). Moreover, the theory, on the surface, as Read argues, seems to promise a revenge-free approach to a whole host of semantic paradoxes. The key for Bradwardine is to distinguish between the claim that the Liar is false from the Liar itself. The propositions appear to be indistinguishable, but they are not. According to Bradwardine, any proposition that ‘says’ of itself that it is false, also ‘says’ of itself that it is true. (As Read points out, this is a subtler thesis than the later Buridanian claim that every claim ‘says’ of itself that it is true.) Restall. In his ‘Curry’s revenge: the costs of non-classical solutions to the paradoxes of self-reference’, Greg Restall discusses the challenges posed by Curry’s paradox to those (non-classical) theories that attempt to preserve Capture and Release, in both Rule and Conditional forms, for truth and related semantic (or logical) notions—e.g. ‘semantical properties’, which serve as the ‘extensions’ of predicates in na¨ıve semantics.²⁶ Restall argues that a Curry conditional is fairly easy to construct unless the language has fairly narrow limits. In particular, a theory that avoids Curry paradox must either reject ‘large disjunctions’, various (otherwise natural) forms of distribution, or the transitivity of entailment. As Restall notes, whatever option is rejected, sound philosophical motivation must accompany the rejection. Scharp. In his wide-ranging ‘Aletheic vengeance’, Kevin Scharp argues that the Liar’s revenge teaches us, among other things, that truth is an inconsistent concept the best theory of which implies that typical truth rules are ‘constitutive’ of truth but nonetheless invalid. Scharp argues that the best (inconsistency) theory of truth takes truth to be a confused concept (in a technical sense), but is a theory that does not use our concept of truth at all. Indeed, Scharp proposes that the proper approach to truth is one that ﬁnds other—non-confused—notions to play the truth role(s). Shapiro. In his ‘Burali-Forti’s revenge’, Stewart Shapiro turns the focus from the Liar paradox to the Burali-Forti paradox, which, he argues, has its own revenge issues. (Using the later von Neumann account, which came after Burali-Forti, the paradox, in short, is that the set of all ordinals satisﬁes all that’s required to be an ordinal, in which case, the successor of , namely + 1, is strictly greater than . But, being ²⁶ Restall doesn’t use the term ‘semantical properties’, but he clearly has this under discussion. (Some philosophers refer to the target entities as ‘na¨ıve sets’, but sets ultimately have little to do with the matter. If we let mathematicians tell us the ‘nature’ of sets—and they’ll likely do so by axiomatizing away Russell problems—we still have to ﬁnd a theory of ‘semantical properties’, the entities that play the familiar role in semantics, namely, those objects ‘expressed’ by any meaningful predicate and ‘exempliﬁed’ by an object just if the given predicate is ‘true of ’ the object.)

Prolegomenon to Future Revenge / 19 itself an ordinal, + 1 must be in , giving the result that < + 1 ≤ , which is impossible.) Shapiro presents the paradox and a variety of ways of dealing with it. He argues that each option faces severe problems, leaving the matter open. Simmons. In his ‘Revenge and context’, Keith Simmons ﬁrst distinguishes between (what he calls) direct revenge and second-order revenge. The former variety is the (what one might call ‘ﬁrst-order’) variety: we already have a stock of semantic terms, and they generate paradox. In particular, as with the ticked sentence in §1.3, one is naturally inclined to classify it as ‘neither true nor false’, but this (at least prima facie) implies untruth, and the paradox remains. One is stuck in direct revenge: an inability to classify the sentence as one thinks it ought to be classiﬁed—but cannot be so classiﬁed, on pain of inconsistency. But, now, one introduces new, technical machinery to deal with the direct revenge problem: one calls the Liar a bugger, or unstable, or whatnot. Second-order revenge emerges with this new machinery, and one is, again, unable to classify the (new) Liars as they ‘ought’ to be (in some sense). Simmons argues that, while there is still work to be done, his ‘singularity theory’ of semantic notions deals not only with direct revenge in a natural way; it also holds the promise of resolving second-order revenge.

Appendix Since many of the chapters in this volume presuppose familiarity with so-called ﬁxedpoint languages, and, in particular, paracomplete languages (see below), this appendix is intended as a user-friendly sketch of the (or a) basic background picture. In particular, I sketch a basic Kripkean picture [11], although I take liberties in the setting up.²⁷ I focus on the non-classical interpretation of Kripke’s (least ﬁxed point) account. My aim is only to give a basic philosophical picture and a sketch of the formal model. I focus on the semantic picture. Philosophical picture One conception of truth has it that truth is entirely transparent, that is, a truth predicate Tr(x) in (and for) our language such that Tr(α) and α are intersubstitutable in all (non-opaque) contexts, for all α in the language. This conception comes with a guiding metaphor, according to which ‘true’ is introduced only for purposes of generalization. Prior to introducing the device, we spoke only the ‘true’-free fragment. ²⁷ This appendix is a very slightly altered version of a section from the much larger [2], which provides more references.

20 / JC Beall (Similarly for other semantic notions/devices, e.g., ‘denotes’, ‘satisﬁes’, ‘true of’, etc.) For simplicity, let us assume that the given ‘semantic-free’ fragment (hence, ‘true’-free fragment) is such that LEM holds.²⁸ Letting L0 be our ‘semantic-free fragment’, we suppose that α ∨ ¬α is true for all α in L0 .²⁹ Indeed, we may suppose that classical semantics—and logic, generally—is entirely appropriate for the fragment L0 . But now we want our generalization-device. How do we want this to work? As above, we want Tr(α) and α to be intersubstitutable for all α. The trouble, of course, is that once ‘is true’ is introduced into the language, various unintended—and, given the role of the device, paradoxical—sentences emerge (e.g. the ticked sentence in §1.2 above).³⁰ The paracomplete idea, of which Kripke’s is the best known, is (in effect) to allow some instances of α ∨ ¬α to ‘fail’.³¹ In particular, if α itself fails to ‘ground out’ in L0 , fails to ‘ﬁnd a value’ by being ultimately equivalent to a sentence in L0 , then the α-instance of LEM should fail. (This is the so-called least ﬁxed point picture.) Kripke illustrated the idea in terms of a learning or teaching process. The guiding principle is that Tr(α) is to be asserted exactly when α is to be asserted. Consider an L0 -sentence that you’re prepared to assert—say, ‘1 + 1 = 2’ or ‘Max is a cat’ or whatever. Heeding the guiding principle, you may then assert that ‘1 + 1 = 2’ and ‘Max is a cat’ are true. In turn, since you are now prepared to assert (1)

‘Max is a cat’ is true

the guiding principle instructs that you may also assert (2)

‘ ‘Max is a cat’ is true’ is true.

And so on. More generally, your learning can be seen as a process of achieving further and further truth-attributions to sentences that ‘ground out’ in L0 . (Similarly for falsity, which is just truth of negation.) Eventually, your competence reﬂects precisely

²⁸ This assumption sets aside the issue of vagueness (and related sorites puzzles). I am setting this aside only for simplicity. The issue of vagueness—or, as some say, ‘indeterminacy’, in general—is quite relevant to some paracomplete approaches to truth. See [4], [13], [21]. ²⁹ This assumption is not essential to Kripke’s account; however, it makes the basic picture much easier to see. ³⁰ With respect to formal languages, the inevitability of such sentences is enshrined in G¨odel’s so-called diagonal lemma. (Even though the result is itself quite signiﬁcant, it is standardly called a lemma because of its role in establishing G¨odel–Tarski indeﬁnability theorems. For user-friendly discussion of the limitative results, and for primary sources, see [18]. For a general discussion of diagonalization, see [19].) ³¹ NB: The sense in which instances of α ∨ ¬α ‘fails’ is modeled by such instances being undesignated (in the formal model). (See ‘Formal model’ below.) How, if at all, such ‘failure’ is expressed in the given language is relevant to ‘revenge’, but I will leave chapters of this volume to discuss that.

Prolegomenon to Future Revenge / 21 the deﬁning intersubstitutivity—and transparency—of truth: that Tr(α) and α are intersubstitutable for all α of the language. But your competence also reﬂects something else: namely, the failure to assert either α or ¬α, for some α in the language. To see the point, think of the above process of ‘further and further truth-attributions’ as a process of writing two (very, very big) books—one, The Truth, the other The False. Think of each stage in the process as completing a ‘chapter’, with chapter zero of each book being empty—this indicating that at the beginning nothing is explicitly recorded as true (or, derivately, false). Concentrate just on the process of recording atomics in The Truth. When you were ﬁrst learning, you scanned L0 (semantic-free fragment) for the true (atomic) sentences, the sentences you were prepared to assert. Chapter one of The Truth comprises the results of your search—sentences such as ‘Max is a cat’ and the like. In other words, letting ‘I(t)’ abbreviate the denotation of t, chapter one of The Truth contains all of those atomics α(t) such that I(t) exempliﬁes α, a ‘fact’ that would’ve been recorded in chapter zero had chapter zero recorded the true semantic-free sentences. (For simplicity, if α(t) is an L0 -atomic such that I(t) exempliﬁes α, then we’ll say that I(t) exempliﬁes α according to chapter zero. In the case of ‘Max is a cat’, chapter zero has it that Max exempliﬁes cathood, even though neither ‘Max is a cat’ nor anything else appears in chapter zero.) In the other book, The False, chapter zero is similarly empty; however, like chapter zero of The Truth, the sentences that would go into The False’s chapter zero are those (atomic) L0 -sentences that, according to the world (as it were), are false—e.g., ‘1 + 1 = 3’, ‘Max is a dog’, or the like.³² If α(t) is a false L0 -atomic, we’ll say that according to chapter zero, I(t) exempliﬁes ¬α (even though, as above, chapter zero explicitly records nothing at all). In turn, chapter one of The False contains all of those atomics α(t) such that, according to chapter zero, I(t) exempliﬁes ¬α (i.e., the L0 -atomics that are false, even though you wouldn’t say as much at this stage). And now the writing (of atomics) continues: chapter two of The Truth comprises ‘ﬁrst-degree truth-attributions’ and atomics α(t) such that, as above, I(t) exempliﬁes α according to chapter one, sentences like (1) and ‘Max is a cat’. In turn, chapter three of The Truth comprises ‘second-degree’ attributions, such as (2), and atomics α(t) such that (as it were) t is α according to chapter two. And so on, and similarly for The False. In general, your writing-project exhibits a pattern. Where Ii (Tr) is chapter i of The Truth, the pattern runs thus: Ii+1 (Tr) = Ii (Tr) ∪ {α(t) : α(t)is an atomic and I(t) exempliﬁes α according to Ii (Tr)} ³² For convenience, we’ll also put non-sentences into The False. Putting non-sentences into The False is not essential to Kripke’s construction, but it makes things easier. Obviously, one can’t write a cat but, for present purposes, one can think of The False as a special book that comes equipped with attached nets (wherein non-sentences go), a net for each chapter.

22 / JC Beall Let S comprise all sentences of the language. With respect to The False book, the pattern of your writing (with respect to atomics) looks thus: Ii+1 (F) = Ii (F) ∪ {α(t) : α(t) is an atomic and I(t) ∈ /S or I(t) exempliﬁes ¬α according to Ii (Tr)} So goes the basic process for atomics. But what about compound (molecular) sentences? The details are sketched below (see ‘Formal model’), but for now the basic idea is as follows (here skipping the relativizing to chapters). With respect to negations, ¬α goes into The True just when α goes into The False. (Otherwise, neither α nor ¬α ﬁnds a place in either book.) With respect to conjunctions, α ∧ β goes into The False if either α or β goes into The False, and it goes into The True just if both α and β go into The True. (Otherwise, α ∧ β ﬁnds a place in neither book.) The case of disjunctions is dual, and the quantiﬁers may be treated as ‘generalized conjunction’ (universal) and ‘generalized disjunction’ (existential). This approach to compound sentences reﬂects the so-called Strong Kleene scheme, which is given below (see ‘Formal model’). Does every sentence eventually ﬁnd a place in one book or other? No. Consider an atomic sentence λ, like the ticked sentence in §1.2, equivalent to ¬Tr(λ). In order to get λ into The True book, there’d have to be some chapter in which it appears. λ doesn’t appear in chapter zero, since nothing does. Moreover, λ doesn’t exemplify anything ‘according to chapter zero’, since chapter zero concerns only the L0 -sentences (and λ isn’t one of those). What about chapter one? In order for λ to appear in chapter one, λ would have to be in chapter zero or be such that λ exempliﬁes ¬Tr(x) according to chapter zero. But for reasons just given, λ satisﬁes neither disjunct, and so doesn’t appear in chapter one. The same is evident for chapter two, chapter three, and so on. Moreover, the same reasoning indicates that λ doesn’t appear in The False book. In general, Liar-like sentences such as the ticked sentence in §1.2 will ﬁnd a place in one of our books only if it ﬁnds a place in one of the chapters Ii (Tr) or Ii (F). But the ticked sentence will ﬁnd a place in Ii (Tr) or Ii (F) only if it ﬁnds a place in Ii−1 (Tr) or Ii−1 (F). But, again, the ticked sentence will ﬁnd a place in Ii−1 (Tr) or Ii−1 (F) only if it ﬁnds a place in Ii−2 (Tr) or Ii−2 (F). And so on. But, then, since I0 (Tr) and I0 (F) are both empty, and since—by our stipulation—something exempliﬁes a property according to I0 (Tr) only if the property is a non-semantic one (the predicate is in L0 ), the ticked sentence (or the like) fails to ﬁnd a place in either book. Such a sentence, according to Kripke, is not only ungrounded, since it ﬁnds a place in neither book, but also paradoxical—it couldn’t ﬁnd a place in either book.³³ ³³ The force of couldn’t here is made precise by the full semantics, but for present purposes one can think of couldn’t along the lines of on pain of (negation-) inconsistency or, for that matter, on pain of being in both books (something impossible, on the current framework).

Prolegomenon to Future Revenge / 23 So goes the basic philosophical picture. What was wanted was an account of how, despite the existence of Liars, we could have a fully transparent truth predicate in the language—and do so without triviality (or, in Kripke’s case, inconsistency). The foregoing picture suggests an answer, at least if we eventually have a chapter Ii (Tr) such that Tr(α) is in Ii (Tr) if and only if α is in Ii (Tr), and similarly a chapter for The False. What Kripke (and, independently, Martin–Woodruff) showed is that, provided our ‘writing process’ follows the right sort of scheme (in effect, a logic weaker than classical), our books will contain such target chapters, and in that respect our language can enjoy a (non-trivial, indeed consistent) transparent truth predicate. Making the philosophical picture more precise is the job of formal, philosophical modeling, to which I now brieﬂy (and somewhat informally) turn.

Formal model For present purposes, I focus on what is known as Kripke’s ‘least ﬁxed point’ model (with empty ground model). I leave proofs to cited works (all of which are readily available), and try to say just enough to see how the formal picture goes. Following standard practice, we can think of an interpreted language L as a triple L, M, σ , where L is the syntax (the relevant syntactical information), M is an ‘interpretation’ or ‘model’ that provides interpretations to the non-logical constants (names, function-symbols, predicates), and σ is a ‘semantic scheme’ or ‘valuation scheme’ that, in effect, provides interpretations (semantic values) to compound sentences.³⁴ Consider, for example, familiar classical languages, where the set V of ‘semantic values’ is {1, 0}. In classical languages, M = D, I , with D our (non-empty) domain and I an ‘interpretation-function’ that assigns to each name an element of D (the denotation of the name), assigns to each n-ary function-symbol an element of Dn −→ D, that is, an n-ary function from Dn into D, and assigns to each n-ary predicate an element of Dn −→ V, a function—sometimes thought of as the intension of the predicate—taking n-tuples of D and yielding a ‘semantic value’ (a ‘truth value’). The extension of an n-ary predicate F (intuitively, the set of things of which F is true) contains all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 1. The classical valuation scheme τ (for Tarski) is the familiar one according to which a negation is true (in a given model) exactly when its negatum is false (in the given model), a disjunction is true (in a model) iff one of the disjuncts is true (in the model), and existential sentences are treated as generalized disjunctions.³⁵ ³⁴ For present purposes, a semantic scheme or valuation scheme σ is simply some general deﬁnition of truth (falsity) in a model. For more involved discussion of semantic schemes, see [8]. ³⁵ I assume familiarity with the basic classical picture, including ‘true in L’ and so on. To make things easier, I will sometimes assume that we’ve moved to models in which everything in the domain has a name, and otherwise I’ll assume familiarity with standard accounts of ‘satisﬁes α(x) in L’.

24 / JC Beall Classical languages (with suitably resourceful L) cannot have their own transparent truth predicate. Paracomplete languages reject the ‘exhaustive’ feature implicit in classical languages: namely, that a sentence or its negation is true, for all sentences. The standard way of formalizing paracomplete languages expands the interpretation of predicates. Recall that in your ‘writing process’ some sentences (e.g. Liars) found a place in neither book. We need to make room for such sentences, and we can expand our semantic values V to do so; we can let V = {1, 12 , 0}, letting the middle value represent (for ‘modeling’ purposes) the status of sentences that found a place in neither book. Generalizing (but, now, straining) the metaphor, we can think of all n-ary predicates as tied to two such ‘big books’, one recording the objects of which the predicate is true, the other the objects of which it is false. On this picture, the extension of a predicate F remains as per the classical (containing all n-tuples of which the predicate is true), but we now also acknowledge an antiextension, this comprising all n-tuples of which the predicate is false. This broader picture of predicates enjoys the classical picture as a special case: namely, where we stipulate that, for any predicate, the extension and antiextension are jointly exhaustive (the union of the two equals the domain) and, of course, exclusive (the intersection of the two is empty). Concentrating on the so-called Strong Kleene account [36],³⁶ the formal story runs as follows. We expand V, as above, to be {1, 12 , 0}, and so our language Lκ = L, M, κ is now a so-called three-valued language (because it uses three semantic values).³⁷ Our designated values—intuitively, the values in terms of which validity or consequence is deﬁned—are a subset of our semantic values; in the Strong Kleene case, there is exactly one designated element, namely 1. A (Strong Kleene) model M = D, I is much as before, with I doing exactly what it did in the classical case except that I now assigns to n-ary predicates elements of Dn −→ {1, 12 , 0}, since V = {1, 12 , 0}. Accordingly, the ‘intensions’ of our paracomplete (Strong Kleene) predicates have three options: 1, 12 , and 0. What about extensions? As above, we want to treat predicates not just in terms of extensions (as in the ³⁶ This is one of the paracomplete languages for which Kripke proved his deﬁnability result. Martin–Woodruff proved a special case of Kripke’s general ‘ﬁxed point’ result, namely, the case for so-called ‘maximal ﬁxed points’ of the Weak Kleene scheme, or Weak Kleene languages. ³⁷ Kripke [11] made much of emphasizing that ‘the third value’ is not to be understood as a third truth value or anything else other than ‘undeﬁned’ (along the lines of Kleene’s original work [10] ). I will not make much of this here, although what to make of semantic values that appear in one’s formal account is an important, philosophical issue. (Note that if one wants to avoid a three-valued language, one can let V = {1, 0} and proceed to construct a Kleene-language by using partial functions (hence, the standard terminology ‘partial predicates’) for interpretations. I think that this is ultimately merely terminological, but I won’t dwell on the matter here.

Prolegomenon to Future Revenge / 25 classical languages) but also antiextensions. The extension of an n-ary predicate F, just as before, comprises all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 1. (Again, intuitively, this remains the set of objects of which F is true.) The antiextension, in turn, comprises all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 0. (Again, intuitively, this is the set of objects of which F is false.) Of course, as intended, an interpretation might fail to put x in either the extension or antiextension of F. In that case, we say (in our ‘metalanguage’) that, relative to the model, F is undeﬁned for x.³⁸ Letting F+ and F− be the extension and antiextension of F, respectively, it is easy to see that, as noted above, classical languages are a special case of (Strong Kleene) paracomplete languages. Paracomplete languages typically eschew inconsistency, and so typically demand that F+ ∩ F− = ∅, in other words, that nothing is in both the extension and antiextension of any predicate. In this way, paracomplete languages typically agree with classical languages. The difference, of course, is that paracomplete languages do not demand that F+ ∪ F− = D for all predicates F. But paracomplete languages allow for such ‘exhaustive constraints’, and in that respect can enjoy classical languages as a special case. To see the close relation between classical languages and Strong Kleene, notice that κ, the Strong Kleene valuation-scheme, runs as follows (here treating only ¬, ∨, and ∃). Where VM (α) is the semantic value of α in M (and, for simplicity, letting each object in the domain name itself), and, for purposes of specifying scheme κ, treating V as standardly (linearly) ordered: K1. VM (¬α) = 1 − VM (α). K2. VM (α ∨ β) = max(VM (α), VM (β)). K3. VM (∃x α(x)) = max{VM (α(t/x)) : for all t ∈ D}. The extent to which classical logic is an extension of a given paracomplete logic depends on the semantic scheme of the language.³⁹ Since κ, as above, is entirely in keeping with the classical scheme except for ‘adding an extra possibility’, it is clear that every classical interpretation is a Strong Kleene-interpretation (but not vice versa).⁴⁰ ³⁸ A common way of speaking is to say that, for example, F(t) is ‘gappy’ with respect to I(t). This terminology is appropriate if one is clear on the relation between one’s formal model and the target notions that the model is intended to serve (in one respect or other), but the terminology can also be confusing, since, e.g., in the current Strong Kleene language, one cannot truly assert of any α that α is ‘gappy’, i.e. ¬Tr(α) ∧ ¬Tr(¬α). (This issue arises in various chapters in the current volume.) ³⁹ Here, perhaps not altogether appropriately, I am privileging model theory over proof theory, thinking of ‘logic’ as the semantic consequence relation that falls out of the semantics. This is in keeping with the elementary aims of the essay, even though (admittedly) it blurs over a lot of philosophical and logical issues. ⁴⁰ Note that in classical languages, VM (A) ∈ {1, 0} for any A, and the familiar classical clauses on connectives are simply (K1)–(K3).

26 / JC Beall Let us say that an interpretation veriﬁes a sentence α iff α is designated (in this case, assigned 1) on that interpretation, and that an interpretation veriﬁes a set of sentences iff it veriﬁes every element of . We deﬁne semantic consequence in familiar terms: α is a consequence of iff every interpretation that veriﬁes also veriﬁes α. I will use ‘SK ’ for the Strong Kleene consequence relation, so understood. Let us say that a sentence α is logically true in Lκ exactly if ∅ Sκ α, that is, iff α is designated (assigned 1) in every model. A remarkable feature of Lκ is that there are no logical truths. To see this, just consider an interpretation that assigns 12 to every atomic, in which case, as an induction will show, every sentence is assigned 12 on that interpretation. Hence, there’s some interpretation in which no sentence is designated, and hence no sentence designated on all interpretations. A fortiori, LEM fails in Strong Kleene languages.⁴¹ And now an answer to one guiding question becomes apparent. What we want is a model of how our language can be non-trivial (indeed, consistent) while containing both a transparent truth predicate and Liar-like sentences. In large part, the answer is that our language is (in relevant respects) along Strong Kleene lines, that the logic is weaker than classical logic. Such a language, as Kripke showed, can contain its own (transparent) truth predicate. The construction runs (in effect) along the lines of the ‘big books’ picture. For simplicity, let Lκ be a classical (but nonetheless Strong Kleene) language such that L (the basic syntax, etc.) is free of semantic terms but has the resources to describe its given syntax—including, among other things, having a name α for each sentence α. (In other words, I assigns to each n-ary predicate an element of Dn −→ {1, 0}, even though the values V of Lκ also contain 12 .) What we want to do is move to a richer language the syntax Lt of which contains Tr(x), a unary predicate intended to be a transparent truth predicate for the enriched language. For simplicity, assume that the domain D of Lκ contains all sentences of Lt .⁴² Think, brieﬂy, about the ‘big books’ picture. One can think of each successive ‘chapter’ as a language that expands one’s ofﬁcial record of what is true (false). More formally, one can think of each such ‘chapter’ of both books as the extension and antiextension of ‘true’, with each such chapter expanding the interpretation of ‘true’. Intuitively (with slight qualiﬁcations about chapters zero), one can think of Ii+1 (Tr) as explicitly recording what is true according to chapter Ii (Tr). The goal, of course, is to ﬁnd a ‘chapter’ at which we have Ii+1 (Tr) = Ii (Tr), a ‘ﬁxed point’ at which anything ⁴¹ This is not to say, of course, that one can’t have a Strong Kleene—or, in general, paracomplete—language some proper fragment of which is such that α ∨ ¬α holds for all α in the proper fragment. (One might, e.g., stipulate that arithmetic is such that α ∨ ¬α holds.) ⁴² This is usually put (more precisely) as that the domain contains the G¨odel-codes of all such sentences, but for present purposes I will skip over the mathematical details.

Prolegomenon to Future Revenge / 27 true in the language is fully recorded in the given chapter—one needn’t go further. Thinking of the various ‘chapters’ as languages, each with a richer interpretation of ‘true’, one can think of the ‘ﬁxed chapter’ as a language that, ﬁnally, has a transparent truth predicate for itself. Returning to the construction at hand, we have our Strong Kleene (but classical) ‘ground language’ Lκ that we now expand to Ltκ , the syntax of which includes that of Lκ but also has Tr(x) (and the resulting sentences formable therefrom). We want the new language to ‘expand’ the ground language, and we want the former to have a model that differs from the latter only in that it assigns an interpretation to Tr(x). For present purposes, we let It , the interpretation function in Ltκ , assign (∅, ∅) to Tr(x), where (∅, ∅) is the function that assigns 12 to each element of Dt . (Hence, the extension and antiextension of Tr(x) in Ltκ are both empty.) This is the formal analogue of ‘chapter zero’. The crucial question, of course, concerns further expansion. How do we expand the interpretation of Tr(x)? How do we move to ‘other chapters’? How, in short, do we eventually reach a ‘chapter’ or language in which we have a transparent truth predicate for the whole given language? This is the role of Kripke’s ‘jump operator’. What we want, of course, are ‘increasingly informative’ interpretations (Ti + , Ti − ) of Tr(x), but interpretations that not only ‘expand’ the previous interpretations but also preserve what has already been interpreted. If α is true according to chapter i, then we want as much preserved: that α remain true according to chapter i + 1. This is the role of the ‘jump operator’, a role that is achievable given the so-called monotonicity of Strong Kleene valuation scheme κ.⁴³ The role of the jump operator is to eventually ‘jump’ through successive interpretations (chapters, languages) Ii (Tr) and land on one that serves the role of transparent truth—serves as an interpretation of ‘is true’. As above, letting Ii (Tr) be a function (Ti + , Ti − ) yielding ‘both chapters i’, the goal is to eventually ‘jump’ upon an interpretation (Ti + , Ti − ) such that + − (Ti + , Ti − ) = (Ti+1 , Ti+1 ). Focusing on the ‘least such point’ in the Strong Kleene setting, Kripke’s construction proceeds as above. We begin at stage 0 at which Tr(x) is interpreted as (∅, ∅), and we ⁴³ Monotonicity is the crucial ingredient in Kripke’s (similarly, Martin–Woodruff ’s) general result. Let M and M be paracomplete (partial) models for (uninterpreted) L. Let FM+ be the extension of F in M, and similarly FM+ for M . (Similarly for antiextension.) Then M extends M iff the models have the same domain, agree on interpretations of names and function signs, and FM+ ⊆ FM+ and FM− ⊆ FM− for all predicates F that M and M interpret. (In other words, M doesn’t change M’s interpretation; it simply interprets whatever, if anything, M left uninterpreted.) Monotonicity property: A semantic (valuation) scheme σ is monotone iff for any α that is interpreted by both models, α’s being designated in M implies its being designated in M whenever M extends M. So, the montonicity property of a scheme ensures that it ‘preserves truth (falsity)’ of ‘prior interpretations’ in the desired fashion.

28 / JC Beall + − , Ti+1 ) deﬁne a ‘jump operator’ on such interpretations:⁴⁴ Tr(x) is interpreted as (Ti+1 + + − at stage i + 1 if interpreted as (Ti , Ti ) at the preceding stage i, where, note well, Ti+1 comprises the sentences that are true (designated) at the preceding stage (chapter, − language) i, and Ti+1 the false sentences (and, for simplicity, non-sentences) at i. Accordingly, we deﬁne the ‘jump operator’ JSK thus:⁴⁵ + − , Ti+1 ) JSK (Ti + , Ti − ) = (Ti+1

The jump operator yields a sequence of richer and richer interpretations that ‘preserve prior information’ (given monotonicity), a process that can be extended into the transﬁnite to yield a sequence (T0+ , T0− ), (T1+ , T1− ), . . . , (Tγ+ , Tγ− ), . . . deﬁned (via transﬁnite recursion) thus:⁴⁶ Jb. Base. (T0+ , T0− ) = (∅, ∅). Js. Successor. (Tγ++1 , Tγ−+1 ) = JSK ((Tγ+ , Tγ− )). Jl. Limit. For limit stages, we collect up by unionizing the prior stages: + − + − Tε , Tε (Tλ , Tλ ) = ε 1, ‘Pn (x)’ is the nth ‘pathogical’ predicate.

40 / Roy T. Cook the newest truth-value, and we also add a new conditional. Each conditional is a better approximation of the ‘ideal’ but unattainable conditional which satisﬁes all of the standard axioms and inference rules, including modus ponens and conditional proof.¹³ We assume that G¨odel coding is carried out such that, for any sentence in Lα , its G¨odel code (written < >) relative to Lα , if it has one, is the same as its G¨odel code relative to Lβ for any β > α. Note that there will be languages Lβ such that some sentences of Lβ will not receive G¨odel codes (since if β is large enough, Lβ is uncountable, yet we only have a countable inﬁnity of numerals to serve as codes). We construct a model for each language as follows: The model of L0 is just any classical model of set theory < D, I > where D is the domain and I is an interpretation function mapping sentences onto {t, f}. For each Lα (α ≥ 0) we construct an n+3-valued Kripke-style ﬁxed point model < D, Iα > recursively as follows. The truth-values (i.e. the range of Iα ) are: {t, f, n} ∪ {pβ : β ≤ α}

where t is the value true, f is false, pβ is the β th ‘pathological’ value, and n is a placeholder value given to sentences involving the application of a semantic predicate to a numeral that either is not the G¨odel code of any sentence or is the G¨odel code of a sentence which has not yet been added to the language (i.e. n is the value sentences get when they have not yet received a ‘real’ value). In other words, we do not assign a legitimate truth-value to the sentence ‘5 is true’ (or its negation) if 5 is not the G¨odel code of any sentence (in any of our languages), nor does the sentence ‘57568489 is true’ obtain a legitimate truth-value in L3 if the sentence coded by 57568489 is a sentence which contains the predicate ‘P3 (x)’. Given an assignment of semantic values to the atomic formulas in a language Lα , the interpretation of the logical connectives is determined as follows: I( ∧ )

= min{I(), I()} relative to the ordering:¹⁴ n < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < f < t

I( ∨ )

= max{I(), I()} relative to the ordering: f < t < p1 < p2 < . . . < pβ < pβ+1 < . . . < pα < n

¹³ The fact that one cannot have a wholly satisfactory conditional and a wholly satisfactory truth predicate in the same language is well known, and, like Field (2003), we opt here for retaining the na¨ıve notion of truth, and settling for an account of the conditional (and thus the biconditional) which does not provide everything that we might wish for. In the ﬁnal section of the chapter the present approach, which posits a never-ending series of conditionals that provide better and better approximations of the ‘ideal’ conditional, is contrasted with Field’s approach, where a single, rather complicated, conditional is constructed in the hopes of getting the closest approximation possible. ¹⁴ The clauses for the binary connectives are just a generalization of the weak Kleene scheme (Kleene (1952)). Interestingly, the strong Kleene scheme will not work in this framework.

Embracing Revenge / 41 I(¬) =

For β < α: I( →β )

f if I() = t t if I() = f I() otherwise. = n

if I() = n or I() = n

max{I(), I()} relative to the ordering: f < t < p1 < p2 < . . . < pβ < pβ+1 < . . . < pα < n if I() = n and I() = n and I() = pδ or I() = pδ where δ > β t

if I() = n and I() = n and for any δ > β, I() = pδ and I() = pδ and I() ≤ I() relative to the ordering: f < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < t

f

otherwise.

The universal and existential quantiﬁers are treated as generalized versions of conjunction and disjunction respectively. We now construct a sequence of models < D, Iα β > via transﬁnite recursion (the ﬁnal model < D, Iα > for each language Lα will be the least ﬁxed point in this sequence). Given the clauses above, we need merely to specify the interpretation of atomic sentences in each model in the series.: For any atomic sentence of Lα+1 : Base: Iα+1 0 ()

=

n Iα () pα+1

if = T(n) or = F(n) or = Pi (n) where n is not the G¨odel code of a sentence in Lα+1 if is a sentence of Lα . otherwise.

Successor: If = T(< >) then: Iα+1 β+1 ()

=

Iα+1 β ()

If = F(< >) then: Iα+1 β+1 () = t = f = Iα+1 β ()

if Iα+1 β () ifIα+1 β () otherwise

= =

f t

= =

pδ t

If = Pδ (< >) for some δ < α, then: Iα+1 β+1 ()

= = =

t f Iα+1 β ()

if Iα+1 β () ifIα+1 β () otherwise

42 / Roy T. Cook If = T(< >) and = F(< >) and = Pδ (< >) then: Iα+1 β+1 ()

=

Iα+1 β ()

Limit: = max{Iα+1 β (): β < γ } relative to the partial ordering: For any δ : n < pγ , pγ < t, pγ < f, and pγ < pδ iff δ < γ ¹⁵

Iα+1 γ ()

All operators, quantiﬁers, etc. in the above scheme are monotonic with respect to the partial ordering of truth-values utilized in the limit case above. Thus, the sequence of models described above will have a ﬁxed point—that is—a model < D, Iα+1 β > such that: < D, Iα+1 β > = < D, Iα+1 β+1 > < D, Iα+1 >, our model of Lα+1 , is just the minimal such ﬁxed point (for the mathematical details of such ﬁxed point constructions see Kripke (1972) or Fitting (1986)). We extend the series of models into the transﬁnite by adding the additional clause handling atomic sentences in the base case for Lγ where γ a limit: Base: Iγ 0 ()

=

max{Iβ (): β < γ } relative to the partial ordering: For any δ : n < pγ , pγ < t, pγ < f, and pγ < pδ iff δ < γ

Applying the same clauses as before for successor and limit superscripts, we are again guaranteed the existence of a ﬁxed point. The minimal such ﬁxed point is our model < D, Iγ >, of Lγ . A few observations regarding the formal theory are in order (proofs have been omitted in order to keep this chapter concise and accessible): [1] In any language Lα where α ≥ 1, T(< >) and are interchangeable salve valutate (i.e. the truth predicate is ‘transparent’, that is, T(< >) and always receive the same semantic value in < D, Iα >). [2] For any languages Lα and Lα+β , and any sentence in Lα (and thus in Lα+β ), Iα () = Iα+β (). In other words, once a sentence is expressible in one of our languages, and it receives a genuine truth-value (i.e. one other than n) in the model of that language, its truth-value does not change in models of extensions of that language.¹⁶ ¹⁵ In addition to guaranteeing the existence of a ﬁxed point, the monotonicity of our valuation scheme relative to this ordering guarantees that the maximum referred to in this clause exists. ¹⁶ A critic might be tempted to object to the present view on the ground that, as we pass from one language to another, the meaning of the truth predicate changes (which might, in turn, push us towards a view more like that sketched in Williamson (1998)). Fact [2] goes some ways towards assuaging such worries. The extension of the truth predicate does change from one language to the next, but only by adding new instances—no sentence can ‘change’ from true to some other value. Such changes in extension do not imply a change in meaning, however, any more than the production of the present volume changed the meaning of ‘published anthology’.

Embracing Revenge / 43 [3] No model < D, Iα > makes true all instances of the T-schema: T(< >) ↔ on any deﬁnition of the biconditional.¹⁷ Given [1], however, this is clearly not due to a fault in the truth predicate, but a failure to express a suitable conditional (and thus biconditional). The seriousness of this omission is lessened by the fact that each of languages makes true all instances of the T-schema for sentences of earlier languages. In other words, letting: ↔α

=df

( →α ) ∧ ( →α )

Lα+1 makes true all instances of the T-schema: T(< >) ↔α where is a sentence of Lα . [4] Given [3], it is not surprising that we can express, within Lα+1 , the complete semantic theory for Lα .¹⁸ For example, letting ‘Sentα (x)’ abbreviate the unary arithmetic predicate that holds of n if and only if n is the G¨odel number of a sentence in Lα , and letting ‘Conj(x, y, z) abbreviate the ternary arithmetic predicate expressing ‘x is the code of the conjunction of the sentences that y and z code’, we can, within L3 , formulate the clause for conjunction in L2 as follows. First, we need to deﬁne slightly altered versions of our truth and pathological predicates: T(α) (< >) F(α) (< >) P(α,β) (< >)

=df =df =df

T(< >) ↔α (T(< >) ↔α T(< >))¹⁹ F(< >) ↔α (T(< >) ↔α T(< >)) Pβ (< >) ↔α (T(< >) ↔α T(< >))

Intuitively, P(α,β) () gets the value t if gets the α th pathological value, gets the value pδ if δ > α and gets the value pδ , gets the value n if get the value n, and gets f otherwise (similarly for T(α) (< >) and F(α) (< >)). We can then formulate the clause for conjunction as: (∀x)(∀y)(∀z)((Sent2 (x) ∧ Conj(x, y, z)) →2 ((T(2) (x) ↔2 (T(2) (y) ∧ T(2) (z)))∧ (P(2,2) (x) ↔2 (P(2,2) (y) ∨ P(2,2) (z)))∧ ¹⁷ This is because none of our conditionals validates every instance of ( → ). For any sentence in Lα , however, ( →α+1 ) is a theorem (in Lα+1 ). ¹⁸ In order to describe the semantics for languages Lα , where α is inﬁnite, we make us of the fact that Lα+1 contains a satisfaction predicate Sat(x, y) expressing the relation ‘x is the code of a predicate satisﬁed by y’ and a predicate Pth (x,, y) expressing the relation ‘x is the code of the yth pathological predicate’. ¹⁹ T(α) is the α-level ‘strong’ truth predicate, which, when applied to a sentence , receives true if is true, receives the α + 1th pathological value if does, and is false otherwise. Notice that these predicates, unlike the ofﬁcial truth predicate, are not transparent. If is the simple Liar (and so receives the ﬁrst pathological value), then T(< >) also recieves the ﬁrst pathological value, while T(α+1) (< >) is false.

44 / Roy T. Cook (P(2,1) (x) ↔2 ((P(2,1) (y) ∨ P(2,1) (z)) ∧ (¬P(2,2) (y) ∧ ¬P(2,2) (z))))∧ (F(2) (x) ↔2 ((F(2) (y) ∧ F(2) (z)) ∧ (¬P(2,1) (y) ∧ ¬P(2,1) (z) ∧ ¬P(2,2) (y) ∧ ¬P(2,2) (z))))))

While the formal theory, as sketched, accurately models the philosophical picture as described in previous sections, there are a number of ways in which we could modify the details. Two are worth mentioning here. The ﬁrst way is to take, instead of the minimal ﬁxed point, the maximal intrinsic ﬁxed point (see Kripke (1972)). While studying the properties of various sorts of ﬁxed points in this iterated version of Kripke’s construction would no doubt provide us with a better understanding of the general framework as a whole, none of [1] through [4] above depend taking the minimal ﬁxed point (we only need the weaker claim that we have taken some ﬁxed point). So the choice between the minimal ﬁxed point and the maximal intrinsic ﬁxed point (or between these and some other ﬁxed point) will depend on one’s attitude towards ungrounded but determinate sentences such as: D:

D is either true or false.

The second way in which we might alter the present account is by adding more than one pathological truth-value at each extension of the language. Motivation for this idea is not hard to ﬁnd. Recall that in the ﬁrst section we saw that the Liar required a third truth-value because it fell into the category of sentences where ‘what is says is the case if and only if what it says is not the case.’ Given that truth and falsity are no longer being treated as exhaustive, this status is distinct from truth and falsity themselves. But it is not obvious that the status of the Liar, looked at in this way, is the same as that of other problematic ungrounded sentences. For example, the Truth Teller ‘This sentence is true’ falls into the ‘what it says is the case if and only if what it says is the case’ category, and the determinate sentence D above falls into the ‘what it says is the case if and only if either what it says is the case or what it says is not the case’. Given that truth and falsity are not exhaustive (and thus, we cannot assume that, for every sentence, either what it says is the case, or what it says is not the case), it is not obvious that these two categories are identical to the status of the Liar sentence.²⁰ Of course, the last paragraph is a bit rough and loose. Nevertheless, it does indicate that there might be good reasons for exploring the idea that extending our language by adding a new semantic predicate might introduce more than one new truth-value. Although adding such additional truth-values at each stage in the above construction ²⁰ In future work I intend on ﬂeshing out this idea by treating the different truth-values as intimately connected to (and thus, within the semantics, representable by) the directed graphs associated with the different patterns of referential dependency exhibited by these different ‘sorts’ of pathological sentence. For an initial stab in this direction, one can consult Cook (2004).

Embracing Revenge / 45 greatly complicates an already complicated picture, the framework could be extended in this manner without greatly affecting the ﬁnal shape of the account (in particular, versions of [1] through [4] will still hold). At this point, however, we will rest content with merely having pointed out the possibility of such extensions of the basic picture.

2.4 Dodging Revenge On the account just sketched, the Revenge Problem is not a problem, but instead provides the crucial insight motivating the account: Given a language L, if we can completely describe the semantics of L, then we have (knowingly or not) extended our language to a new language L’ (where such an extension involves not only adding to the set of wffs, but adding an additional truth-value that those wffs can receive). The semantic theory of L, however, as expressed in L’ is not sufﬁent for L’ itself. In order to describe its semantics, we must extend the language again (and as a result extend the collection of truth-values as well). And so on. In hacker-speak, the Revenge Problem is no longer a bug—it is now a feature, exemplifying the indeﬁnite extensibility of the concepts language, statement, and truth-value. One advantage of this view is that there are no real limitations on what can be expressed. At any stage in the series of languages, we are free to extend the language we presently speak in order to describe all the facts (semantic and otherwise), although in doing so we might introduce additional truth-values (and thus enable ourselves to ‘access’ more facts). But further extensions will allow us to speak of those as well, and so on. Of course, given this lack of expressive restriction, a critic might be forgiven for thinking that Revenge is likely to reappear. After all, if we can, at any stage in the game, extend our semantic resources in order to describe all of the (currently accessible) facts, then what is to stop us from extending the language so as to contain all possible semantic predicates at once?²¹ In fact, doesn’t the language used in the previous section, in describing the formal theory, amount to a language that does just that? But once we have allowed ourselves to extend the language in this way, there is nothing to stop us from forming the Super-Strengthened Liar: SUP: This sentence is either false or has one of the pathological values. ²¹ The comments of the next few paragraphs also explain why we cannot at any point add operators such as ‘exclusion’ negation to our language (i.e. a connective ∗ such that ∗ (P) is false if P is true, and true otherwise). Thanks go to Stewart Shapiro for pointing this out.

46 / Roy T. Cook It is not difﬁcult to see that such a sentence would be inconsistent in our formal theory, were it expressible. Fortunately for our account, however, there are good reasons for thinking that, contrary to appearances, SUP is not expressible in any language.²² Actually, there are two separate claims involved in this global version of the Revenge Problem which need to be distinguished, so each can be dealt with in turn. The ﬁrst is that there is nothing to stop us from extending our language so that we can talk about all possible truth-values at once. The second claim is that we have already done this, in the previous section when formulating our formal account of the semantics. Regarding the ﬁrst claim, the initial answer is easy: there are reasons why we cannot formulate a language that contains every possible truth-value. Anytime we attempt to add a predicate ‘is a pathological truth-value’ to a language Lα , we end up extending the language, forming a new language Lα+β . The semantics of Lα+β will require at least one more truth-value than that of Lα , and our new predicate will only be satisﬁed (i.e. receive the value t) by (G¨odel codes of) sentences which receive values p1 , p2 , . . . pδ for some δ < α + β. This is the very lesson that the Revenge Problem teaches us: That any attempt to construct a language that allows us to talk about ‘all’ semantic values (in the sense of containing predicates for each value) brings a new pathological value into the picture, one which is not described by the language in question. In other words, the indeﬁnite extensibility of the concept statement prevents us from every being able to say things about all statements (or, derivatively, about all truth-values). At this point, however, we run into the second aspect of the objection. How can we claim that we can never talk about all truth-values at once, so the criticism goes, when we obviously quantiﬁed over all of them in the formal account given in the previous section? The easy answer to this question is that it misrepresents what exactly the formal model is doing. In particular, this objection confuses describing a language and using that very same language. The formal semantics presented in §2.3 is a description (i.e. a model, in the intuitive sense of model) of a sequence of possible language extensions. No semantic predicates are used in describing this mathematical structure—the account is (or, can be reformulated) within ﬁrst-order set theory. As a result, the formal model (can) occur in our (actual) base language corresponding to L0 (and, in fact, this is the proper place ²² Actually, this is a bit of a sloppy way of putting it, since we have expressed SUP, in English, just a few lines previously. More precisely, SUP is expressible (we just expressed it), but it does not say what we think it does. There are intricate questions looming here, connected to whether or not we are, in any sense, ‘allowed’ to extend our language in inconsistent ways. If, however, we assume that such a predicate as that used in SUP is added to our language, and that the language remains consistent, then the predicate does not express what we (in some sense) intended it to express. The situation is exactly analogous to the fact that we cannot add the term ‘the set of all sets’ to our language and expect it to consistently have its ‘intended’ meaning.

Embracing Revenge / 47 for such theorizing). Thus, the account of the formal semantics does not occur within a language that uses all of the semantic notions which it describes as occurring in the hierarchy.²³ Nevertheless, the critic might continue, if our base language L0 contains the semantic theory as described in the previous section, could we not extend the language by adding the predicate: is a sentence which receives one of the pathological values described in L0 . Furthermore, once we have added such a predicate to our language, what is to stop us from using this predicate to formulate a version of the Super Strengthened Liar? The answer to this two-part question is of course a two-part answer. First, and most easily, we can grant to the critic that he is free to add such a predicate to his language (thus in a sense ‘skipping’ the step by step individual extensions of the language as described in the formal theory). The answer to the second question, however, is more complicated, and requires our getting a bit clearer on what we mean by the phrase ‘pathological value described in L0 ’. Remember that L0 (plus its interpretation) contains a set theory at least as strong as ﬁrst-order ZFC. In addition, it describes a series of languages Lα for every ordinal α. Thus, there will be a least ordinal, call it π, such that there is a model of L0 with π set theoretic ranks, and every model of π has at least π ranks. As a result, the theory of L0 only guarantees the existence of all ordinals less than π (although it will have models with more ordinals as well). As a result, L0 will only guarantee the existence of π distinct languages, with a corresponding collection of π distinct truth-values. Since a theory only ‘describes’ those objects that are guaranteed to exist according to that theory (or, at least, we assuming as much here), we should interpret the phrase ‘pathological value described in L0 ’ as being satisﬁed only by pathological values pβ where β < π. As a result, adding the suggested predicate to the language extends the language to a sub-language of Lπ which requires the same truth-values for its interpretation as Lπ itself (the language in question will be a sub-theory of Lπ since the language we obtain—a natural language—is countable while Lπ is not). Lπ (and the new semantic value pπ ) are not described in our (present) base theory.²⁴ ²³ Something similar to this approach is found in Field (2003). Although our accounts differ signiﬁcantly, I owe much to the careful study of this chapter. ²⁴ Actually, it is this very phenomenon that justiﬁes our carrying out the construction of the previous section into the transﬁnite. We did not do so because it is likely, or even possible, that we might some day master a language with an uncountably inﬁnite vocabulary such as those high up in the hierarchy. Rather, the thought is that we might, through tricks such as the one considered here, extend our language in such a way as to require the same collection of truth-values as is required by one of the uncountable languages in the hierarchy (even if the actual language we are using remains countable).

48 / Roy T. Cook So, given a particular set theory in the base language L0 , we can extend our language past all of the languages described in the formal semantics developed in L0 by adding this predicate. This pushes us ‘up’ to a language beyond any described in our L0 semantics. But of course we wish to be able to provide a semantics for this language as well. The solution is to strengthen the set theory at the base level. If we add an additional axiom guaranteeing the existence of the ordinal π, then our new base theory will allow us to formulate a new account of the formal semantics which is identical to the original one other than the fact that it implies the existence of more languages (and more pathological values to accompany them). We can of course repeat the process, adding other semantic predicates, and obtaining even stronger languages as a result. In principle, there is no limit to such extensions (other than those imposed by our ﬁnite lifespans, etc.). Thus, in a sense we can never provide a complete account of all of possible extensions of our base language, since for any such account (formulated in a particular set theory T ), we can add a predicate: is a sentence which receives one of the pathological values described in T. which extends the language past what can be described by T. This is not a ﬂaw in our formal semantics, however, but instead reﬂects a well-known feature of set theory. Our formal semantics entails that there is a language Lα for each ordinal α. But for any consistent set theory T, there is a stronger consistent set theory T’ such that T’ implies the existence of more ordinals than does T. Since we can never formulate a set theory which implies the existence of all possible ordinals, we can never formulate a formal semantics for our account which implies the existence of all possible extensions of our language (and corresponding truth-values) in some absolute sense of the word ‘all’. While no single set theory implies the existence of all ordinals, however, there seems to be no reason to doubt that, for any ordinal, there is a set theory that implies its existence. As a result, for any possible extension of our language, we can formulate a semantics for it (by utilizing a suitably strong set theory in the base theory). Earlier we drew an analogy between the indeﬁnite extensibility of the concept ordinal and the indeﬁnite extensibility of the concept language (and the corresponding indeﬁnite extensibility of truth-value). The previous few paragraphs suggest, however, that there is more to this than just an analogy—in fact, the indeﬁnite extensibility of our language just is the indeﬁnite extensibility of the ordinals. This insight promises fruitful connections between the semantic and set-theoretic paradoxes.²⁵ ²⁵ One such connection involves the fact that certain very powerful versions of the Strengthened Liar, such as SUP above, seem (on the present account) to entail (at least indirectly) large cardinal axioms. These connections will be explored in future work.

Embracing Revenge / 49 Thus, the lesson to learn from the Revenge Problem is just this: What we can say (and the semantic values that what we say can receive) is indeﬁnitely extensible in exactly the way the ordinals are. This implies that there is no language in which we can say everything. It does not imply, however, that there is something (coherent) which we cannot say in any language.

2.5 Truth, the Conditional, and Field Something should be said, at this point, regarding the conditional, or, on the present view, the conditionals, plural. The formal theory presented in section 2.3 (and the informal account motivating it) provides, not a single conditional, but a new conditional for each extension of the language. The reason for this move, odd looking though it might be at ﬁrst glance, is the need to avoid Curry’s Paradox. The arithmetic version of Curry’s construction can be carried out as follows: Given an arbitrary sentence , we can diagonalize to obtain a sentence X where: X ↔ (T < X >→ ) If (a) our truth predicate is transparent (i.e. for any , and T < >are intersubstitutable in all non-opaque contexts), (b) our conditional satisﬁes the inference rule modus ponens, and (c) our conditional validates the contraction axiom ((A → (A → B)) → (A → B)), then we can prove : [1] [2] [3] [4] [5] [6] [7]

X → (T < X >→ ) X → (X → ) (X → (X → )) → (X → ) (X → ) (T < X >) → ) X

Diagonalization [1], Transparency of Truth Contraction [2], [3], Modus Ponens [4], Transparency of Truth [5], Diagonalization [4], [6] Modus Ponens

Note that if we deﬁne negation in terms of the conditional and a primitive absurd sentence ⊥: ¬ =df → ⊥ then the Liar paradox is merely a special case of Curry’s paradox.²⁶ ²⁶ The fact that the Liar Paradox can be seen as a special instance of Curry’s Paradox, and thus in some sense the problems with semantic paradox in general hinge crucially on the conditional, is vastly underappreciated. Greg Restall’s contribution to this volume, however, is a notable exception to this.

50 / Roy T. Cook Thus, the real problem highlighted by the semantic paradoxes is that we cannot (on pain of triviality) have both: (a) (b)

A transparent truth predicate. A conditional that uniformly satisﬁes the standard axioms and rules of inference for the conditional.

Since a transparent truth predicate is one of the main motivations for the present account, it follows that we cannot have a single conditional uniformly satisfying the rules usually attributed to the conditional. Denying the validity of basic principles such as contraction and modus ponens is a bitter pill to swallow, however, so the obvious move is to give up as ‘little’ of the standard rules for the conditional as possible. There seem to be two ways in which this can be accomplished. The ﬁrst way is to attempt to provide a single conditional that comes as close as possible to the standard classical account. Hartry Field’s important (2003) (and its successors, including his contribution to this volume) represents such an approach. Field’s theory contains a transparent truth predicate, and in addition it contains a conditional that is well behaved when applied to non-pathological sentences. The main drawback with this approach, however, is that the conditional cannot satisfy all instances of the standard axioms and rules for the conditional.²⁷ In particular, contraction can fail when the subformulas involved fail to receive classical values (as happens with the instance of contraction relevant to Curry’s paradox). On the other hand, however, we can give up on the idea of a single conditional, and instead accept that the concept conditional is itself indeﬁnitely extensible. This is the approach taken above. In each extension of our language, we obtain (or, at least, we can obtain) a new conditional that is a better approximation to our intuitive ideas regarding ‘if. . .then. . .’. Each conditional satisﬁes all the axioms and rules we would expect, at least when applied to sentences of earlier languages. In particular, every instance of both modus ponens and contraction are validated. This last claim, of course, needs some explanation, given that if we have both modus ponens and contraction without any restrictions whatsoever, then we can (as we have already seen) reconstruct Curry’s paradox. The point is this: all instances of the inference rule modus ponens are valid, since (assuming we deﬁne validity in the standard way as truth-preservation in models) no matter which instance of the conditional one takes, if it is true, and its antecedent is true, it follows that its consequent must be true. In addition, given any sentences A and B, there will be some ²⁷ This is not to say that his account does not do an admirable job, from a technical perspective, of validating as many of the standard rules as is possible.

Embracing Revenge / 51 ordinal α such that ((A → (A → B)) → (A → B)) is a theorem.²⁸ Curry’s paradox is blocked, however, since the conditional →α validating the relevant instance of contraction occurs later in the hierarchy than does the conditional occurring in the Curry sentence we obtain through diagonalization. This would seem to be the main advantage of the present view over Field’s approach, at least if one of our motivations (secondary, perhaps, to a transparent truth predicate) is to retain as much of the traditional account of the conditional as is possible. On the present account we need not give up any instance of any standard rule or axiom for the conditional (we just need to remember that not all instances of all rules or axioms are valid for all conditionals!).²⁹ Despite this difference, both views (Field’s and the present one) represent instances of a family of views which take as their primary motivation the salvation of a single, uniﬁed truth predicate that is transparent (at least in non-opaque contexts). Both views accept the fact that an account achieving this (and containing some ²⁸ Proof sketch: Let α be the least ordinal such that A and B both occur in Lα . Consider the Lα+1 sentence: ((A →α (A →α B)) → (A →α B)) Since A and B receive, as truth-values, either t, f, or pβ for some β ≤ α, the relevant portion of the satisfaction clause for this conditional is: I( →α )

=

t f

I() ≤ I() relative to the ordering : f < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < t otherwise.

Assume, for reductio, that this formula fails to be true, that is: I(A →α (A →α B)) > I(A →α B) Then: I(A →α (A →α B)) = t And: I(A →α B) = f The former implies that: I(A) ≤ I(A →α B) And the latter implies: I(B) < I(A) Since partial orders are transitive, we obtain: I(B) < I(A →α B) = f But this is impossible, since f is the minimal element of the ordering. ²⁹ In the end, however, the choice of one of these accounts over the other, or of some third view over both, should depend, not on technical merits, but on philosophical motivation.

52 / Roy T. Cook sort of reasonable conditional) will require stratiﬁcation of some sort (this is what Tarski had right), at least if we wish out ﬁnal account to be able, in some sense, to ‘completely’ characterize all semantically problematic sentences. Field chooses to ﬁnd the stratiﬁcation in a hierarchy of stronger and stronger ‘deﬁnite’ truth predicates, while on the present view the same role is played by a hierarchy of stronger and stronger conditionals (plus the various semantic predicates that accompany them). Nevertheless, despite the vastly different details, the general philosophical viewpoint seems roughly the same.³⁰

References Cook, R. (2004). ‘Patterns of Paradox’, Journal of Symbolic Logic 69: 767–74 Dummett, M. (1993). The Seas of Language, Oxford,: Clarendon Press Field, H. (2003). ‘A revenge-immune solution to the semantic paradoxes’, Journal of Philosophical Logic 32: 132–177 Fitting, M. (1986). ‘Notes on the mathematical aspects of Kripke’s theory of truth’, Notre Dame Journal of Formal Logic 27: 75–88 Godel, K., (1992). On Formally Undecidable Propositions. New York,: Dover Kleene, S. (1952). Introduction to Metamathematics, Amsterdam,: North Holland Kripke, S. (1975). ‘Outline of a theory of truth’, Journal of Philosophy 72 (1975), 690-716; reprinted in Robert L. Martin (ed.), Recent Essays on Truth and the Liar Paradox, Oxford, Clarendon Press (1984), pp. 53–81 Russell, B. (1906). ‘On some difﬁculties in the theory of transﬁnite numbers and order types’, Proceedings of the London Mathematical Society 4, 29–53 Shapiro, S., and Wright, C. (2006). ‘All things indeﬁnitely extensible’, in A. Rayo and G. Uzquiano (eds.), Absolute Generality, Oxford: Oxford University Press (2006) Tarski, A. (1933). ‘The concept of truth in the languages of the deductive sciences’, Prace Towarzystwa Naukowego Warszawskiego, Wydzial III Nauk Matematyczno-Fizycznych 34, Warsaw; English translation in Alfred Tarski, Logic, Semantics, Metamathematics, Papers from 1923 to 1938, John Corcoran (ed.), Indianapolis,: Hackett Publishing Company, (1983), pp. 152–278 Williamson, T. (1998). ‘Indeﬁnite extensibility’, Grazer Philosophische Studien 55: 1–24 ³⁰ Earlier versions of this chapter were presented at The Ohio State University and Arche: The AHRC Centre for the Philosophy of Logic, Language, Mathematics, and Mind, and the current version has beneﬁted much from the comments, criticisms, and congenial atmosphere found there. Thanks are also due to JC Beall, Stewart Shapiro, Timothy Williamson, and an anonymous referee for additional comments or guidance. A special debt is owed to the students in my Spring 2006 undergraduate Leibniz seminar at Villanova University, who were forced to listen to early versions of these ideas before and after class, and without whose prodding I might never have started, or ﬁnished, the chapter.

3 The Liar Paradox, Expressibility, Possible Languages Matti Eklund

3.1 Introduction Here is the liar paradox. We have a sentence, (L), which somehow says of itself that it is false. Suppose (L) is true. Then things are as (L) says they are. (For it would appear to be a mere platitude that if a sentence is true, then things are as the sentence says they are.) (L) says that (L) is false. So, (L) is false. Since the supposition that (L) is true leads to contradiction, we can assert that (L) is false. But since this is just what (L) says, (L) is then true. (For it would appear to be a mere platitude that if things are as a given sentence says they are, the sentence is true.) So (L) is true. So (L) is both true and false. Contradiction. In the literature there is a bewildering variety of purported solutions to the liar paradox. I will not discuss any of these purported solutions in any detail. Instead I will further problematize the question of what a solution should achieve. I will bring up a somewhat neglected cluster of problems connected with the liar paradox. These problems remain even if one of the solutions to the liar paradox currently on offer would succeed perfectly in solving the problem it is designed to solve. They arise when we consider what possible languages there are.¹ Often in discussions of the liar paradox, truth ¹ (1) I will throughout conceive of languages as necessarily existing abstract objects. If this assumption should be rejected, my conclusions about this will just have to be reformulated as claims about what

54 / Matti Eklund predicates are treated as if they were applicable only to sentences (in contexts), and then only to sentences of one language. Clearly this is a simpliﬁcation. The English language predicate ‘true’ can be applied to sentences of all sorts of languages, to utterances of all sorts of languages, and, perhaps primarily, to propositions. As the arguments here will show, the simpliﬁcation is not innocent, but is an important distortion. I will approach the problem of possible languages via an issue that has attracted some attention in the liar literature: that of the liar paradox and expressibility. As is well known, a consequence of many of the most popular and most widely discussed solutions to the liar paradox is that some properties are deemed inexpressible in natural language. Theorists disagree about just which notions, if any, are inexpressible in natural language; how problematic or not such consequences are; and how best to conceive of the inexpressibility. In section 3.2, I will outline the main views on the liar paradox and expressibility. Sections 3.3 and 3.4 will be devoted to critical discussion of the most important of these views. Section 3.5 will focus on the issue of what possible languages there are. In section 3.6, I will discuss how serious revenge problems really are.

3.2 Inexpressibility and Revenge Problems Whatever is in the end the correct account of the liar paradox, the liar reasoning undoubtedly establishes what is sometimes called Tarski’s theorem: in no language whose logic is classical and which can talk about itself to a sufﬁcient extent can there be a predicate that satisﬁes the T-schema, where for a predicate to satisfy the T-schema is for a valid schema to result when this predicate is substituted for the ‘T’ in s is T iff p, (where instances of this schema are obtained by putting sentences for p and names of the corresponding sentences for s). A natural thought is that Tarski’s theorem entails that natural language is expressively limited. The reasoning would be that natural language satisﬁes the there could be. (2) There are independent problems concerning the notion of all languages. Focus, more narrowly on a problem concerning talk of all possible predicates. (If there is a problem concerning all possible predicates, there is a problem concerning all possible languages.) If, no matter what objects the xs are, there is a predicate true of exactly the xs, there are more possible predicates than there are objects—a contradiction, if predicates are objects. We must either think of quantiﬁcation over predicates and languages as restricted, or else take the space of languages and predicates as smaller. Serious though these problems are, I will set them aside as orthogonal to the problems that I will mainly focus on.

The Liar Paradox, Expressibility, Possible Languages / 55 conditions stated in Tarski’s theorem: so the property which satisﬁes the T-schema is not expressible in natural language, where a property satisﬁes the T-schema iff it is such that any predicate which expresses it satisﬁes the T-schema. (Notice that here and throughout I will use ‘expresses’ for the relation between predicates and the corresponding properties.) There are two ways to resist this reasoning. One is to deny that natural language does satisfy the conditions of Tarski’s theorem (the relevant type of sentence cannot be meaningfully formed or the logic of natural language is not classical).² The other is to deny that there is a property which satisﬁes the T-schema. The thought in the latter case would be that the property simply doesn’t exist. However, if there can be a richer metalanguage with a predicate which behaves in the right way semantically—speciﬁcally, which has the right semantic features to express the property—then it is hard to maintain that the property does not exist. (Here, and throughout, I will assume an ‘abundant’ ontology of properties.)³ Someone who takes Tarski’s theorem to show that there is a property which cannot be expressed in natural language might further hold that any predicate expressing the property of being true would have to satisfy the T-schema: so what would be established is that truth cannot be expressed in natural language.⁴ Sometimes, e.g. in Robert Martin’s (1984a), this reasoning is attributed to Tarski himself. This suggested conclusion is on the face of it bizarre. The claim is that truth cannot be expressed in natural language. But is truth not expressed by the predicate ‘true’, undoubtedly a natural language predicate? The threat that the liar reasoning entails that natural language is expressively limited comes up even if we deny that natural language satisﬁes the conditions of Tarski’s theorem, as in fact many theorists writing about the liar do. Most popular theories of the liar paradox seem to have the conclusion that certain properties—including properties which are expressed by predicates employed in these theories—cannot, on pain of consistency, be expressed in natural language but only in a richer metalanguage. Take ﬁrst theories of the liar which attempt to solve the paradox by saying that the liar sentence has a semantic status somehow intermediate between truth and falsity. The most famous theory of this kind is Saul Kripke’s (1975). Sometimes this intermediate status is conceived of as a third truth-value; sometimes, as in Kripke, it is rather conceived of as ‘unsettled’, or the absence of a truth-value. I will call sentences with this intermediate status neuter; and I will for simplicity talk about this ² There are non-trivial problems concerning going back and forth between what holds for formal theories and natural languages. ³ Below I will present a justiﬁcation of this choice and explain why it does not beg any important questions. ⁴ Or, focusing on language-speciﬁc truth predicates as the literature does, that truth in L cannot be expressed in L, for L a natural language.

56 / Matti Eklund as a truth-value, even if this begs otherwise important questions concerning how the intermediate status is best conceived. A liar sentence of the kind we started talking about can consistently be said to be neuter. But a purported solution of this general kind immediately invites the strengthened liar. Consider a new liar sentence which says of itself that it is untrue. A sentence that is neuter would appear to be, among other things, untrue. But then we are obviously back in paradox. There are different ways around this problem. But one popular way out is to say that in a three-valued setting, both ‘not’ and ‘true’ are ‘weak’: that they take neuter into neuter.⁵ If a sentence S is neuter, so is its negation and so is a sentence which says of S that it is true. If the negation sign and the object language truth predicate work this way, then a sentence which says of itself that it is not true is after all not paradoxical: it can consistently be ascribed the value neuter. Note it is not enough for paradox to be avoided that ‘not’ and ‘true’ be in this way weak: so long as there is any construction at all in the object language that takes both neuter and false into truth and takes truth into falsity—so long as strong negation, as it is often called, can be expressed at all—we land in paradox. The position must be that no construction in the object language can express that a sentence has some semantic status other than truth: as I will put it, that a sentence is untrue. Kripke also holds that untruth can be expressed only in a richer metalanguage. Second, consider the revision theory of truth, defended most prominently in Anil Gupta and Nuel Belnap (1993).⁶ The revison theorist retains classical logic but holds that sentences like the liar sentence cannot stably be assigned any truth-value. This theorist cannot, it seems, allow that ‘not stably true’ of the object language expresses what we would na¨ıvely take this predicate to express. For then consider a liar sentence that says of itself that it is not stably true. Like Kripke, Gupta and Belnap hold that some semantic notions needed in a semantic theory for a natural language can be expressed only in a richer metalanguage: that natural language is not semantically self-sufﬁcient. Gupta (1997) argues at some length that the ideal of avoiding this consequence may well be unattainable.⁷ In general, the situation is this. A standard kind of revenge problem for purported solutions to the liar paradox is the problem that given the expressive resources used to solve the solution to the liar in its simple form, a new paradox can be formed. The standard form of the revenge problem is this: the expressive resources of our language allow us to exhaustively and exclusively divide sentences into the true ones and the rest. If our language has sufﬁcient expressive resources to state ⁵ See e.g. Kripke (1975), Soames (1999), and Maudlin (2004). ⁶ Earlier presentations of the revision theory are Belnap (1982), Gupta (1982), and Herzberger (1982 and 1982a). ⁷ Gupta (1997), pp. 439 ff.

The Liar Paradox, Expressibility, Possible Languages / 57 an exhaustive and exclusive division of all sentences into the true ones and the rest, paradox can be reinstated. Just let our new liar sentence say of itself that it belongs to the rest.⁸ One common way to avoid a threatening revenge problem is to deny semantic self-sufﬁciency. The revenge problem arises only if it is assumed that the semantic theory can be formulated in the object language. It is only then that a new liar sentence can be formulated in the object language. But it is often regarded as an embarrassment for a theory of the liar paradox if it is forced to, so to speak, push some predicates into such a metalanguage. That a theory is so forced can often seem suspect already on intuitive grounds: it does not seem as if I expand my language when I use a construction expressing strong negation or when I use an expression expressing the property of not being stably true. There are also some arguments in the literature to the effect that any theory which is so forced simply must be unacceptable: natural language must be regarded as semantically self-sufﬁcient. One argument is suggested by Vann McGee (1994 and 1997): human language lies within the reach of human understanding and hence it must be possible to state a correct theory of human language in a human language.⁹ However, even if we agree on the assumption here, the conclusion that we must be able to give a semantic theory for a natural language L in L does not follow. Whereas the assumption might entail that there is some possible natural language where we can give a semantics for English, and generally that for every natural language there is some natural language where we can give a semantic theory for it, this does not mean that English possesses those resources. The assumption at most entails that for every natural language L there is some language we are capable of employing in which a semantics for L can be given. Graham Priest (1990) presents a different argument for semantic self-sufﬁciency. He says, ‘The whole point of solutions to the liar paradox (as opposed to reformist suggestions as to how to change our language) is to show that our semantic discourse (about truth, etc.) is, appearances notwithstanding, consistent. An attempt to show this which produces more and more discourse, not in its own scope, therefore fails.’¹⁰ This is not a convincing argument either. First, even if Priest is entirely right, one might think that the proper conclusion to draw might be that a solution of the kind described is unattainable. (Compare again Gupta’s stance.) Second, there would be a problem if the further discourse ‘produced’ could not be plausibly believed to be ⁸ Priest (1987), p. 29, makes this point well. ⁹ See McGee (1994), p. 628, and (1997), p. 402. f; compare too Gupta (1997), pp. 440. ff, who criticizes the argument. McGee comes closest to actually endorsing the argument discussed in the main text in his (1994); the discussion in his (1997) is considerably more guarded. ¹⁰ Priest (1990), p. 202.

58 / Matti Eklund consistent. But there is no immediate reason why the hierarchy of metalanguages that the Kripkean is saddled with should fail to be consistent.¹¹, ¹² In the above discussion I have, inter alia, suggested a number of different views on the implications of the liar paradox regarding the expressive limitations of natural languages. Let me now be more systematic: RADICAL INEXPRESSIBILITY (RI). Some seeming ordinary semantic properties, like truth, are in fact not expressible in ordinary natural languages, but only in a richer metalanguage. INEXPRESSIBILITY (I). Some semantic properties—including semantic properties we need to be able to express in an adequate semantic theory of natural language—are expressible only in a metalanguage. SEMANTIC SELF-SUFFICIENCY (SS). All semantic properties we need to be able to express in an adequate semantic theory of natural language are expressible already in natural language. WEAK UNIVERSALITY (WU). All properties are expressible in natural language; or at any rate: the liar paradox casts no doubt on this claim. (It is the qualiﬁcation that warrants the ‘weak’.)

(RI) entails (I) but not vice versa. (WU) entails (SS) but not vice versa. Martin ascribes view (RI) to Tarski. View (I) has the best claim to being orthodoxy: it is subscribed to by both Kripke (1975) and Gupta and Belnap (1993). McGee’s and Priest’s arguments are arguments for (SS). McGee is explicitly doubtful about (WU). Although Priest does not outright endorse (WU), some of the moves he makes are explicable only on the assumption that he holds this stronger view. For example, having shown to his satisfaction that Boolean negation (or, to take this in terms of properties, the property that a sentence has when its Boolean negation is true) need not be expressed in a semantic theory for natural language, Priest anyway sees it as incumbent upon himself to provide an argument for why Boolean negation just is not there to be expressed, and that it is not an expressive limitation of English that English does not possess the resources to express it. Hartry Field should also be mentioned along with McGee and Priest as someone defending a view of type (SS) or (WU). Field’s theory of truth is designed to meet the requirement of self-sufﬁciency, and Field takes this to be a supremely important consideration in its favor.¹³ ¹¹ It is of course unclear what it is for a ‘discourse’ or a ‘language’ to be consistent or inconsistent. But this unclarity is a problem for Priest, not for me. ¹² Similar remarks as apply to McGee and to Priest seem to me to apply to the argument for self-sufﬁciency given in Scharp (manuscript). Scharp gives an argument for why a theory of truth for a language needs to be internalizable: roughly, that even if it cannot be given in our actual natural language L there must be some expanded version of natural language, L+, where it can be given and such that the theory also applies to L+. The argument is that if this internalizability requirement is not met the theory cannot be both ‘descriptively correct’ and ‘descriptively complete’. Even if Scharp is right about this, the conclusion may well be that a semantic theory of the kind described simply cannot be had. ¹³ See Field (2003a), (2003b), (2005a), (2005b), (this volume).

The Liar Paradox, Expressibility, Possible Languages / 59 I will in my discussion focus primarily on (I) and (WU). Given that the arguments for semantic self-sufﬁciency are unpersuasive, I fail to see (SS) as a principled position. As for (RI), I happen to think that a brief dismissal of it as absurd is too quick—more on this at the very end of the Chapter—but I will anyway not focus much on it. The problems I will discuss regarding the more popular view (I) apply in a similar way to (RI). In section 3.3, I will discuss (I). In section 3.4, I will discuss the viability of (WU).

3.3 Universality I will approach (I) via consideration of the question of the universality of natural language.¹⁴ The notion of universality was introduced by Tarski: A characteristic feature of colloquial language . . . is its universality. It would not be in harmony with the spirit of this language if in some other language a word occurred which could not be translated into it; it could be claimed that ‘if we can speak meaningfully about anything at all, we can also speak about it in colloquial language’.¹⁵

What does ‘universality’ mean exactly? Instead of considering what Tarski and other theorists might have meant by it let me introduce my own—admittedly rough—characterization (which however seems to be in the spirit of Tarski): A language L is universal iff for every property there is a predicate of L that expresses it.¹⁶ A few remarks on this characterization are in order: First, for simplicity I talk only about the expressibility of properties. In a properly general characterization of the expressive power of natural languages we should consider not only properties, but also logical operations, objects, etc. Second, given the focus on properties, metaphysical questions concerning the nature and abundance of properties become relevant. On a conception of properties given which there are only very few properties (a sufﬁciently ‘sparse’ conception of properties, in common jargon) the claim of universality can be (relatively) trivial, and not the interesting claim it is intended to be. (Consider for instance a view on which the only properties there are, are properties corresponding ¹⁴ The classic discussion of universality is Tarski (1935/83); see also Fitch (1946 and 1964), Gupta (1997), Herzberger (1970), Martin (1976), McGee (1991 and 1997), Priest (1984 and 1987), and Simmons (1993). ¹⁵ Tarski (1935/83), p. 164. It is because Tarski says these things that (RI) cannot unproblematically be attributed to him. ¹⁶ The characterization of universality employs unrestricted quantiﬁcation over properties. Such quantiﬁcation is obviously problematic. Although the problems concerning the possibility of unrestricted quantiﬁcation over properties are not unrelated to the liar paradox, I will not here attempt to directly address those problems.

60 / Matti Eklund to the predicates employed in microphysics.) As already mentioned, I will throughout assume an abundant conception of properties, given which, roughly speaking, there are maximally many properties. A justiﬁcation for this strategy is that the talk of properties can be regarded as a convenient way of talking about what objects fall under a predicate. Third, there are questions about how to individuate languages. In logic, languages are individuated very ﬁnely: add another constant and you have a new language. Outside logic, however, natural languages are individuated more liberally: a language can undergo signiﬁcant changes in its vocabulary and its syntax and still we happily speak of it as the same language after these changes. I will here individuate natural languages ﬁnely. The justiﬁcation for this is straightforward. For ‘English is universal’ to express the substantive claim it is meant to express, it must not be sufﬁcient for its truth that for every property, English as it currently exists could be expanded with a predicate which expresses this property, while still we would call it the same language. Fourth, there are some questions regarding just how to understand ‘expressed’. These questions I will soon get to. Now, despite what unclarities may remain in the given characterization of universality, the characterization ought to be clear enough that it should seem intuitively implausible that all natural languages are universal. Consider, say, eighteenthcentury English. Do we really want to say that eighteenth-century English contained the resources for expressing the property of being a quark, or the property of being an inaccessible cardinal? Moreover, is it not perfectly possible that there are properties that simply are determinately outside our cognitive reach, in the way that (I would suppose) the property of being an inaccessible cardinal is outside the cognitive reach of a gorilla? But there is a complication. Stick with the example of eighteenth-century English. Intuitively, it is in some respects limited in expressive resources. It cannot express the property of being an inaccessible cardinal. So there is some property that can be expressed in some natural language that cannot be expressed in eighteenth-century English. But suppose the tallest man on Earth at midnight 1 July 2016 employs the predicate ‘is an inaccessible cardinal’. Then consider the predicate of eighteenthcentury English ‘falls under the predicate employed by the tallest man on Earth at midnight 1 July 2016’. This predicate is true of exactly the inaccessible cardinals: and so, arguably, expresses the property of being an inaccessible cardinal. Hence it appears that our verdict on eighteenth-century English must be revised: it is after all universal.¹⁷ There is an immediate objection to what has just been argued. The objection is that what is shown is at most that for every property φ there is a predicate of ¹⁷ Objection: did eighteenth-century English really contain linguistic vocabulary like ‘falls under’ and ‘predicate’? Reply: The details are irrelevant. What I take for granted is that eighteenth-century English contained some linguistic vocabulary that could be employed for the same purpose.

The Liar Paradox, Expressibility, Possible Languages / 61 eighteenth-century English which, given that the circumstances are propitious, as used on a particular occasion is true of exactly what has φ. But this is not enough for the predicate to express φ. The predicate ‘falls under the predicate employed by the tallest man on Earth at midnight 1 July 2016’ does not express untruth, even if the predicate the tallest man on Earth then used did express this property; rather, it expresses the property of falling under the predicate employed by the tallest man on Earth at midnight 1 July 2016. That is different. However, while there is a distinction here, it is actually for present purposes possible to disregard it. Let us say that a property φ is indirectly expressible in a language L iff there is a predicate F of L such that for some context c, an utterance of F by a speaker of L is such that for all x, ‘F(x)’ and ‘x has φ’ have the same truth-value. It would be wrong, for the reason just noted, to identify indirect expressibility as expressibility.¹⁸ But all that the present discussion needs is indirect expressibility of the relevant properties. The reason is that provided that given that a certain theory of truth demands that property not be expressible in English, it also demands that this property not be indirectly expressible in English. Let me take a few more examples to illustrate further what is at issue, before returning to the lessons of the liar paradox. Suppose, as earlier suggested, that there are properties beyond our cognitive reach: they are too complex, too fantastic, too divine, what have you. These too are properties we can express in our ordinary language, if what I have suggested with respect to the previous examples is correct. Consider predicates of the form ‘falls under the predicate employed at t by the most F creature in the universe’ (Where F might be intelligent, or old, or morally praiseworthy, or what have you.) Some of these predicates might be true of exactly what has one of these properties which supposedly are beyond our ken. It is unlikely, perhaps, that for every cognitively unreachable property (as we might call these properties) there will in fact be a predicate of our language which is in fact true of whatever has that property, as things stand. (Some cognitively unreachable properties need not be thought of at any time by any creature in the universe, etc.) But for all that, each of these properties may still be indirectly expressible. Compare too an objection Keith Simmons presses against taking natural languages to be universal: for every set in the ZF hierarchy there is a distinct concept (the concept of being a member of that set), and ‘[g]iven certain assumptions about natural languages (in particular, about upper limits on the size of vocabularies and the length of sentences), these concepts would outrun the expressive capacity of any natural ¹⁸ Compare the notion of ‘loose speaker expressibility’ characterized in Hofweber (2006): a property φ is loosely speaker expressible in L iff there is a predicate F of L and a context C such that that an utterance of F (by a speaker of L) in C expresses φ.

62 / Matti Eklund language’.¹⁹ This is a convincing argument against the possibility of the expressibility of all properties in natural language. However, Simmons’ reasoning leaves open that these concepts should all be indirectly expressible in natural language.²⁰ Return now to the liar paradox. The present considerations can appear to have the consequence that any theory of the liar paradox that relies on kicking some predicates up into the metalanguage can be easily refuted. Let U be the relevant metalanguage predicate. Then in the object language we will have predicates of the form ‘falls under the predicate U of the language spoken by so-and-so at time so-and-so’. Whatever is expressed by U will be indirectly expressible in the object language. But then we are back in paradox: let P, in a given context, indirectly express what U expresses, and consider a sentence which says of itself that it is P. (Call this kind of version of the liar paradox an interlanguage paradox.) However, it would be much too hasty to conclude from this reasoning that no view of type (I) can be maintained. For the reasoning relies on the assumption that for any predicate F, F and ‘falls under F’ are coextensive—as I will put it, that ‘falls under’ is transparent (generally, that satisfaction predicates like ‘falls under’, ‘is true of’, ‘satisﬁes’, etc. are transparent). While it is natural to hold that ‘falls under’ is transparent, liar-like reasoning—speciﬁcally, Grelling’s paradox²¹—calls this into doubt. More speciﬁcally, the assumption that ‘falls under’ is transparent is in the same category as the assumption that a sentence S and the corresponding sentence ‘S is true’ always have the same truth-value (that ‘true’ is transparent). Both assumptions are called into doubt by liar reasoning. But although there is no immediate refutation of views of type (I) in the ofﬁng, there is a nearby puzzle that deserves stressing. Although some take the liar reasoning to show that the truth and satisfaction predicates of English are not transparent, for example Kripke (1975) and others following up on his work have shown that there can be languages which are like natural languages whose truth and satisfaction predicates are transparent. One way the truth and satisfaction predicates of a language L can be transparent is if L lacks an untruth predicate. This is for instance the way that Kripkean theories of truth achieve transparency. The languages he describes ¹⁹ Simmons (1993), p. 15. Simmons talks about the expressibility of concepts; I prefer to think of the issue as about the expressibility of properties. ²⁰ Once the notion of indirect expressibility is introduced, one can introduce a corresponding notion of indirect universality: a language L is indirectly universal iff for every property there is a predicate of L that indirectly expresses it. What I have argued is that some immediate reasons for doubt about whether English is universal do not readily show that English fails to be indirectly universal. There may, for all I am concerned to argue, be reasons to doubt whether English really is indirectly universal. All I am concerned with is the indirect expressibility of those properties some theorists say for liar-related reasons can only be expressed in a metalanguage. ²¹ Grelling’s paradox: Let a predicate be heterological just in case it is not true of itself. Is ‘heterological’ true of itself or not?

The Liar Paradox, Expressibility, Possible Languages / 63 can contain transparent truth and satisfaction predicates on pain of not containing untruth predicates, as earlier discussed. If they did, we would be back in paradox, as is well known. But here we must be careful. Can there be a language L which contains transparent truth and satisfaction predicates on pain of not containing an untruth predicate?²² There is a difﬁculty here. There cannot be such a language L, if (i) there is some other language L∗ with an untruth predicate for languages of the kind to which L, if it exists, belongs (say, natural languages), and (ii) L’s truth and satisfaction predicates apply to sentences and predicates of languages other than L; speciﬁcally, of L∗ (if it exists). Now, if we think of L as something like a natural language, L would appear to have to satisfy (ii). It then comes down to condition (i). We are in a curious position. For L can exist if L∗ doesn’t and L∗ can exist if L doesn’t. Which of these two languages exists? I will return below to the signiﬁcance of this kind of puzzle.²³ For now let me just note the following. The puzzle does not have to do with what we should say about actual natural languages. The question is about a hypothetical natural language L, like actual ones, except for the possible difference that its truth and satisfaction predicates are transparent and on pain of consistency it does not contain an untruth predicate. One might well think that L, whether actually used or not, certainly exists. But now it turns out that L does not exist, if there is another language L∗ , with an untruth predicate for languages like L. The sort of puzzle I have called attention to does not refute view (I) on semantic self-sufﬁciency and universality. All it shows is that such a view can be hostage to facts about what possible languages there are.

3.4 Priest on Boolean Negation Turn now to views of type (WU): views according to which the liar paradox does not have the consequence that some notions are inexpressible. The discussion will be ²² Note the formulation. The question is not about the possibility of transparent truth and satisfaction predicates, simpliciter. It is about one particular way of achieving transparency. ²³ It deserves noting that while the puzzle presents problems for Kripkean theories of truth, it presents no problem for more old-fashioned, Tarskian theories, denying that truth in L can be expressed in L. Two points deserve noting. (i) A consistent Tarskian ought to say the same thing about ‘true of ’, ‘satisﬁes’, and ‘falls under’ as about ‘true’. (ii) A Tarskian obviously wouldn’t hold that L can express truth in other languages. (After all, truth in a natural language can for the Tarskian be expressed in a metalanguage. But a Tarskian will obviously want to hold that there cannot be pairs of languages L and L such that L can express truth in L and L can express truth in L. It should be clear that no interlanguage paradox can force the Tarskian to accept that truth and satisfaction in L can at least be indirectly expressed in L.

64 / Matti Eklund focused on Graham Priest’s defense of a view of this type. Priest is a dialetheist, and this informs his discussion. But even though dialetheism is a radical view, it should be clear that the points I make regarding Priest generalize. A dialetheist holds that there are true contradictions. The liar paradox presents one of the main arguments for dialetheism. The idea is that the solution to the liar paradox is to recognize true contradictions (the liar sentence is both true and not true). For the view that there are true contradictions to get off the ground, negation must not work in such a way that from a contradiction everything follows (it must not satisfy ex falso quodlibet); negation must, in other words, be paraconsistent. For otherwise the dialetheist would have to hold that all propositions are true. One of Priest’s main arguments for a dialetheist solution to the liar paradox is this. The liar reasoning forces us to make a choice between embracing expressive incompleteness (hence to adopt view (RI) or (I)) or embracing inconsistency. Nondialetheist theories of the liar commit their proponents to embracing the former alternative. But that is untenable. Hence the latter alternative, dialetheism, is forced upon us.²⁴ There are a few different lines of resistance to an argument like this. One is to deny that non-dialetheic theories have the consequence that natural languages are expressively incomplete. A second is to deny that this consequence is particularly damaging. A third, which I will focus on in this section, is to argue that the dialetheism does not in fact avoid expressive incompleteness. Such an argument can take different forms. It can be argued that the dialetheist cannot allow a predicate which expresses that a sentence is only untrue. For consider then a liar sentence which says of itself that it is only untrue. The standard dialetheist strategy, of saying that liar sentences are both true and untrue, does not seem to be applicable here. For if this sentence is both true and untrue, then it must be only untrue. But if it is only untrue, then it is true after all.²⁵ Another argument, which is the one I will focus on, since it is the one that Priest (1990) is about, is to the effect that the dialetheist cannot allow the expressibility of Boolean negation. As Priest himself states the problem, ‘If [Boolean negation] is allowed then, using the T-scheme and self-reference in the usual way, we can produce a sentence equivalent to its own Boolean negation, and hence deduce a Boolean contradiction, whence everything follows by Boolean EFQ [ex falso quodlibet]’.²⁶ Talk about the inexpressibility of Boolean negation is on its face different from talk about inexpressibility of properties, since a negation sign does not express a property. ²⁴ See e.g. Priest (1990), p. 202. ²⁵ The point of the argument (which has been presented by others, see e.g. Parsons (1990)) is just that dialetheism faces its own expressibility problem, prima facie as serious as the expressibility problems that arise for other theories. Naturally, dialetheists have responses to these problems; see e.g. Priest (1995). ²⁶ Priest (1990), p. 203.

The Liar Paradox, Expressibility, Possible Languages / 65 But I will slide over this difference, since the same remarks that apply to the question of the possibility of having a logical operator that expresses Boolean negation apply to the question of the possibility of having a predicate that expresses the corresponding property (the property a sentence has when its Boolean negation is true). Priest’s way out is to deny that there is such a thing as Boolean negation. It just is not there to be expressed. So the supposed fact that no expression of English expresses it indicates no expressive limitation in English. His argument for this perhaps surprising conclusion proceeds by way of criticizing possible attempts to show that there must be such a thing as Boolean negation. Let me brieﬂy summarize how Priest argues. ²⁷ Suppose ﬁrst that someone says that we can state rules of inference characterizing Boolean negation, and therefore Boolean negation exists. To this the reply is that we can state rules of inference for Prior’s ‘tonk’—the disjunction introduction rule and the conjunction elimination rule—and this does not mean that there is an associated operation expressed by the connective.²⁸ A more sophisticated version of the appeal to rules of inference would have it that the rules of inference for Boolean negation satisfy the condition for successful introduction of a connective while those for ‘tonk’ do not. But a standard condition is conservativeness: and a connective satisfying the rules of inference for Boolean negation will not satisfy this condition, if it is introduced into a dialetheist language with a truth predicate satisfying the T-schema (precisely because it makes the language trivial).²⁹ A different suggestion is to say that the rules of inference manage to characterize an operation if they are demonstrably sound. Concerning this, Priest argues that any argument to the effect that the relevant rules of inference are sound must itself employ Boolean negation, and so is question-begging.³⁰ As Priest is quite clear about, the force of his argument against the existence of Boolean negation relies on there being an independent case for thinking that dialetheism is true of English.³¹ Consider the point about how the rules of inference for Boolean negation fail to satisfy the conservativeness criterion. That point relies on assumptions about what is in the language prior to the introduction of Boolean negation. Priest argues that this fact about the dialectic is no cause for concern. ²⁷ My discussion of this argument will follow Priest (1990). The argument is also given in Priest (2005), ch. 5, which however focuses less on the liar paradox. Another type of argument for why some things we thought were expressible at least in some possible language do not in fact exist is that given in Maudlin (2004), passim. Maudlin goes from the assumption that ‘truth and falsity are always rooted in the state of the world’ to the claim that ‘if a sentence is true or false, then either it is a boundary sentence, made true or false by the world of non-semantic facts, or it is semantically connected to at least one boundary sentence, from which its truth value can be traced’, (p. 49), and from this Maudlin concludes that, e.g., there cannot be such a thing as strong negation. Not only is our negation not strong; in no language can negation behave that way. I cannot adequately discuss Maudlin’s stance here. Sufﬁce it to say that I agree with the observation of Gupta (2006) that Maudlin’s claim does not follow from the initial assumption. ²⁸ Priest (1990), p. 204. ²⁹ Ibid., pp. 204 f. ³⁰ Ibid., pp. 205–8. ³¹ Ibid., p. 209.

66 / Matti Eklund I have two remarks to make about Priest’s argument. First, note that if Priest’s argument for why a dialetheist need not appeal to inexpressibility is successful, then a non-dialetheist can make use of essentially the same argument, provided she does not need for the statability of her view that the property deemed inexpressible in the object language is expressible in the metalanguage. Field’s theory is arguably a non-dialetheist theory which meets this condition.³² Second, more importantly, if we widen our perspective to look at possible languages, I think we can see that, despite Priest’s claims to the contrary, there is a problem with the structure of Priest’s argument. Let me explain. Consider a hypothetical linguistic community such that the speakers of this community have implicitly made a semantic decision to use an expression to mean Boolean negation (where by their having made an ‘implicit semantic decision’ I simply mean that they have come to use, and come to conceive of, their negation sign as obeying the principles governing Boolean negation). Surely such a linguistic community is conceivable. And considering such a community begs no question against Priest. His argument is to the effect that Boolean negation does not exist. All that follows from this with respect to any actual or hypothetical linguistic community is that any attempt to express Boolean negation will fail. Insofar as we are convinced by Priest’s arguments that dialetheism is the correct story about our language—and suppose for argument’s sake that we are—then from our perspective, it seems that our negation is not Boolean. Given this, an argument along Priest-style lines can be given that Boolean negation does not exist. But from the perspective of this hypothetical community, it appears that their negation is Boolean and an argument analogous to that Priest gives can be given for taking something we take to exist (perhaps truth taken as satisfying the T-schema) not to exist. What is the truth of the matter? Here are some straightforward suggestions. (a) We are simply right and they are simply wrong (or, for that matter, they are simply right and we are simply wrong). (b) It is somehow objectively indeterminate what exists and what does not exist. (c) Both types of negation really do exist: it is just that they are expressible in different languages. (d) The underlying metaphysical assumptions ought to be questioned. ³² Field (this volume) argues that a fully general property of determinacy—such as otherwise would cause revenge problems for his theory—in fact does not exist. The reason, brieﬂy, is that not only does Field’s object language not contain the resources to deﬁne the relevant notion of determinacy: more generally, one cannot extrapolate from the resources of the object language to make intelligible such a notion of determinacy. In the main text I focus on Priest’s argument rather than Field’s because Priest’s argument seems more theoretically interesting. Why should Field’s reasoning make us the least inclined to conclude that the relevant notion does not exist? (Scharp (this volume) discusses this problem with Field’s argument at greater length. Compare too Yablo (2003), pp. 328 f.)

The Liar Paradox, Expressibility, Possible Languages / 67 The suggestions are all very problematic. Consider ﬁrst some problems with (a), which is the suggestion Priest would be forced to embrace. First, the situation seems clearly to be symmetric. Second, the consequence that whole linguistic communities might in this way be wrong about meaning is not easy to swallow. It is not hard to imagine cases where whole communities have mistaken views about what the expressions of their language mean. But a case like this would be different. Here a community would not just embrace the theory that, say, their negation sign expresses Boolean negation. Their whole use of this predicate would also point this way—they actually employ the relevant inference rules, etc.—and yet they could be wrong. Or that is what embracing (a) would have us accept. Third, setting these questions of knowledge and ignorance aside, one might think that it would be odd if meaning facts were hostage to metaphysics in the way suggested by this alternative: metaphysical facts about what propositions there are, not readily knowable to ordinary speakers, are held to determine what speakers mean. Let me elaborate on the second and third objections. We are, to be sure, well accustomed to the idea of semantic and psychological externalism, theses to the effect that the meanings of a speaker’s linguistic expressions and the contents of her mental states can be in part determined by factors external to the speaker. The arguments from Kripke, Hilary Putnam, and Tyler Burge are very familiar and taken to be persuasive.³³ Moreover, David Lewis has made popular the idea that some entities in the world are more natural, and hence more eligible to be meant and referred to, than are others.³⁴ This is another route to externalism. It might be thought that in light of these points, the third objection should not be very serious. But Kripkean and Putnamean appeals to the causal environment are beside the point when we are talking about expressions like ‘true’ and ‘not’. And the appeal to the social environment in Putnam’s ‘elm’/‘beech’ case and in Burge’s arthritis case is beside the point when we are talking about whole linguistic communities being wrong. The way in which the external world would matter to the meanings of speakers’ expressions under hypothesis (a) is most similar to the way that Lewis takes the world to matter. A Lewisian appeal to naturalness can certainly be made also with respect to expressions that purport to refer to abstract objects. But there is a crucial difference between the kind of dependence on the external world that obtains if Lewis is right and the dependence that (a) would saddle us with. For Lewis stresses that he does not claim that nothing could mean grue or quus. As he notes, such a claim would on the face of it be absurd. All he claims is that if the facts about a speaker’s dispositions with respect to her use of ‘+’ do not determine whether this sign as the speaker uses it means plus or quus, then the greater naturalness of the former decides in its favor. By contrast, what we are now asked to accept is that it would be absolutely impossible for anyone to mean Boolean negation by a sign. ³³ See Kripke (1980), Putnam (1975), and Burge (1979).

³⁴ Lewis (1983) and (1984).

68 / Matti Eklund Someone might defend (a) from the objections pressed by attempting to argue that our intuitions about logical matters not only tell us what we do and do not mean by various logical signs, but also about what operations there are and are not in logical reality. So for instance, in so far as our intuitions are best respected by a semantics for our language given which negation is paraconsistent instead of classical, this can be taken to tell us something not only about our language but about what logical operations there are, and hence what possible languages there can be. If something like this is right, then the objection to suggestion (a) that the envisaged situation is perfectly symmetric is completely beside the point. This appeal to what members of a hypothetical linguistic community would think and intuit is as misbegotten as an objection to reliance on what our senses deliver which appeals to what hypothetical creatures whose senses would deliver contrary things would be. It is of course correct that if intuition is viewed in the way here described, the objections I have presented to suggestion (a) are beside the point. But how plausible is this view on logical intuition? (Further remarks relevant to this issue follow in the next section.) So suggestion (a), which Priest is forced to embrace, faces serious problems. However, it is not evident that this means that Priest has been shown wrong. For the other suggestions likewise face serious problems. Turn ﬁrst to suggestion (b). This is appropriately fair when it comes to the question of who is right and who is wrong. But how make sense of the postulated indeterminacy? The indeterminacy cannot be semantic, given the way semantic indeterminacy is normally understood. For in order for an expression to be semantically indeterminate as between two different meanings (the way that, for instance, a vague expression is sometimes held to be semantically indeterminate as between different precisiﬁcations) the different meanings must exist. And the problem we are currently dealing with is that the respective meanings cannot possibly coexist. So the indeterminacy in question must be ontological. But ontological indeterminacy is widely regarded with suspicion: what does it even mean for the world to be in and of itself indeterminate?³⁵ As for (c), taking this route would mean abandoning Priest’s universality view for a view of type (I). Moreover, given that the dialetheist’s reasons for excluding Boolean negation apply equally in the case of a predicate expressing the property corresponding to Boolean negation, the concerns above stressed with respect to (I) would apply here too. These negative reﬂections may make us attracted to (d): we may think that it is the associated metaphysical picture that leads us wrong in the ﬁrst place. But rejecting the associated metaphysical picture—of there being these things, properties ³⁵ Notice too that here we are talking about a very special kind of purported ontological indeterminacy: indeterminacy in what logical operations there are. Postulating indeterminacy here is quite different from, say, suggesting that clouds and mountains are ontologically indeterminate.

The Liar Paradox, Expressibility, Possible Languages / 69 and operations, which our predicates and logical connectives express—does not get around the problem. For what Priest is fundamentally concerned with when arguing that there is no such thing as Boolean negation is that there can be no language with an expression that functions in thus-and-such a way. This question too may be metaphysical—it concerns what languages there can be—but if it is metaphysical, it is unavoidably so. It does not presuppose a potentially objectionable ontology of entities supposedly expressed by predicates and by logical operators. The situation is curious. Alternatives (a)–(d) appear to be exhaustive. But they all seem objectionable. This is a paradox. There is no clean objection here to what Priest says about Boolean negation, for although the view he is committed to, (a), seems bad, it is not clearly a worse view than the others. But Priest’s position is far from unproblematic.

3.5 Possible Languages I have discussed views (I) and (WU) on the liar paradox and expressibility, and somewhat tentatively presented some problems with both views. Let me now broaden the perspective a bit, before in the next section returning to the so-called revenge problem often raised in discussions of the liar paradox. Consider two different questions raised by the liar reasoning: The actual-language question: what is the correct semantic theory of ‘true’, ‘not’ and other key expressions in the liar reasoning? The possible-language question: what possible languages are there?

The actual-language question is more often discussed. The reason for the focus on this question is clear. The liar reasoning is carried out in ordinary natural language. It presents a puzzle about how ordinary language works. Hence it is only to be expected that those theorists who seek to solve the liar paradox are concerned to get natural language right. The formal theories that are developed are meant to be accurate models of natural language. But the liar paradox also has signiﬁcant implications for the possible-language question. For instance, Tarski’s theorem can be regarded as a theorem about what possible languages there are. In this section I will focus on the possible-language question. For one thing, this is very much an additional question, both difﬁcult and philosophically signiﬁcant, and it is not clear that it is solved even by an otherwise perfectly adequate solution to the liar paradox. For another, consideration of the possible-language question is of consequence for the actual-language question. What if the way I take English actually to be does not correspond to the way that any language could possibly be?

70 / Matti Eklund What possible languages are there? Many are happy to embrace a principle of plenitude for abstract objects, according to which (to put things intuitively) all the pure abstracta that coherently can exist also do exist. In the case of mathematical entities, the consequences of such a principle of plenitude are relatively straightforward. But in the case of languages, it is harder to see what the consequences of a principle of plenitude are.³⁶ Take a semantic theory according to which the semantics of English is somehow fuzzy-valued (degree-theoretic). Does this semantic theory describe a possible language at all? One kind of reason for doubt is this. Even if, given a principle of plenitude, there will of course be an abstract object some of whose constitutive parts or elements—the would-be sentences—are somehow mapped onto real numbers, the question is whether this is a possible language. It is a possible language only if these real numbers can adequately be regarded as, in some sense, truth-values (or, generally, semantic values of sentences). This it is possible to deny, consistently with ﬁrm adherence to a principle of plenitude. A classic discussion of this is Michael Dummett’s (1959), where it is argued that for general reasons having to do with the speech act of assertions, assertions can only be either correct or incorrect, and since truth and falsity corresponds to correctness and incorrectness in assertions, any appeal to ‘multivalence’—there being truth-values distinct from truth and falsity—must be denied. On what we may call a liberal view on what languages there are, there are languages corresponding to quite different classical and non-classical semantics. On a restrictive view, only languages corresponding to a very restricted class of semantics exist. Even if a very restrictive view is correct, there are further problems with respect to what possible languages there are. For it may be—as brought up above in connection with the interlanguage paradox—that whether a given, considered by itself seemingly perfectly possible, language really exists depends on what other languages there are. For future reference, let us say that there are two possible types of restrictions on what possible languages there are: Dummettian restrictions, having to do with e.g. what can intelligibly be said about the truth-values of sentences (I call them ‘Dummettian’ since Dummett is the theorist who has done the most to put them in the foreground—but I want to include under this general label also considerations Dummett himself would refuse to endorse) and liar-related restrictions, having to do with what liar-type reasoning entails. The obvious methodology for dealing with the actual-language question posed by the liar paradox is to compare proposed accounts of the paradox with intuitions we have. ‘Intuitions’ is used in many senses in the philosophical literature. But here it ³⁶ As stated in footnote 1, I here presuppose that languages are pure abstracta. If this is denied, the relevant question is instead about what languages there can be. The same considerations are still relevant.

The Liar Paradox, Expressibility, Possible Languages / 71 would appear that we are dealing with semantic intuitions: e.g. intuitions about the truth-values of various sentences containing ‘true’ and other semantic predicates. The correct answer to the actual-language question posed by the liar would standardly be assumed to be: that which best respects our semantic intuitions.³⁷ It is certainly possible that our semantic intuitions are inconsistent. In that case, our language is arguably the language that, so to speak, is the language ‘nearest’ our intuitions.³⁸ Attention to the possible-language question complicates this picture. First, even if our semantic intuitions were to be best respected by a theory that postulates for English a many-valued or fuzzy semantics it can be that theoretical considerations show that there are Dummettian restrictions on what languages there are, so that no language has such a semantics. Second, there are liar-related restrictions, such as illustrated by the interlanguage paradox. Can there be a language L, otherwise like a natural language, with transparent truth and satisfaction predicates?—Only if this language is not what we may call seriously limited: only if there is no predicate U of some possible language such that the property expressed by U cannot, on pain of paradox, be expressed in L. There can consistently be a language L of the kind described. But L cannot exist if there is another language as described. And such a language can also consistently exist. Focus now on liar-related restrictions. Suppose that, as earlier outlined, there are two different languages L and L∗ both of which can consistently exist but which are incompatible. How can we know which of L and L∗ exists? One need not be in general skeptical of the idea of reliable substantive metaphysical intuitions to feel the force of the problem here. For even if there are reliable substantive metaphysical intuitions, problems remain. First, such intuitions do not seem to be what we actually draw on when evaluating particular proposed solutions to the liar paradox: we seem rather to rely on ordinary semantic intuitions. Second, more to the point, when it comes to choices like that between a language like English except for the possible difference that its truth predicate is transparent and a language where untruth can be expressed, is it really plausible that we have any intuitions that speak to the issue of which language it is that exists? We can perhaps have strong intuitions that speak to the issue of which ³⁷ I have elsewhere argued that the liar paradox shows that natural language is inconsistent, in the sense that the paradox shows that principles such that it is part of semantic competence to accept them are jointly inconsistent (See especially my (2002).) What is mentioned in the text as a possibility is something considerably weaker. There is a distinction between on the one hand semantic intuitions, in the sense merely of intuitions we actually have about semantic matters, and on the other hand what competence requires us to accept. ³⁸ There are problems in spelling out what ‘nearness’ here amounts to. But there is no way of getting around this issue. It seems rather obvious that our semantic intuitions are inconsistent—that is why the liar paradox is a paradox—and so long as that does not mean that we have failed to endow our expressions with meaning we are forced into the ‘nearness’ talk.

72 / Matti Eklund of these languages it is that we speak, provided both exists, but that is a different matter. Compare the case of other abstract objects; say, mathematical entities. To make this case analogous with the language case, suppose that no plenitude principle, according to which some pure mathematical entities exist if they can consistently exist, is true. Arguably, we have intuitions that promise to speak to the issue of which mathematical entities then exist. If so, then presumably what we think is that the entities quantiﬁed over in relatively natural mathematical theories like ZF and PA exist but the entities quantiﬁed over in seemingly unnatural theories do not, where perhaps Quine’s NF is the prime example of an unnatural theory. Of course serious questions can be raised about the reliability of our intuitions about these matters. My point is just that in this case it can at least be reasonably urged that we have the relevant intuitions. By contrast, do we even have intuitions about what type of negation really exists? If there are sufﬁcient Dummettian restrictions, then maybe the liar-related restrictions do not present this sort of problem. For maybe there are not then two otherwise possible but incompatible languages. But in lieu of actual arguments to the effect that there are sufﬁcient Dummettian restrictions, this is only a pious hope.³⁹

3.6 Revenge Problems Let us now return to revenge problems. Such problems are often assumed to be lethal to the purported solutions that face them. Are they? This question is obviously closely tied to the questions that the discussion thus far has focused on, about universality and about whether certain properties simply fail to exist. There are two main ways to deal with purported revenge problems. One obvious way to deal with a revenge problem is to deny semantic self-sufﬁciency. This is the way of dealing with a revenge problem that corresponds to views (RI) and (I) on the ³⁹ Let me also stress an independent reason to be concerned with the interlanguage paradoxes. If indeed our semantic intuitions are inconsistent, as suggested above, and the correct semantic theory of (the relevant fragment of ) English is just the theory that does the best job of capturing these inconsistent intuitions, then the following possibility should easily suggest itself: it is simply semantically indeterminate what is the semantic value of ‘true’ (and of other key expressions in the liar reasoning and its variants). There are different semantic theories assigning different semantic values to ‘true’, such that nothing about whatever determines the semantic value of ‘true’ determines between these theories. The point is naturally put in terms of possible languages. Nothing determines, among these possible languages, which one it is that we speak. There is a range of languages such that, for all that is determined about what is our actual language, any one of them could be our natural language. Call these possible languages ‘candidate-languages’. The relevance of the interlanguage paradoxes is that they show that we cannot take for granted that these candidate-languages all exist; perhaps at most one does, in which case the semantic values of the relevant expressions can be determinate after all.

The Liar Paradox, Expressibility, Possible Languages / 73 expressive power of natural language. One problem with such a strategy is that it can seem that the relevant property clearly is expressible already in English. (Consider a view according to which truth is not expressible in English, or according to which it is not expressible in English that a sentence is unsettled or false.) Call this type of problem incredulity. A second problem is that of whether, for instance because of the reasoning of the interlanguage paradox discussed earlier, it just is not coherent to maintain that some property is expressible only in another language: for the relevant property may anyway be indirectly expressible, and that is sufﬁcient for problems to arise. Call this second problem the interlanguage problem. The interlanguage problem arises for some, but by no means all, views of this general kind. Given that general arguments for semantic self-sufﬁciency like those of McGee and Priest are ﬂawed, I do not see what other problem besides these two could threaten to arise generally for this type of way of dealing with the revenge problem. In fact, I suspect that even those discussions of revenge problems that on the face of it focus on self-sufﬁciency are motivated by what I called the incredulity problem: the problem is not ascent into a richer metalanguage per se but rather that the property said to be expressible only in this other language is one that we seem already to be able to express. Those who want to avoid having to kick up properties into the metalanguage because of the liar paradox—that is, those who do not accept views (RI) or (I)—seek to show that no revenge problems arise for their theories. Normally this is done by an attempt to demonstrate that the expressive resources employed in stating the solution can be allowed into the object language with no untoward consequences: that our language can be taken to be self-sufﬁcient. Suppose this is successfully done for some given theory of the liar. This does not necessarily avoid either the incredulity problem or the interlanguage problem. Take ﬁrst the incredulity problem. Even if none of the expressions employed in a particular account of the liar has to be conceived of as belonging to a richer metalanguage, the theory can still have the consequence that some property that intuitively seems expressible in English in fact is not expressible. So the incredulity problem can still arise.⁴⁰ What is more, even if the theory’s account of the truth and satisfaction predicates is consistent and even if no untoward consequences follow from allowing the expressive resources employed in stating the solution into the object language, the resources for an interlanguage liar paradox may still be available, so long as (i) there is some language with a predicate expressing a property which on pain of contradiction cannot be expressed by any predicate in the object language, and (ii) such a property is thereby indirectly expressible already in the object language. Let me close by making a few brief remarks on the incredulity problem. One might think that the incredulity problem is anyway the really hard one: that all otherwise ⁴⁰ See again the discussion of Field’s theory of truth in Scharp (this volume).

74 / Matti Eklund acceptable theories of the liar paradox end up deeming inexpressible in a given natural language properties that clearly are expressible in that language. There are two ways to respond to this suspicion. One is of course to attempt to devise a theory which does not deem any intuitively expressible properties inexpressible. A second, more original and theoretically more involved suggestion involves being more careful about expressibility. Consider the kind of view on the liar paradox I have myself elsewhere defended—an inconsistency view, I will here call it.⁴¹ According to this view, the liar paradox arises because meaning-constitutive principles for expressions of our language are inconsistent, where a meaning-constitutive principle is something it is part of semantic competence with the relevant expressions to be disposed to accept.⁴² I will not here attempt to defend this sort of view. But what I wish to stress is that this sort of view provides the materials for a promising response to the incredulity problem. On an inconsistency view one can say that the meaning-constitutive principles for the logical connectives are that they satisfy the inference rules characteristic of classical logic and a meaning-constitutive principle for the truth predicate is that it satisfy the T-schema. Tarski’s theorem then shows that the semantic values of the relevant expressions cannot possibly make true all the relevant meaning-constitutive principles. The semantic values of the relevant expressions are then what come closest to satisfying the relevant meaning-constitutive principles, possibly given other constraints. Suppose—I am only concerned to give a ‘model’ here—that the truth predicate fails to satisfy the T-schema. It is then fairly natural to say, absurd though it may sound, that truth fails to be expressible in English. There is no predicate of English that expresses what the truth predicate aims to express; no predicate that satisﬁes the meaning-constitutive principles for the truth predicate. And even if saying that this means that truth fails to be expressible in English is unnecessarily paradoxical, that does not affect the main point. The inconsistency view provides the materials for a way of getting around incredulity problems. We can respect the idea that truth obviously is expressible in English by noting that there is a predicate of English which aims to express it (or, to cash this out, which is governed by the right meaning-constitutive principles). We can respect the result that truth does not seem to be expressible in English by noting that (on the assumptions mentioned) there is no predicate of English whose semantic value satisﬁes the meaning-constitutive principles for ‘true’. Let an expression misﬁre iff its semantic value fails to satisfy the associated meaningconstitutive principles. The point is then that on an inconsistency view some expressions will misﬁre. Speciﬁcally, one can think that there can be properties such ⁴¹ See Eklund (2002). ⁴² There are many important matters concerning formulation that I here slide over. For instance, it is far from clear that this talk of dispositions to accept is what best describes what semantic competence involves.

The Liar Paradox, Expressibility, Possible Languages / 75 that any expression of a given language that aim to express this property will misﬁre. Such a property is in the most straightforward sense not expressible in the language in question. But while the property is inexpressible there may be predicates that aim to express it. Notice that if this is how the incredulity problem is dealt with, it is far from clear why a view of type (I) should be preferable to a view of type (RI). For the reason that a type (I) view would be preferable is that it is more palatable that a seemingly ‘technical’ property should be expressible only in a richer language than that an ‘ordinary’ property which ordinary speakers seem clearly able to express should be expressible only in a richer language. But given the route just outlined, a property deemed inexpressible is one that at the same time is claimed that we aim to express. And it is far clearer that we aim to express an ‘ordinary’ property such as that of being true than that we actually succeed in this aim.⁴³

References Barwise, Jon, and Etchemendy, John (1987). The Liar: An Essay on Truth and Circularity, Oxford University Press, Oxford Beall, JC (ed.) (2003). Liars and Heaps: New Essays on Paradox, Clarendon Press, Oxford Beall, JC, and Armour-Garb, Bradley (eds.) (2005). Deﬂationism and Paradox, Oxford University Press, Oxford Belnap, Nuel (1982). ‘Gupta’s rule of revision theory of truth’, Journal of Philosophical Logic 11: 103–16 Burge, Tyler (1979). ‘Individualism and the mental’. In Peter French, Theodore Uehling, and Howard Wettstein (eds.), Midwest Studies in Philosophy IV, University of Minnesota Press, Minneapolis, pp. 73–121 Eklund, Matti (2002). ‘Inconsistent languages’, Philosophy and Phenomenological Research 64: 251–75 Field, Hartry (2003a). ‘A revenge-immune solution to the semantic paradoxes’, Journal of Philosophical Logic 31: 1–27 (2003b). ‘The semantic paradoxes and the paradoxes of vagueness’, in Beall (2003) (2005a). ‘Is the liar sentence both true and false?’, in Beall and Armour-Garb (2005) ⁴³ Note that if the view sketched in the last few paragraphs is accepted, and view of type (RI) is acceptable, then many of the statements in this chapter will have to be taken with a ‘grain of salt’ (compare Frege (1892), p. 192): for some property that I aim to express will not in fact be expressible. But in contrast with other theorists who ask not to be begrudged a pinch of salt, I have a theoretical explanation of what is going on when my expressions do not in fact express what they are meant to express. An earlier version of this chapter was presented as a paper at a conference on paradoxes at the University of Latvia, November 2005. Thanks to the conference participants for good discussion. I also wish to thank Dan Korman, Agust´ın Rayo, and Kevin Scharp for helpful comments on earlier versions of this chapter.

76 / Matti Eklund Field, Hartry (2005b). ‘Variations on a theme by Yablo’. In Beall and Armour-Garb (2005) this volume, ‘Solving the paradoxes, escaping revenge’ Fitch, Frederic (1946). ‘Self-reference in philosophy’, Mind 55: 64–73 (1964). ‘Universal metalanguages for philosophy’, Review of Metaphysics 17: 396–402 Frege, Gottlob (1892). ‘On concept and object’. In Michael Beaney (ed.), The Frege Reader, Blackwell, Oxford (1997), pp. 181–93 Gupta, Anil (1982). ‘The liar paradox’, Journal of Philosophical Logic 11: 1–60. Reprinted in Martin (1984) (1997). ‘Deﬁnition and revision: a response to McGee and Martin’. In Villanueva (1997), pp. 419–43 (2006). ‘Truth and Paradox: Solving the Riddles, by Tim Maudlin’, Mind 115: 163–5 Gupta, Anil, and Belnap, Nuel (1993). The Revision Theory of Truth, MIT Press, Cambridge, Mass. Herzberger, Hans (1970). ‘Paradoxes of grounding in semantics’, Journal of Philosophy 67: 145–67 (1982). ‘Notes on naive semantics’, Journal of Philosophical Logic 11: 61–102. Reprinted in Martin (1984) (1982a). ‘Naive semantics and the liar paradox’, Journal of Philosophy 79: 479–97 Hofweber, Thomas (2006). ‘Inexpressible properties and propositions’. In Dean Zimmerman (ed.), Oxford Handbook of Metaphysics, vol. 2 Kripke, Saul (1975). ‘Outline of a theory of truth’, Journal of Philosophy 72: 690–716. Reprinted in Martin (1984) (1980). Naming and Necessity, Harvard University Press, Cambridge, Massa. Lewis, David (1983). ‘New work for a theory of universals’, Australasian Journal of Philosophy 61: 343–77 (1984). ‘Putnam’s paradox’, Australasian Journal of Philosophy 62: 221–36 Martin, Robert L (1976). ‘Are natural languages universal?’, Synth`ese 32: 271–91 (1984). Recent Essays on Truth and the Liar Paradox, Clarendon Press, Oxford (1984a). ‘Introduction’. In Martin (1984), pp. 1–8 Maudlin, Tim (2004). Truth and Paradox: Solving the Riddles, Clarendon Press, Oxford McGee, Vann (1991). Truth, Vagueness and Paradox, Hackett Publishing Company, Indianapolis (1994). ‘Afterword: truth and paradox’. In Robert M. Harnish (ed.), Basic Topics in the Philosophy of Language, Prentice Hall, Englewood Cliffs, NJ, pp. 615–33 (1997). ‘Revision’, in Villanueva (1997), pp. 387–406 Parsons, Terence (1990). ‘True contradictions’, Canadian Journal of Philosophy 20: 335–53 Priest, Graham (1984). ‘Semantic closure’, Studia Logica 43: 117–29 (1987). In Contradiction, Kluwer Academic Publishers, Dordrecht (1990). ‘Boolean negation and all that’, Journal of Philosophical Logic 19: 201–15 (1995). ‘Beyond gaps and gluts: reply to Parsons’, Canadian Journal of Philosophy 25: 57–66 (2005). Doubt Truth to Be a Liar, Oxford University Press, Oxford Putnam, Hilary (1975). ‘The meaning of ‘‘meaning’’ ’. In Keith Gunderson (ed.), Language, Mind, and Knowledge, University of Minnesota Press, Minneapolis, pp. 131–93 Scharp, Kevin: this volume, ‘Aletheic vengeance’ manuscript, ‘Truth and internalizability’

The Liar Paradox, Expressibility, Possible Languages / 77 Simmons, Keith (1993). Universality and the Liar, Cambridge University Press, Cambridge Soames, Scott (1999). Understanding Truth, Oxford University Press, Oxford Tarski, Alfred (1935/83). ‘The concept of truth in formalized languages’. In John Corcoran (ed.), Logic, Semantics, Metamathematics: Papers from 1923 to 1938, 2nd edn, Indianapolis: Hackett Publishing Company. English translation by J. H. Woodger of ‘Der Wahrheitsbegriff in Formalisierten Sprachen’, Studia Philosophica 1 (1935) Villanueva, Enrique (ed.) (1997). Truth, Ridgeview, Atascadero, Calif. Yablo, Stephen (2003). ‘New grounds for naive truth theory’. In Beall (2003), pp. 312–30

4 Solving the Paradoxes, Escaping Revenge Hartry Field

It is ‘the received wisdom’ that any intuitively natural and consistent resolution of a class of semantic paradoxes immediately leads to other paradoxes just as bad as the ﬁrst. This is often called the ‘revenge problem’. Some proponents of the received wisdom draw the conclusion that there is no hope of any natural treatment that puts all the paradoxes to rest: we must either live with the existence of paradoxes that we are unable to treat, or adopt artiﬁcial and ad hoc means to avoid them. Others (‘dialetheists’ ) argue that we can put the paradoxes to rest, but only by licensing the acceptance of some contradictions (presumably in a paraconsistent logic that prevents the contradictions from spreading everywhere).¹ I think the received wisdom is incorrect. In my effort to rebut it, I will focus on a certain type of solution to the paradoxes. This type of solution has the advantage of keeping the full Tarski truth schema (T) True(A ) ↔ A ¹ This latter view is only reasonable if ‘revenge’ is less of a worry for inconsistent solutions to the paradoxes than for consistent ones. I think myself that advocates of inconsistent solutions face a prima-facie revenge problem, and doubt that they can escape it without employing the devices I suggest in this chapter on behalf of certain consistent solutions. But that is a matter for another occasion.

Solving the Paradoxes, Escaping Revenge / 79 (and more generally, a full satisfaction schema). This has a price, namely that we must restrict both the law of excluded middle and the law connecting A → B with ¬A ∨ B, but we can carve the restrictions narrowly enough so that ordinary reasoning (e.g. in mathematics and physics) is unaffected.² I’ll call solutions of this type G-solutions. (If you want to think of the ‘G’ as standing for ‘good’ I won’t stop you.) The literature contains several demonstratively consistent solutions of this sort; for purposes of this chapter there is no need to choose between them. (I will give an informal introduction to this type of solution in sections 1, 3, and 4, and a formal account in section 5.) It will turn out that any such solution generates certain never-ending hierarchies of sentences that may seem ‘increasingly paradoxical’ (roughly speaking, it is harder to ﬁnd a theory that satisfactorily treats later members of the hierarchy than to ﬁnd one that satisfactorily treats earlier members); but the G-solution gives a consistent treatment of each member of each such hierarchy. The existence of these hierarchies prevents certain kinds of revenge problems from arising: certain attempts to state revenge problems simply involve going up a level in a hierarchy all levels of which have been given a non-paradoxical treatment. Still, there are certain ‘vindictive strategies’ (strategies for trying to ‘get revenge’ against solutions to the paradoxes) that G-solutions may seem to be subject to. I’ll argue that the most popular such strategy is based on a misunderstanding of the signiﬁcance of model-theoretic semantics. But there is a far more interesting strategy for which this is not so. As mentioned, a G-solution generates certain never-ending hierarchies of apparently paradoxical sentences which however are each successfully treated by the account. But shouldn’t it be possible to ‘break out of the hierarchies’ to get paradoxes that are not resolved by the account? Or to put it another way: If we can’t ‘break out of the hierarchies’ within the language that our solution to the paradoxes treats, isn’t that simply due to an expressive limitation in that language? That, I think, is the most difﬁcult revenge worry for G-solutions to deal with. Some of the worries about ‘breaking out of the hierarchies’ turn out to be intimately connected to K¨onig’s paradox of the least undeﬁnable ordinal, a paradox in the same ballpark as those of Berry and Richard. The G-solutions provide a consistent treatment of these deﬁnability paradoxes. The treatment of K¨onig’s paradox will be an important element in my argument that we are unable to ‘break outside the hierarchies’, but that this does not reﬂect an expressive limitation of the language. ² Also, the equivalence between A → B and ¬A ∨ B will hold on the assumption of excluded middle for A and for B.

80 / Hartry Field

Part One Introductory Discussion 1 The paradoxes and excluded middle Imagine that we speak a ﬁrst-order language L: it has the usual connectives and quantiﬁers, and it contains no ambiguous terms and no indexicals. It is to be a very rich language, powerful enough to express all our mathematics, including the richest set theory we currently know how to develop. It should be able to talk about its own expressions and their syntax; though we needn’t actually make this a separate requirement, since as G¨odel showed we can use arithmetical surrogates. If we like we can also assume that L can express all our current claims about the physical world too, though this will not really matter to the problem to be discussed. Finally, L should contain terms like ‘true’ and ‘true of’. For present purposes we needn’t worry about how such terms apply to sentences and formulas in languages other than our own, so we may as well assume that they have been restricted to apply only to the sentences and formulas of L. These assumptions about our language L are enough to generate paradoxes (or rather, apparent paradoxes). Some of them, like the Liar paradox, arise from the fact that by any of a number of well-known routes we can construct self-referential sentences: sentences that attribute to themselves any property you like. For instance, the Liar paradox arises from any sentence that directly or indirectly asserts its own untruth; let Q be some such sentence, and Q its standard name.³ Since Q asserts its own untruth, it certainly seems that Q ↔ ¬True(Q ) had better be part of our overall theory. In addition, it seems that our theory of truth ought to include every ‘Tarski biconditional’ , i.e. every instance of the schema (T) mentioned earlier; hence in particular, True(Q ) ↔ Q. ³ There is a familiar distinction between contingent and non-contingent Liar sentences. If the sentence ‘Nothing written on the ﬁrst blackboard manufactured in 2005 is true’ is written on a blackboard that (perhaps unbeknownst to the writer) was the ﬁrst to be manufactured in 2005, it is a contingent Liar: given the contingent facts about blackboard-manufacture it in effect asserts its own untruth. Non-contingent Liar sentences assert their own untruth independent of such empirical facts. Some of the formulations below are only strictly correct for non-contingent Liars, but could easily be generalized to apply to contingent Liars too. (For instance, in the next sentence of the text, replace ‘be part of our overall theory’ by ‘follow from our overall theory together with the empirical facts’.) The distinction between the two kinds of Liar sentences will make no important difference.

Solving the Paradoxes, Escaping Revenge / 81 But then if the conditional, and the biconditional deﬁned from it in the obvious way, are at all reasonable, we can infer (∗ )

True(Q ) ↔ ¬True(Q ).

And being of form B ↔ ¬B, this leads to contradiction in classical logic. There are similar paradoxes that don’t require the construction of self-referential sentences. For instance, just as our theory of truth ought to include the instances of Schema (T), so our theory of satisfaction ought to include the instances of the following schema: (S) For all x, x satisﬁesA(v) if and only if A(x) (where to say that x satisﬁes A(v) is the same as saying that A(v) is true of x).⁴ In the special case where A(v) is the formula ‘v does not satisfy v’, this yields that for all x, v does not satisfy v satisﬁes x if and only if x does not satisfy x, and hence (∗∗ )

v does not satisfy v satisﬁes itself if and only if it does not satisfy itself. ⁵

Again, (∗∗ ) is of form B ↔ ¬B and hence leads to contradiction in classical logic. The ‘G-solutions’ that I’ll be considering accept these derivations of (∗ ) and (∗∗ ). But unlike ‘dialetheic’ views (e.g. [16] ), they do not accept contradictions (sentences of form C ∧ ¬C). So they must reject all arguments that would take us (for arbitrary B) from B ↔ ¬B to a sentence of form C ∧ ¬C. I think that the most revealing way of trying to argue from B ↔ ¬B to a contradiction is as follows: (i) Assume both B ↔ ¬B and B. Then by modus ponens, ¬B; so B ∧ ¬B. (ii) Assume both B ↔ ¬B and ¬B. Then by modus ponens, B; so B ∧ ¬B. ⁴ (S) should really be called (S1 ): it is the schema for the satisfaction predicate ‘satisﬁes1 ’ that applies to formulas with exactly one free variable, and there is an analogous schema (Sn ) for each satisfaction predicate ‘satisﬁesn ’ that applies to formulas with exactly n free variables. But (S1 ) can be taken as basic: in any language rich enough to code ﬁnite sequences we can artiﬁcially deﬁne the higher satisfaction predicates in terms of 1-place satisfaction, in a way that guarantees the schemas for the former if we have (S1 ): e.g. to say that ‘v1 is larger than v2 ’ is satisﬁed by o1 and o2 in that order is in effect to say that ‘u is an ordered pair whose ﬁrst member is larger than its second’ is satisﬁed by o1 , o2 . We can similarly reduce truth to satisfaction: to say that ‘Snow is white’ is true is in effect to say that ‘Snow is white and u = u’ is satisﬁed by everything (or equivalently, by something). (T) then falls out of (S1 ), so (S1 ) can be taken as the sole basic schema. But it’s more natural to talk in terms of truth than satisfaction, so I’ll keep on talking about (T). ⁵ A natural abbreviation for ‘v satisﬁes itself’ would be ‘v is onanistic’. But for some reason ‘v is homological’ has caught on instead, with ‘heterological’ for ‘non-onanistic’.

82 / Hartry Field (iii) Since B ∧ ¬B follows both from the assumptions B ↔ ¬B and B and from the assumptions B ↔ ¬B and ¬B, then it follows from the assumptions B ↔ ¬B and B ∨ ¬B. (Reasoning by cases.) (iv) But B ∨ ¬B is a logical truth, so B ∧ ¬B follows from B ↔ ¬B alone. I now further stipulate that G-solutions accept both modus ponens and reasoning by cases (aka disjunction elimination). So they take the reasoning to be valid through step (iii). What G-solutions question is the use of the law of excluded middle in step (iv). Unlike intuitionists, though, G-theorists take excluded middle to be perfectly acceptable within standard mathematics, physics, and so forth; it is only certain reasoning using truth and related concepts that are affected.⁶ There is a verbal issue here about exactly how this point should be put. One way to put it is to say that excluded middle is literally valid in some contexts like mathematics, but invalid outside that domain. But it might be thought that the ‘topic neutrality’ of logic implies that if excluded middle can’t be accepted everywhere then it can’t be taken as literally valid anywhere. Even so, this doesn’t undermine the claim that it is effectively valid⁷ in contexts like mathematics: if one accepts all instances of the schema A ∨ ¬A that don’t contain ‘true’, then even if one doesn’t claim that they are logical truths one can reason from them just as a classical logician reasons in mathematics and physics. So it really makes no difference in which of the two ways we talk. Another argument from B ↔ ¬B to a contradiction runs as follows: after step (i) as above, we conclude that B ↔ ¬B entails ¬B by a reductio rule (that if X and B together entail ¬B, then X alone entails ¬B); that result and (ii) then give the contradiction. But the most obvious argument for that reductio rule is based on the law of excluded middle (together with reasoning by cases). The argument is that if X and B together entail ¬B, then since X and ¬B certainly entail ¬B, it follows that X and B ∨ ¬B entails ¬B; and since B ∨ ¬B is a logical truth, this means that X entails ¬B. So I will ⁶ Actually advocates of G-solutions might want to further restrict excluded middle, e.g. by disallowing its application to certain sentences containing vague concepts; and indeed it is not out of the question to regard certain mathematical concepts such as ‘ordinal’ as having a kind of ‘indeﬁnite extensibility’ that is akin to vagueness. Still, for purposes of this chapter I assume that excluded middle applies unrestrictedly within standard mathematics. Another plausible restriction of excluded middle is to sentences containing normative concepts like ‘appropriate’ or ‘reasonable’; this is relevant to certain ‘doxastic paradoxes’ involving, for instance, sentences asserting that it is not appropriate to believe them. But such paradoxes are outside the scope of this chapter. ⁷ ‘Effectively valid’ means ‘in effect valid’: it has nothing to do with effective procedures. Similarly I’ll use ‘effectively classical’ to mean ‘in effect classical’, i.e. excluded middle holds even if not as a logical law.

Solving the Paradoxes, Escaping Revenge / 83 assume that in giving up (or restricting) excluded middle we give up (or restrict) this reductio rule as well. Admittedly, this reductio rule is valid in intuitionist logic even in absence of excluded middle, so I can’t say that we are compelled to give up the reductio rule if we give up excluded middle. But intuitionist logic does not evade the paradoxes, so we had best not follow its lead.⁸ My point is that there is a natural response to the Liar paradox which sees this kind of reductio reasoning as depending on the law of excluded middle and both as needing restriction; and that is the response that G-solutions adopt.

2 Trying to preserve classical logic Weakening classical logic to deal with the paradoxes is obviously not something to be done lightly, and there are questions about how to understand the proposal, some of which I will address in the next section. But ﬁrst I’d like to brieﬂy survey the options for handling the paradoxes within classical logic. One reason for doing this is to make evident the appeal of the non-classical approach, and another is to facilitate a later discussion of the ‘hierarchies of paradoxical sentences’ that arise within G-solutions. In classical logic, the reasoning of the Liar paradox can easily be turned into a proof of the following disjunction: Either (i) Q is true, but ¬Q or (ii) Q is not true, but Q. At this point, classical theorists have three options. (Of course, there is also the possibility of remaining agnostic between the options, but that is of no particular interest.) The ﬁrst option is to choose disjunct (i). This would seem quite unattractive: doesn’t calling Q true while saying ‘nonetheless, ¬Q’ deprive the notion of truth of signiﬁcance? The second option is to choose disjunct (ii). This seems on its face almost equally unattractive: if one holds that Q is not true, what is one doing holding Q? ⁸ Intuitionists tend to motivate the reductio rule by way of the law ¬(A ∧ ¬A) (sometimes misleadingly called the ‘law of non-contradiction’ ). But to anyone who accepts the deMorgan law ¬(A ∧ B) |= ¬A ∨ ¬B, this version of the ‘law of non-contradiction’ simply amounts to ¬A ∨ ¬¬A, a slightly restricted version of excluded middle that few who reject excluded middle would accept. (That’s why dialetheists who accept excluded middle accept ¬(A ∧ ¬A), making clear that it does not adequately capture the principle that we should reject contradictions.) The intuitionist argument for reductio thus turns on their rejection of the deMorgan law.

84 / Hartry Field The third option is to accept the disjunction of (i) and (ii) while ruling out as absurd the acceptance of either disjunct. (It is because the acceptance of either disjunct is viewed as absurd that this is really a third option, distinct from agnosticism between the ﬁrst two options). This third option takes the acceptance of either (i) or (ii) to be absurd, on the ground that commitment to A requires commitment to A being true and conversely; but it nonetheless allows commitment to A ∨ ¬A. Now, many people think that if one accepts a disjunction of two options each of which would be absurd to accept, one has already accepted an absurdity. Indeed, that principle appears to be built into classical logic: it is the principle of reasoning by cases (or disjunction elimination), to which attention was called above. This third option is based on rejecting that principle, except in restricted form.⁹ So it is probably best thought of as only a semi-classical option: it does accept all the validities of classical logic, but disallows natural applications of disjunction elimination and some of the other standard meta-rules. These three options seem to be the only possibilities for keeping the validities of classical logic without accepting contradictions.¹⁰ Admittedly, one could insist with Tarski that the predicate ‘true’ should be given a hidden subscript, or that its extension ⁹ The restricted form is that if together with A entail C by classical rules, and together with B entail C by classical rules, then together with A ∨ B entail C. The third option can accept that, but cannot accept the generalization to ‘entailment’ by the truth rules (that commitment to A requires commitment to True(A ) and conversely). And these truth rules must have a quasi-logical status on the third option, since it was only by holding acceptance of (i) and of (ii) to be absurd that the view differentiated itself from agnosticism between the ﬁrst two options. ¹⁰ I know of no one who has seriously proposed taking the ﬁrst option. Classical and semi-classical logicians who do technical work on the paradoxes mostly tend to prefer the third option: see [10] and [14]; also [9], where ﬁve of the nine types of theories discussed fall under option three. The option of choice among non-specialists seems to be option two, but some specialists prefer it as well, e.g. [2] and [13]. (If the description of the latter as a classical theory seems surprising, see [8].) What about Kripke’s seminal [11]? That’s more complicated since Kripke offers a model-theoretic semantics with no instructions on how to read the theory off the semantics. But if we interpret him as suggesting that though the extension of ‘True’ is a ﬁxed point, the logic is classical, then his theory also falls under option two. An alternative and I think more attractive interpretation of Kripke is to take the set of acceptable sentences to coincide with the extension of ‘True’: they are both the contents of the same ﬁxed point. But if the ﬁxed points are based on a Kleene semantics, this gives a non-classical logic, and so is not germane to the discussion in this section. (This way of interpreting Kripke has been advocated in [21]—not altogether consistently, in my view, since Soames talks in terms of truth-value gaps, which seems to presuppose the classical logic interpretation. [19] clearly distinguishes the two ways of getting a theory from a Kripkean ﬁxed point, in the distinction between the theories there called KF and KFS.) On the non-classical reading of Kripke, his solution is similar in sprit to the G-solutions under discussion in this chapter; however, the nonclassical logic one obtains from this way of reading Kripke is unsatisfactorily weak, since Kleene semantics has no serious conditional. G-solutions do much better in this regard.

Solving the Paradoxes, Escaping Revenge / 85 vary with context. Still, given classical logic (even in the weak sense that includes only the validities and not the meta-rules), the above three options are the only consistent ones when the subscript and context are held ﬁxed.¹¹ A problem with all of the classical and semi-classical solutions is that they prevent the notion of truth from fulﬁlling its generally accepted role. The standard story about why we need a notion of truth ([18], [12]) is that we need it to make certain kinds of generalizations. For instance, the only way to generalize over (Snow is white) → ¬¬(Snow is white) (Grass is green) → ¬¬(Grass is green)], is to ﬁrst restate them in terms of truth and then generalize using ordinary quantiﬁers: For every sentence, if it is true, so is its double negation. But this says what we want it to say only if we assume the intersubstitutivity of True(A ) with A: that is, the principle Intersubstitutivity: If Y results from X by replacing some occurrences of A with True(A ), then X and Y entail each other. [This needs to be restricted to cases where the substitution is into contexts that aren’t quotational, intentional, etc.; but I’ll take the language L to contain no such contexts.]

This principle entails the truth schema in classical logic, indeed in any logic in which A ↔ A is a logical truth. So the three classical and semi-classical options all reject the intersubstitutivity principle. Thus they fail to satisfy the purpose of the notion. For instance, relative to the assumption that what Jones said was exactly A1 , . . . ,An , we want If everything Jones said is true then to be equivalent to If A1 and . . . and An then

.

This requires that the True(Ai ) be intersubstitutable with the Ai inside the conditional, but that won’t in general be so on any of the classical and semi-classical theories. The semi-classical theory does better than the fully classical ones: it allows for intersubstitutivity of True(A ) with A in more contexts. Indeed the fully classical ones don’t even allow substitutivity for unembedded occurrences: True(A ) and A can’t be mutually entailing in a classical theory that includes disjunction elimination (as we’ll see in the next section). But though the semi-classical theories do better, that ¹¹ I’m putting aside solutions to the Liar paradox based on unmotivated syntactic restrictions that prevent the formation of self-referential sentences. Very strong syntactic restrictions are required for this, and the solutions are of little interest since they do not generalize to the heterologicality paradox.

86 / Hartry Field isn’t good enough. An advantage of G-solutions is that not only do they accept the Tarski schema, they accept the full Intersubstitutivity Principle. I conclude this section with some further remarks on the second classical option; in particular, on a version of the second classical option that invokes a hierarchy of truth predicates. This will play a role later in the chapter, where I will compare it to a hierarchy of strengthenings of a single truth predicate that arises in G-solutions. A common theme among proponents of the second option is that the schema (T) holds for all sentences that ‘express propositions’, where to ‘express a proposition’ is to be either true or false, i.e. to either be true or have a true negation. (On this view, expressing a proposition is much stronger than being meaningful: it would be hard to argue that the ‘contingent Liar sentences’ of note 3 aren’t meaningful, but proponents of this option take them not to express propositions.) So instead of (T) we have (RT) [True(A ) ∨ True(¬A )] → [True(A ) ↔ A]. It is easily seen that this is equivalent to the left-to-right half of (T), i.e. to (LR) True(A ) → A.¹² Obviously, then, a decent theory of truth containing (RT)/(LR) needs to contain vastly more. (It’s compatible with (RT)/(LR) that nothing is true; or that only sentences starting with the letter ‘B’ are true; etc.) The crucial question for such a theory is, how are we to ﬁll it out without leading to paradox? It turns out that however we try to ﬁll it out, we are led to the conclusion that basic principles of the truth theory itself fail to be true. (They also fail to be false, so that they come out as ‘not expressing propositions’ .) Of course any non-contingent Liar sentence is itself an assertion of the theory that the theory asserts not to be true, but it presumably doesn’t count as one of the basic principles of the theory. But what does count as a basic principle of the theory is (RT), or its equivalent (LR). And Montague [15] showed that with very minimal extra assumptions, one can derive from (LR) a conclusion of form ¬True[True(M ) → M ], i.e. that some speciﬁc instance of (LR) isn’t true. Most people regard it as a serious defect of a theory that it declares central parts of itself untrue; and saying that these parts ‘don’t express propositions’ doesn’t appear to help much. This is the point at which the idea of a hierarchy of truth predicates may suggest itself. The idea is that we don’t have a general truth predicate, but only a hierarchy of predicates ‘trueα ’, where the subscripts are notations for ordinal numbers (in a ¹² In proving a given instance of schema (RT) from schema (LR) one uses two instances of the latter, one for A and one for ¬A.

Solving the Paradoxes, Escaping Revenge / 87 suitably large initial segment of the ordinals that has no last member). The Montague theorem then shows that the principles of the theory of truthα aren’t trueα , but the idea is to try to ameliorate this by saying that they’re trueα+1 . Call such a view a stratiﬁed truth theory. Besides their artiﬁciality, stratiﬁed truth theories seriously limit what we can express, in a way that undermines the point of the notion of truth. For instance, suppose we disagree with someone’s overall theory of something, but haven’t decided which part is wrong. The usual way of expressing our disagreement is to say: not all of the claims of his theory are true. Without a general truth predicate, what are we to do? The only obvious idea is to pick some large α, and say ‘Not all of the claims of his theory are trueα ’ . But this is likely to fail its purpose since we needn’t know how large an α we need. (Indeed, there would be strong pressure on each of us to use very high subscripts α even in fairly ordinary circumstances, but however high we make it there is a signiﬁcant risk of it not being high enough to serve our purposes. This was the lesson of the famous discussion of Nixon and Dean in [11]. Nixon and Dean wanted to say that nothing the other said about Watergate was true, and to include those assertions of the other in the scope of their own assertions; but to succeed, each needed to employ a strictly higher subscript than the other.) Actually, the situation is even worse than this. For suppose that we want to express disagreement with a stratiﬁed truth theorist’s overall ‘theory of truth’ (i.e. the theory he expresses with all of his ‘trueα ’ predicates), but that we haven’t decided which part of that theory is wrong. Here the problem isn’t just with knowing how high an α to pick; rather, no α that we pick could serve its purpose. The reason is that it’s already part of the stratiﬁed theory that some of its claims aren’t trueα , namely, the principles about truthα ; that’s why the theorist introduced the notion of truthα+1 . So whatever α we pick, we won’t succeed in expressing our disagreement. The problems just mentioned are really just an important special case of a problem that I’ve argued to infect all classical and semi-classical theories: they can’t give truth its proper role as a device of generalization. Except possibly for dialetheic theories, which I will not consider here, restricting excluded middle seems to be the only way to avoid crippling limitations on our notion of truth.

3 More on rejecting excluded middle It is important to note that in classical logic you don’t need anything like the full strength of the truth schema (T) (or the satisfaction schema (S)) to derive contradictions: indeed, if you allow reasoning by cases as well as the classical validities, all that is required is the two assumptions (T-Elim)

A follows from True(A )

88 / Hartry Field and (T-Introd)

True(A ) follows from A

(or the analogous Elimination and Introduction rules for satisfaction). For using these instead of (T), we can easily recast the derivation (i)-(iv) in Section 1 (with True(Q ) as the B) as follows: (i*) By (T-Elim), True(Q ) implies Q,¹³ which is equivalent to ¬True(Q ); hence True(Q ) implies the contradiction True(Q ) ∧ ¬True(Q ); (ii*) ¬True(Q ) is equivalent to Q, which by (T-Introd) implies True(Q ); hence ¬True(Q ) also implies the contradiction True(Q ) ∧ ¬True(Q ). (iii*) Since True(Q ) ∧ ¬True(Q ) follows both from the assumption True(Q ) and from the assumption ¬True(Q ), then it follows from the assumption True(Q ) ∨ ¬True(Q ). (Reasoning by cases.) (iv*) But True(Q ) ∨ ¬True(Q ) is a logical truth, so we have a derivation of the contradiction True(Q ) ∧ ¬True(Q ). (If we strengthened (T-Elim) to the assumption of the conditional True(A ) → A, we could give a derivation that doesn’t involve reasoning by cases.) In fact, we don’t even need the full strength of (T-Introd); we can make do with the weaker assumption (T-Incoherence) A and ¬True(A ) are jointly inconsistent. Inconsistency proof: True(Q ) implies Q by (T-Elim), and ¬True(Q ) implies Q since it is equivalent to Q, so we derive Q using reasoning by cases plus excluded middle. Using the other half of the equivalence between Q and ¬True(Q ), we get Q ∧ ¬True(Q ), which is inconsistent by (T-Incoherence). The fact that the paradox arises from weaker assumptions than (T) is important for two reasons. First and most obviously, it means that if we insist on keeping full classical logic we must do more than restrict (T), we must restrict the weaker assumptions as well. But the second reason it’s important concerns not classical solutions, but G-solutions: it gives rise to an important moral for what G-solutions have to be like. For even though G-solutions take truth to obey the Tarski schema (T), we’ll see that they recognize other ‘truth-like’ predicates (e.g. ‘determinately true’) that don’t obey the analog of (T) but do obey the analogs of (T-Elim) and (T-Introd) (or at the very least, (T-Elim) and (T-Incoherence)). For each truth-like predicate, there is a Liar-like sentence that asserts that it does not instantiate this predicate. ¹³ That is, Q follows from True(Q ). (One reader took my ‘A implies B’ to mean ‘if A then B’, and on this basis accused me here of illicitly extending (T-Elim) to hypothetical contexts; but that is not what I mean by ‘implies’.)

Solving the Paradoxes, Escaping Revenge / 89 Reasoning as in (i∗ ) and (ii∗ ) is thus validated, and since G-solutions accept reasoning by cases without restriction, paradox can only be avoided by rejecting the application of excluded middle to these Liar-like sentences formed from truth-like predicates. In short, G-solutions are committed to the view that there can be no truth-like predicate for which excluded middle can be assumed. (Since excluded middle is to hold within ordinary mathematics and physics, this means that no truth-like predicates can be constructed within their vocabulary.) As we’ll see, the conviction that there must be truth-like predicates obeying excluded middle is one primary source of revenge worries. I close this section by trying to make clear what is involved in restricting the application of excluded middle to certain sentences, e.g. the Liar sentence, when one accepts the intersubstitutivity of True(Q ) with Q. In particular, what is the appropriate attitude to take to the claim True(Q )? According to the sort of solution to the paradoxes I’ve sketched, one must reject the claim that True(Q ) and also reject the claim that ¬True(Q ), since these claims each imply a contradiction relative to any theory of truth that implies the Tarski biconditionals. (One can take rejection as a primitive state of mind, involving at the very least a refusal to accept; a slightly more informative account of rejection can be found in [4] (Section 3).) We must likewise reject the corresponding instance of excluded middle Z:

True(Q ) ∨ ¬True(Q ),

for it too leads to contradiction. And because we reject Z, our refusal to either accept True(Q ) or accept ¬True(Q ) doesn’t seem appropriately described as ‘agnosticism’ about the truth of Q. We would be agnostic about True(Q ) if we believed Z but were undecided which disjunct to believe; but when we reject Z the very factuality of the claim that True(Q ) is being put into question, so our not believing True(Q ) while also not believing ¬True(Q ) isn’t happily described as ‘agnosticism’ . It should be immediately noted that a solution of this sort does not postulate a ‘truth-value gap’ in Q: it does not say that Q is neither true nor false, i.e. that neither Q nor its negation is true. It also does not say that Q is neither true nor not true. Saying that Q is ‘gappy’ or ‘non-bivalent’ in either of these senses would trivially entail that Q is not true, which (by the Tarski biconditionals and modus ponens) leads to contradiction. Since the claim that Q is ‘gappy’ (non-bivalent) leads to contradiction, we must reject it. That isn’t to say that we should believe that Q is bivalent (or that it is not ‘gappy’; these are the same, assuming the equivalence of ¬¬A to A, as I henceforth shall). The claim that Q is bivalent or non-gappy amounts to Z∗ : True(Q ) ∨ True(¬Q ). This in turn amounts to Z (non-truth and falsity turn out to coincide as applied to sentences in the language), and as we’ve seen, Z must be rejected.

90 / Hartry Field If it seems odd that both the claim that Q is gappy and the claim that it is not gappy lead to contradiction, it shouldn’t: from the fact that Gappy(Q ) and ¬Gappy(Q ) each lead to contradiction, all we can conclude is that [email protected] :

Gappy(Q ) ∨ ¬Gappy(Q )

leads to contradiction; so the proper conclusion is that this instance of excluded middle must also be rejected.¹⁴ In other words, the claim that Q is ‘gappy’ has the same status as Q itself has. In particular, just as it is misleading to declare ourselves ‘agnostic’ about the Liar sentence, it is also misleading to declare ourselves ‘agnostic’ about the claim that the Liar sentence is ‘gappy’ or the claim that it is bivalent: for we don’t recognize that there is a fact to be agnostic about. I think it would be a serious problem if there were no way to assert the ‘defective’ status of Q within the language. As we’ll see, there is a way; but it can’t be done by saying that Q suffers a truth-value gap.

4 The Berry–Richard–K¨onig paradox I think that all of the semantic paradoxes turn on excluded middle, though some of them (especially some of the ones involving conditionals) do so in an indirect and unobvious fashion. I will make this precise in Section 5. There I will introduce a language that contains a ‘quasi-classical conditional’ which obeys many of the classical laws for conditionals even in the absence of excluded middle; moreover it reduces to the material conditional when excluded middle is assumed for antecedent and consequent. I will then state a result (proved elsewhere) according to which every semantic ‘paradox’ that can be formulated in this language has a solution that is compatible with the Tarski biconditionals. The solution may depend on the failure of some of the classical laws for the conditional, but that failure will always be traceable to a breakdown in excluded middle for the antecedent or consequent of one of the conditionals in question. We thus diagnose these apparent paradoxes as only apparent, they depend on illicit applications of excluded middle. Of course, the fact that those apparent paradoxes that can be formulated in the language turn out not be genuinely paradoxical does not settle the revenge issue: settling that issue requires considering the possibility of expanding the language to get new paradoxes. I will have a lot to say toward undermining the idea of revenge in later sections. ¹⁴ Indeed whenever one rejects a given instance A ∨ ¬A of excluded middle, one should also reject the instance (A ∨ ¬A) ∨ ¬(A ∨ ¬A), for they are equivalent by very uncontroversial reasoning; hence one should reject Bivalent(A ) ∨ ¬Bivalent(A ). [Reason for the equivalence: ¬(A ∨ ¬A) implies ¬A, so (A ∨ ¬A) ∨ ¬(A ∨ ¬A) implies (A ∨ ¬A) ∨ ¬A, which implies A ∨ ¬A. The other direction is trivial.]

Solving the Paradoxes, Escaping Revenge / 91 First though I will consider how the paradoxes of deﬁnability fare on this sort of view. There are a number of slightly different paradoxes of deﬁnability, the most famous being Berry’s and Richard’s, but they all have the same underlying idea. Because of its relevance later in the chapter, I will focus attention on a variant of the Berry and Richard paradoxes due to K¨onig. Recall that L is a ﬁrst order language adequate to expressing its own syntax, and that contains a satisfaction predicate. From that predicate we can deﬁne ‘L-deﬁnable’: z is L-deﬁnable if and only if there is at least one formula of L (with exactly one free variable) that is satisﬁed by z and by nothing else. Now, L is assumed to be built from a ﬁnite or countably inﬁnite vocabulary, so it contains only countably many formulas; from which it follows that only countably many things are L-deﬁnable. But there are uncountably many ordinal numbers; indeed, uncountably many countableordinalnumbers.So thereare (countable)ordinalnumbers that are not L -deﬁnable. So there is a smallest ordinal number that is not L-deﬁnable, and it must be unique. But then ‘v is an ordinal number that is not L-deﬁnable but for which all smaller ordinals are L-deﬁnable’ is uniquely satisﬁed by this ordinal, so it is L-deﬁnable after all, which is a contradiction. That is the paradoxical line of argument. Any solution to the paradoxes of satisfaction will implicitly contain a solution to this deﬁnability paradox. On classical logic solutions, if the language L contains the predicate ‘satisﬁes’ then certain instances of schema (S) from Section 1 are refutable; and in particular, if we deﬁne ‘L-deﬁnable’ from ‘satisﬁes’ as above, there will be counterinstances to even the more restricted schema (Sdefn ) For all x, x satisﬁes ‘v is L-deﬁnable’ if and only if x is L-deﬁnable. This gives one possible diagnosis of the error in the argument: that it lies in the inference from σ being the uniquely smallest L-undeﬁnable ordinal to ‘v is the smallest L-undeﬁnable ordinal’ being uniquely satisﬁed by σ . But on the approach that I’ve sketched, we are committed to maintaining all instances of (S), and in particular all instances of (Sdefn ). Where then does the reasoning of the paradox go wrong? Where the reasoning goes wrong, I think, is that it makes an implicit application of excluded middle to a formula involving ‘L-deﬁnable’. Excluded middle can be assumed for certain restricted deﬁnability predicates. For instance, let L0 be obtained from L by deleting ‘satisﬁes’ and terms deﬁned from it (such as ‘deﬁnable’) or closely related to it; then excluded middle holds for formulas that contain ‘L0 -deﬁnable’ (as long as they don’t contain problematic terms in addition). There is no even prima facie problem about the least ordinal that is not L0 -deﬁnable, since the description of it just given is in a part of L that goes beyond L0 . Similarly for expansions of L0 in which the application of ‘satisﬁes’ is somehow restricted in a way that guarantees excluded

92 / Hartry Field middle (e.g. a language L1 in which ‘x satisﬁes y’ can occur only in the context ‘x satisﬁes y and y is an L0 -formula’, or a language L2 in which ‘x satisﬁes y’ can occur only in the context ‘x satisﬁes y and y is an L1 -formula’). Given a well-deﬁned hierarchy of such expansions, each of which includes all the vocabulary of the previous, one gets within L a hierarchy of restricted deﬁnability predicates, each more inclusive than the previous. But for deﬁnability in the full language L, the fact that excluded middle must be rejected for ‘satisﬁes’ suggests that it will almost certainly have to be rejected for the predicate ‘L-deﬁnable’ deﬁned from it; and the paradox shows that indeed it does. The implicit application of excluded middle to a formula involving ‘L-deﬁnable’ occurred in the step from (1)

There are ordinal numbers that are not L-deﬁnable

(2)

There is a smallest ordinal number that is not L-deﬁnable.

to To see that the inference from (1) to (2) depends on excluded middle, consider any speciﬁc ordinal β, and suppose that every ordinal less than β is L-deﬁnable. Given this supposition, (2) says in effect (3)

Either β is not L-deﬁnable, or there is an ordinal α>β such that α is not L-deﬁnable and all its predecessors are L-deﬁnable.

But this entails (4)

β is not L-deﬁnable or β is L-deﬁnable;

and so if we reject (4) we must reject (2).¹⁵ But there is no call to reject (1): there are certainly ordinals that are not L-deﬁnable, for uncountable ones can’t be L-deﬁnable (and there may well be sufﬁciently large countable ones which are deﬁnitely not L-deﬁnable too). So the inference from (1) to (2) relies on excluded middle. This resolution of the paradox may seem to have a high cost. For the inference from ‘There are ordinals α such that F(α)’ to ‘There is a smallest ordinal α such that F(α)’ is absolutely fundamental to ordinary set-theoretic reasoning; doesn’t what I’m saying count as a huge and crippling restriction on ordinary set theory? Not at all: ordinary set theory allows sets to be deﬁned only by ‘effectively classical’ properties, that is, properties F for which the generalized law of excluded middle ∀x[F(x) ∨ ¬F(x)] holds. I’m not suggesting any restriction whatever on the ordinary laws of set theory; what I am saying, and what is independently quite obvious, is that one has to be very careful if one wants to extend set theory by allowing properties (or formulas) that aren’t known to be effectively classical into its axiom schemas. ¹⁵ Which isn’t to say that we should accept the negation of (2): that would require (an existential quantiﬁcation of) a negation of excluded middle, which would lead to contradiction.

Solving the Paradoxes, Escaping Revenge / 93 This point is worth elaboration. Standard set theory (ZFC) contains two axiom schemas (the schemas of Separation and Replacement). On a strict interpretation of the theory, the allowable instances of the schemas are just those instances in the language of set theory; however, the ‘impure’ set theory that most of us accept and employ is more extensive than this, it allows instances of the schemas in which physical vocabulary occurs (e.g. we take the separation schema to allow us to pass from the existence of a set of all non-sets to the existence of a set of all neutrinos). But when the law of excluded middle is not assumed to hold unrestrictedly, there is a question of just how far the extension should go. I think a suitable extension of the schema of separation to be the rule (Extended Separation) (∀x ∈ z)(Ax ∨ ¬Ax) ∃y∀x(x ∈ y ↔ x ∈ z ∧ A(x)) (allowing free parameters in the formula A(x)), where any vocabulary at all, including ‘true’, can appear in A(x). Requiring excluded middle as an assumption of separation seems reasonable. Otherwise, we would license sets for which membership in the set depends on whether the Liar sentence is true; given extensionality, this would lead at the very least to indeterminate identity claims between sets, and it isn’t at all clear that paradox could be avoided even allowing that. But Extended Separation as formulated above avoids such oddities, while allowing such sets as the set of true sentences in the ‘true’-free set-theoretic language; it seems to me as much of an extension of separation to the language containing ‘true’ as we ought to want. It is easy to see that if the formula F(x) is allowed to contain non-classical vocabulary, then Extended Separation (together with the fact that every non-empty set has a member of least rank) justiﬁes reasoning from ‘There is at least one ordinal α such that F(α) and such that for all ordinals β

This page intentionally left blank

Revenge of the Liar New Essays on the Paradox

edited by

JC Beall

1

1

Great Clarendon Street, Oxford ox2 6dp Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With ofﬁces in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York the several contributors 2007 The moral rights of the authors have been asserted Database right Oxford University Press (maker) First published 2007 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose this same condition on any acquirer British Library Cataloguing in Publication Data Data available Library of Congress Cataloging in Publication Data Data available Typeset by Laserwords Private Limited, Chennai, India Printed in Great Britain on acid-free paper by Biddles Ltd, King’s Lynn, Norfolk ISBN 978–0–19–923390–8 ISBN 978–0–19–923391–5 10 9 8 7 6 5 4 3 2 1

CONTENTS

Notes on Contributors 1 Prolegomenon to Future Revenge JC Beall

vii 1

2 Embracing Revenge: On the Indeﬁnite Extendibility of Language Roy T. Cook

31

3 The Liar Paradox, Expressibility, Possible Languages Matti Eklund

53

4 Solving the Paradoxes, Escaping Revenge Hartry Field

78

5 Validity, Paradox, and the Ideal of Deductive Logic Thomas Hofweber

145

6 On the Metatheory of Field’s ‘Solving the Paradoxes, Escaping Revenge’ Hannes Leitgeb

159

7 Reducing Revenge to Discomfort Tim Maudlin

184

8 Understanding the Liar Douglas Patterson

197

9 Revenge, Field, and ZF Graham Priest

225

10 Field on Revenge Agust´ın Rayo and P. D. Welch

234

vi / Contents 11 Bradwardine’s Revenge Stephen Read 12 Curry’s Revenge: The Costs of Non-classical Solutions to the Paradoxes of Self-reference Greg Restall

250

262

13 Aletheic Vengeance Kevin Scharp

272

14 Burali–Forti’s Revenge Stewart Shapiro

320

15 Revenge and Context Keith Simmons

345

Index

369

NOTES ON CONTRIBUTORS

JC Beall, Professor of Philosophy, University of Connecticut, and Arch´e Associate Research Fellow, University of St Andrews. In addition to articles and edited volumes on truth, paradox, and related issues, Beall is the author of Logical Pluralism with Greg Restall and the textbook Possibilities and Paradox: An Introduction to Modal and Many-Valued Logic with Bas C. van Fraassen. He is currently ﬁnishing a monograph on transparent truth and paradox (forthcoming with Oxford University Press). When not doing philosophy, Beall enjoys walking in the woods (and trying not to do philosophy), listening to music, and home schooling his cats. Roy T. Cook, Visiting Assistant Professor, Villanova University, and Arch´e Associate Research Fellow, University of St Andrews. Cook’s main research interests are philosophy of mathematics, philosophical logic, mathematical logic, and the philosophy of language. He has published papers on these topics in Mind, The Monist, Journal of Philosophical Logic, Journal of Symbolic Logic, Notre Dame Journal of Formal Logic, Dialectica, and Analysis, among other places. He is also editor of The Arch´e Papers on the Mathematics of Abstraction. Cook is currently working on a comprehensive dictionary of philosophical logic. When not thinking about logic and mathematics, Cook enjoys building sculptures and mosaics out of LEGO bricks, and attempting to keep his three cats from disassembling them. Matti Eklund, Assistant Professor, Sage School of Philosophy, Cornell University. Research interests: metaphysics, philosophy of language, and philosophy of logic. His publications include ‘Inconsistent Languages’ (Philosophy and Phenomenological Research 2002), ‘What vagueness consists in’ (Philosophical Studies 2005) and ‘Carnap and ontological pluralism’ (forthcoming in D. Chalmers, D. Manley, and R. Wasserman (eds.), Metametaphysics, Oxford University Press). His current research focuses on the implications of the liar and sorites paradoxes, and on various issues in metaontology. Hartry Field, Silver Professor of Philosophy, New York University. Field is the author of Science Without Numbers (Blackwell 1980), which won the Lakatos Prize, of Realism, Mathematics and Modality (Blackwell 1989), and of Truth and the Absence of Fact (Oxford 2001). His current research interests include objectivity and indeterminacy, a priori knowledge, causation, and the semantic and set-theoretic paradoxes; he is currently ﬁnishing a monograph on truth and paradox (forthcoming with Oxford University

viii / Notes on Contributors Press). Field hates writing either bibliographical or biographical blurbs, and so leaves that task to the editor. Thomas Hofweber, Associate Professor of Philosophy, University of North Carolina at Chapel Hill. Hofweber mainly works in metaphysics, the philosophy of language and the philosophy of mathematics. Sample publications: Inexpressible Properties and Propositions (Oxford Studies in Metaphysics, vol. 2, 2006), ‘Number determiners, numbers and arithmetic’ (Philosophical Review 2005), ‘Supervenience and object-dependent properties’ (Journal of Philosophy 2005), ‘A puzzle about ontology’ (Nous 2005). At present he is working on a book on the domain and methods of metaphysics, in particular ontology. Hannes Leitgeb, Reader in Mathematical Logic and Philosophy of Mathematics, University of Bristol. Leitgeb’s chief research interests are in philosophical logic, epistemology, cognitive science, and philosophy of mathematics. He has published, among others, in Journal of Philosophical Logic, Synthese, Analysis, Philosophia Mathematica, Erkenntnis, Journal of Logic, Language and Information, Studia Logica, Notre Dame Journal of Formal Logic, Topoi, Logique et Analyse, Artiﬁcial Intelligence. In 2004 he published Inference on the Low Level: An Investigation into Deduction, Nonmonotonic Reasoning, and the Philosophy of Cognition in the Kluwer/Springer Applied Logic series. At present, Leitgeb is trying to resurrect Carnap’s Logical Structure of the World. Apart from philosophy he likes music and marathons. Tim Maudlin, Professor of Philosophy, Rutgers University. Professor Maudlin’s primary interest is in the nature of reality. Recent books include Truth and Paradox and The Metaphysics Within Physics. He is inordinately fond of Belgian wafﬂes with whipped cream. Douglas Patterson, Associate Professor of Philosophy, Kansas State University. Patterson works in the philosophy of language, philosophical logic, and related areas in metaphysics, epistemology, and the philosophy of mind. Patterson is the author of a number of articles on truth and related topics, including ‘Theories of truth and convention T’, Philosophers’ Imprint 2:5 and ‘Tarski, the liar and inconsistent languages’, The Monist 89:1. He is the editor of a collection of essays on Tarski, forthcoming with Oxford University Press, and of Inquiry 50:6 on inconsistency theories of semantic paradox. In his spare time Patterson enjoys hiking and travel, and he was the 1986 Minnesota Junior Men’s Biathlon champion. Graham Priest, Boyce Gibson Professor of Philosophy, University of Melbourne, and Arch´e Professorial Fellow, University of St Andrews. Priest’s chief research interests are logic and related areas, including metaphysics and the history of philosophy (East and West). He has published in nearly all the major philosophy journals. Recent books include, Towards Non-Being, Doubt Truth to be a Liar, and the second edition of In Contradiction (all with Oxford University Press). He is currently ﬁnishing a second

Notes on Contributors / ix volume of his Introduction to Non-Classical Logic. When not doing philosophy, Priest enjoys doing philosophy. Agust´ın Rayo, Associate Professor of Philosophy, MIT, and Arch´e Associate Research Fellow, University of St Andrews. Rayo’s chief research interests are in philosophical logic and philosophy of language. Rayo has published in such areas in various journals, including Nous and the Journal of Symbolic Logic. He recently edited Absolute Generality with Gabriel Uzquiano. He is currently engaged in a research project on content. When he isn’t doing philosophy, Rayo enjoys scuba diving and opera. Stephen Read, Reader in History and Philosophy of Logic, University of St Andrews. His chief research interests are in philosophical logic, medieval logic, and metaphysics, in particular, the notion of logical consequence; and extend from medieval theories in the philosophy of language, mind, and logic, to the more modern concerns of relevance logic and the philosophy of logic. He has published in such areas in History and Philosophy of Logic, Journal of Philosophical Logic, Mind, Philosophy, Vivarium, and others. His books include Relevant Logic (1987) and Thinking about Logic (1995). He is currently working on a critical edition and English translation of Thomas Bradwardine’s Insolubilia. When not doing philosophy, he enjoys opera, cycling, and playing the clavichord. Greg Restall, Associate Professor of Philosophy, University of Melbourne. Restall’s research interests are in logic, metaphysics, and related ﬁelds. Recent books include Logic (a textbook) and Logical Pluralism, with co-author JC Beall, and he is currently working on the connections between proof theory and meaning. Restall enjoys looking after his young son. Kevin Scharp, Assistant Professor of Philosophy, The Ohio State University. Scharp’s main areas of research are philosophy of language and philosophical logic. He has published papers in British Journal for the History of Philosophy, International Journal of Philosophical Studies, and Inquiry. Scharp is editor (along with Robert Brandom) of the forthcoming volume of Wilfrid Sellars’ essays entitled In the Space of Reasons from Harvard University Press. Scharp is currently working on a book on truth and the liar paradox. When not doing philosophy, he loves backpacking, cooking, and listening to music. Stewart Shapiro, O’Donnell Professor of Philosophy, The Ohio State University, and Arch´e Professorial Fellow, University of St Andrews. He specializes in philosophy of mathematics, philosophical logic, and philosophy of language, notably vagueness. His most recent book is Vagueness in Context (Oxford: Oxford University Press, 2007) and he is the editor of the Oxford Handbook of the Philosophy of Mathematics and Logic. He still jogs (if you can still call it that), and likes the Grateful Dead and Incredible String Band. Keith Simmons, Professor of Philosophy, University of North Carolina at Chapel Hill. He has research interests in logic, philosophy of logic, and philosophy of language. He

x / Notes on Contributors is the author of Universality and the Liar (Cambridge University Press) and, with Simon Blackburn, the editor of Truth (in the Oxford Readings in Philosophy series). He is currently at work on a monograph about the paradoxes, and, with Dorit Bar-On, a monograph about truth. P. D. Welch, Professor of Pure Mathematics, University of Bristol. He has research interests in set theory, models of computation, theories of truth, and possible worlds semantics. His work on these and related topics have appeared in many of the leading logic and mathematics journals. Welch is currently working with the The Luxemburger Zirkel on a major research project entitled Logical Methods in Epistemology, Semantics, and Philosophy of Mathematics.

1 Prolegomenon to Future Revenge JC Beall

This chapter attempts to lay out some background to the target phenomenon: the Liar and its revenge. The phenomenon is too big, and the literature (much) too vast, to give anything like a historical summary, or even an uncontroversial sketch of the geography. Accordingly, my aim is simply to lay out a few background ideas, in addition to brieﬂy summarizing the contributed essays. I also try to avoid overlap with the individual chapters’ rehearsals of revenge (including standard references to historical theses, like ‘semantic self-sufﬁciency’), as the chapters do a nice job covering such material. Finally, because some of the ideas are presupposed by many of the chapters in this volume, a gentle sketch of Kripke’s ‘ﬁxed point’ approach to truth is given in an appendix.

1.1 Truth Whatever else it may do, truth is often thought to play Capture and Release. Where Tr(x) is our truth predicate, α a sentence, and α a name of α, Capture and Release are as follows. Capture: α ⇒ Tr(α) Release: Tr(α) ⇒ α When ⇒ is a conditional, we have the Conditional Form of Capture and Release: namely, when conjoined, the familiar T-biconditionals. When ⇒ is a turnstile, we have the Rule Form of Capture and Release, which indicates ‘valid inference’ (in some sense).

2 / JC Beall The names ‘Capture’ and ‘Release’ arise from the fact that Tr(x) captures the information in x, fully storing it for its eventual release. In practice, a familiar—if not the—role of Tr(x) is its release function: an assertion of Tr(α) releases all of the information in α. This is useful for ‘long generalizations’ or ‘blind generalizations’ or the like, many of which generalizations would be practically impossible if we didn’t enjoy a truth predicate that played (at least the rule form of) Capture and Release.¹ That truth plays Capture and Release in Conditional Form is plausible but controversial. That truth plays Capture and Release in at least Rule Form is less controversial, and will henceforth be assumed.²

1.2 The Liar The Liar phenomenon involves sentences that imply their own falsity or, more generally, untruth. By way of example, consider the ticked sentence in §1.2 of this chapter. √ The ticked sentence in §1.2 of ‘Prolegomenon to future revenge’ is not true. Assume that the ticked sentence is true. Release, in turn, delivers that the ticked sentence is not true. Hence, the ticked sentence, given Release (an essential feature of truth), implies its own untruth. Is the ticked sentence untrue? Capture gives reason for pause: that the ticked sentence is not true implies, via Capture, that the ticked sentence is true! The question is: what shall we say about the ‘semantic status’ of the ticked sentence? Answering this question invites the Liar’s revenge.

1.3 The Liar’s Revenge On one hand, the revenge phenomenon—the Liar’s revenge—is not so much a distinct phenomenon from the Liar as it is a witness to both the difﬁculty and ubiquity of Liars. ¹ Depending on the language, Rule Capture and Release is insufﬁcient for a fully transparent truth predicate, one such that Tr(x) and x are intersubstitutable in all (non-opaque) contexts, for all sentences x in the language. By my lights, the Liar phenomenon is at its most difﬁcult incarnation when truth is fully transparent, since any distinction between, for example, ‘Excluded Middle’ and ‘Bivalence’ collapses. But I will set this aside here. See Appendix for one approach to transparent truth, and see Field’s chapter (Chapter 4) for another, stronger approach, as well as relevant discussion. ² A variety of theories reject even Rule Form of Capture and Release, but it will be assumed throughout this ‘introduction’. One of the better known examples of rejecting Rule Capture is so-called Kripke–Feferman [3]. (See too [16].)

Prolegomenon to Future Revenge / 3 On the other hand, ‘revenge’ is often launched as an objection to an account of truth (or a response to Liars). Without intending a stark distinction, I will discuss revenge qua Liar phenomenon and revenge qua objection separately, with most of the discussion on the latter but all of the discussion brief.

1.3.1 The revenge phenomenon The revenge phenomenon arises at the point of classifying Liars. Consider, again, the ticked sentence. As above, classifying the ticked sentence as true results in inconsistency; Release delivers that the ticked sentence is also not true. Likewise, classifying the ticked sentence as not true results in inconsistency; Capture delivers that the ticked sentence is also true. How, then, shall we classify the ticked sentence? A natural suggestion is that the ticked sentence is neither true nor false. The trouble with this suggestion—even apart from logical issues involving negation—is the apparent connection between being neither true nor false and being not true.³ In particular, presumably, we have ntf–nt. ¬Tr(α) ∧ ¬Tr(¬α) ⇒ ¬Tr(α) Again, ⇒ may be a conditional or a turnstile. Either way, the problem at hand is plain. Assume, as per the current suggestion, that the ticked sentence is ‘neither true nor false’. By ntf–nt, we immediately get that the ticked sentence is not true. But, now, we’re back to inconsistency, as Capture, in turn, delivers that the ticked sentence is (also) true. So, while natural, the suggestion that the ticked sentence is neither true nor false is not a promising proposal.⁴ Quick reﬂection leads to a general lesson: whatever category one devises for the ticked sentence, it had better not imply untruth. For example, suppose that one introduces the category bugger for Liars. On this proposal, the ticked sentence is a bugger. Whatever else being a bugger might involve, we had better not have bug if we’re to avoid inconsistency. bug. Bugger(α) ⇒ ¬Tr(α) The trouble with bug is exactly the trouble with ntf–nt. On the current proposal, the ticked sentence is a bugger, in which case, via bug, it is not true. Capture, as before, delivers that the ticked sentence is (also) true, and inconsistency remains. ³ Throughout, I will assume that falsity is truth of negation—i.e., that α is false just if ¬α is true. (This is a standard line, but it might be challenged. Fortunately, in the present context, nothing substantive turns on the issue.) ⁴ I should note that my presentation simpliﬁes matters a great deal. One might postulate a different negation at work in (wide-scope positions in) ntf–nt, thereby complicating matters. Moreover, one might—perhaps with some philosophical motivation—reject ntf–nt altogether. And there are other options, as will be evident in various chapters of this volume.

4 / JC Beall One lesson, then, is that our classiﬁcation of the ticked sentence cannot consistently deliver its untruth. With the lesson in mind, suppose that we classify the ticked sentence as a bugger but, whatever else ‘bugger’ might mean, we reject bug (in both Rule and Conditional Forms). Notwithstanding further details on ‘buggerhood’, this course yields the promise of consistently classifying the ticked sentence (and its negation): it is a bugger. The revenge phenomenon re-emerges. Having, as we’re assuming, consistently classiﬁed the ticked sentence as a ‘bugger’, other Liars emerge to thwart our aims at consistently (and completely) classifying Liars. By way of example, consider the starred sentence. The starred sentence in §1.3.1 of ‘Prolegomenon to future revenge’ is either not true or a bugger. Assume that the starred sentence is true. Release delivers that the starred sentence is not true or a bugger, and hence true and either not true or a bugger. Similarly, that the starred sentence is either not true or a bugger implies, via Capture, that it is true—and, hence, true and either not true or a bugger. Accordingly, given normal conjunction and disjunction behavior, if we have it that the starred sentence is either true, not true, or a bugger, we have either inconsistency (viz., true and not true) or some true buggers. While the latter option, without bug (or similar principles), might afford a consistent theory, it is prima facie objectionable if, whatever else ‘bugger’ might mean, the buggers are thought to be somehow ‘defective’, sentences that ought to be rejected.⁵ The revenge phenomenon, at least in one relevant sense, is as above: it is not so much a separate phenomenon from the Liar as it is what makes the Liar phenomenon challenging. The Liar’s revenge is reﬂected in the apparent hyrdra-like appearance of Liars: once you’ve dealt with one Liar, another one emerges. In short, if one manages to consistently classify a Liar as a such-n-so, another Liar emerges—e.g. a sentence that says of itself only that it’s not true or a such-n-so. Dramatically and very generally put, Liars attempt to wreak inconsistency in one’s language. If the Liar can’t have what she wants, she’ll enlist ‘strengthened’ relatives to frustrate your wants, in particular, your expressive wants. As it is sometimes put, Liars force—or try to force—you to choose between either inconsistently expressing what you want to express or not expressing what you want to express. A quietist advises that we give up on our aim to classify Liars; there are Liars in the language, but there is no ‘semantic category’ in which the ticked sentence may truly be said to reside. Accordingly, whereof one cannot truly classify, thereof one must—or, in any event, might as well—be silent. The virtue of such an approach ⁵ Whether ‘buggers’ should be conceived as defective (in some sense) is an open issue. Field’s chapter (see Chapter 4) is relevant to this issue. See too [1].

Prolegomenon to Future Revenge / 5 is that it avoids revenge (since it doesn’t engage); the salient defect is that it offers no clear account of truth or the paradoxes at all. Against such a ‘proposal’, little can be said, and so won’t. Another—so-called dialetheic—option is to accept the apparent inconsistency engendered by Liars. Provided that our logic tolerates such inconsistency—and part of theproposed lessonoftheLiaristhatourlogicdoestoleratesuchinconsistency—there’s no obvious problem. What Liars teach us, on the dialetheic view, is that truth is inconsistent—that some true sentences have true negations. Whether such a position avoids the Liar’s revenge is an open question.⁶ And there are (many) other options, as subsequent chapters reﬂect. What is uncontroversial is that the revenge phenomenon has fueled, and continues to fuel, work on the Liar phenomenon. This is not surprising, at least if, as suggested, the revenge phenomenon just is the Liar phenomenon—indeed, as above, a witness to the Liar’s ubiquity.

1.3.2 Towards revenge qua objection The literature on truth and paradox exhibits a familiar and ubiquitous pattern: each proposed ‘account of truth’ is followed by a charge of revenge, that the account can’t accommodate such and so a notion (e.g. ‘untruth’, ‘exclusively false’, or whathaveyou) and, in that respect, is thereby inadequate. Indeed, were it not for alleged ‘revenge’ problems, many proposed theories of truth might be objection-free—or, at least, the number of known or cited objections would be greatly diminished. Such ‘revenge’ charges, as said, are often launched as inadequacy objections against proposed accounts of truth. Unfortunately, there is some unclarity about the relevance of such charges, and, more to the point, unclarity with respect to the burden involved in successfully establishing the intended inadequacy result. Without aiming to resolve them, §1.4 brieﬂy discusses some of the given issues involved in revenge qua objection. Before turning to §1.4, two background issues need to be brieﬂy covered.⁷ 1.3.2.1 Incoherent operators By way of background, it is important to see that there are operators that cannot coherently exist if our language enjoys various features. Tarski’s Theorem gives one ⁶ That some sentences are true and false is one thing; however, the dialetheic position is rational only if at least some sentences are just true. The worry is whether the dialetheist can give an adequate account of ‘just true’ without the position exploding into triviality. Some of the chapters have discussion of this point. For a general discussion (and defense) of dialetheism, see [14, 15]. ⁷ I should warn that, from this point forward, my presentation may border on controversial.

6 / JC Beall concrete example of such a result,⁸ but another example might be useful. In particular, suppose, as is plausible, that our language has features F1 and F2. F1. There’s a predicate Tr(x) that ‘obeys’ (unrestricted) Release and Capture in at least Rule Form. F2. ‘Reasoning by Cases’ is valid: if α implies γ , and β implies γ then α ∨ β implies γ , for all α, β, γ . As such, the language, on pain of triviality, has no operator such that both E1 and E2 hold.⁹ E1. α ∨ α E2. α, α ⊥ Suppose that we do have such an operator. Consider a familiar construction, which will be guaranteed via diagonalization, self-reference or the like: a sentence λ that ‘says’ Tr(λ). From E1, we have Tr(λ) ∨ Tr(λ) which yields two cases. 1. Case one: (a) Tr(λ) (b) Release yields:¹⁰ Tr(λ). (c) E2 yields: ⊥ 2. Case two: (a) Tr(λ) (b) Capture yields: Tr(λ) (c) E2 yields: ⊥ The point, for present purposes, is modest but important: there are incoherent notions, notions that cannot coherently exist if our language enjoys various features. While modest, the point is something on which all parties can agree. ⁸ Tarski’s Theorem, in effect, is that (classical) arithmetical truth is not deﬁnable in (classical) arithmetic. For a user-friendly discussion of the theorem and its broader implications, see [20] and, more in-depth, [18]. For a user-friendly discussion of what Tarski’s Theorem does not teach us, see [23], which is also highly relevant to ‘revenge’ issues, in general, and particularly relevant to Field’s proposal (see Chapter 4). ⁹ E1 might be thought of as an exhaustion principle, and E2 as exclusion or explosion principle. Throughout, ⊥ is an ‘explosive’ sentence, one that implies all sentences. ¹⁰ Intersubstitutability of Identicals is also involved here (and at the same place in Case two). This is usually assumed to be valid, but it, like so much in the area, has been challenged. See [17].

Prolegomenon to Future Revenge / 7 A principal question, at the heart of Liar studies, is this: what is our language like, given that it enjoys such and so features? More to the point: assuming that our language has a truth predicate that plays Capture and Release (in at least rule form), what are its other features? One might say that it fails to contain a fully exhaustive device, something that would yield E1, or fails to have any fully explosive device, something that would yield E2. One might, with various theorists, say that F2, in its given unrestricted form, fails for our language. One might say other things. Whatever one says, one aims to give a clear, precise account of the matter—a clear, precise account of what our language is like, given that it has such and so features. This is normally done by way of a ‘formal modeling’. 1.3.2.2 Models and reality Like much in philosophical logic, constructing a formal account of truth is ‘model building’ in the ordinary ‘paradigm’ sense of ‘model’. The point of such a model is to indicate how ‘real truth’ in our ‘real language’ can have the target (logical) features we take it to have—e.g. consistency (or, perhaps, inconsistency but non-triviality), Release and Capture features, perhaps full intersubstitutability of Tr(α) and α. In that respect, formal accounts of truth are idealized models to be evaluated by their adequacy with respect to the ‘real phenomena’ they purport to model.¹¹ Formal accounts (or theories) of truth aim only indirectly at being accounts of truth. What we’re doing in giving such an account is two-fold. 1. We construct an artiﬁcial model language—one that’s intended to serve as a heuristic, albeit idealized, model of our own ‘real’ language—and, in turn, give an account of how ‘true’ behaves in that language by constructing a precise account of truth-in-that-language. 2. We then claim that the behavior of ‘true’ in our language, at least in relevant, target respects, is like the behavior of the truth predicate in our model language. By far the most dominant approach towards the ﬁrst task—viz. constructing one’s model language—employs a classical set theory. One reason for doing so is that classical set theory is familiar, well-understood, and generally taken to be consistent. A related reason is that, in using a classical set theory, one’s formal account of truth ¹¹ Theories, like McGee’s [13], that purport not to be ‘descriptive’ but, rather, ‘revisionary’ or ‘normative’, are not typically subject to ‘revenge’-charges to the same extent that ‘descriptive’ theories are, and so are not the chief concern here. On the other hand, McGee aims to give a revisionary theory (not to be confused with revision theory) that aims to stay as close to the phenomena—our ‘real language’—as possible. In that respect, ‘revenge’ objections might well arise.

8 / JC Beall can be more than merely a heuristic picture; it can also serve as a ‘model’ in the technical sense of establishing consistency.¹² That a classical set theory is used in constructing our artiﬁcial language serves to emphasize the heuristic, idealized nature of the construction. We know that, due to paradoxical sentences, there’s no truth predicate in (and for) our ‘real language’ if our real language is (fully) classical.¹³ But the project, as above, is to show how we can have a truth predicate in our ‘real language’, despite such paradoxical sentences. And the project, as above, is usually—if not always—carried out in a classical set theory. Does this mean that the project, as typically carried out, is inexorably doomed? Not at all. Just as in physics, where idealization is highly illuminating despite its distance from the real mess, so too in philosophical logic: the classical construction is illuminating and useful, despite its notable idealization. But it is idealized, and, pending argument, on the surface only heuristic. That’s the upshot of using classical set theory.

1.4 Comments on Revengers’ Revenge A quick glance at the Liar literature will indicate that ‘revenge’ is often invoked as a problem for a given theory of truth and paradox. For present purposes, a revenger is one who charges ‘revenge’ against some proposed account of truth. The principal issue of this section is the burden of revenge—the burden that revengers carry. The chapters in this volume will tell their own (and not necessarily compatible) story on this issue.

1.4.1 Too easy revenge As above, in giving a formal theory of truth, one does not directly give a theory of truth; rather, one gives a theory of Lm -truth, an account, for some formal ‘model language’ Lm , of how Lm ’s truth predicate behaves, in particular, its logical behavior. By endorsing a formal theory of truth, one is endorsing that one’s own truth predicate is relevantly like that, like the truth predicate in Lm , at least with respect to various phenomena in question—for example, logical behavior. ¹² In paraconsistent contexts, the aim is basically the same, except that the target result is non-triviality despite negation-inconsistency. In the more dominant non-paraconsistent cases, the aim is also non-triviality, but that’s ensured by consistency. ¹³ The same applies, of course, if the truth predicate has an extension: the extension isn’t really a classical set. Every classical set S is such that x ∈ S ∨ x ∈ / S, which, given paradoxical sentences, results in inconsistency. (The point is independent of ‘size’ issues. Classical proper classes are likewise such that x ∈ C ∨ x ∈ / C.) If T is the extension of Tr(x) and T is a set, a sentence λ that ‘says’ λ ∈ /T makes the point—assuming, as is plausible, suitable ‘extension’ versions of Capture and Release (e.g., α ⇒ α ∈ T, etc.).

Prolegomenon to Future Revenge / 9 Revenge qua objection—revenger’s revenge—is an adequacy objection. Typically, the revenger charges that a given ‘model language’ is inadequate due to expressive limitation. Let L be our ‘real language’, English or some such natural language, and let Lm be our heuristic model language. Let ‘Lm -truth’ abbreviate ‘the behavior of Lm ’s truth predicate’. In broadest terms, the situation is this: we want our (heuristic) Lm , and in particular Lm -truth, to illuminate relevant features of our own truth predicate, to explain how, despite paradoxical sentences, our truth predicate achieves the features we take it to have. Revenge purports to show that Lm achieves its target features in virtue of lacking expressive features that L itself (our real language) appears to enjoy. But if Lm enjoys the target features only in virtue of lacking relevant features that our real L enjoys, then Lm is an inadequate model: it fails to show how L itself achieves its target features (e.g. consistency). That, in a nutshell, is one common shape of revenge. Consider a familiar and typical example, namely, Kripke’s partial languages.¹⁴ Let Lm , our heuristic model language, be such a (ﬁxed point) language constructed via the Strong Kleenescheme.¹⁵ Inconstructing Lm ,weuse—inourmetalanguage—classical set theory, and we deﬁne truth-in-Lm (and similarly, false-in-Lm ), which notions are used to discuss Lm -truth (the behavior of Lm ’s truth predicate). Moreover, we can prove—in our metalanguage—that, despite paradoxical sentences, a sentence Tr(α) is true-in-Lm exactly if α is true-in-Lm . The familiar revenge charge is that Lm , so understood, is not an adequate model; it fails to illuminate how our own truth predicate, despite paradoxical sentences, achieves consistency. In particular, the revenger’s charge is that Lm -truth achieves its consistency in virtue of Lm ’s expressive poverty: Lm cannot, on pain of inconsistency, express certain notions that our real language can express. Example: suppose that Lm contains a predicate ϕ(x) that deﬁnes {β : β is not true-in-Lm }. And now, where λ says ϕ(λ), we can immediately prove—in the metalanguage—that λ is true-in-Lm iff ϕ(λ) is true-in-Lm iff λ is not true-in-Lm . Because—and only because—we have it (in our classical metalanguage) that λ is true-in-Lm or not, we thereby have a contradiction: that λ is both true-in-Lm and not. But since we have it that truth-in-Lm is consistent (given consistency of classical set theory in which Lm is constructed), we conclude that Lm cannot express ‘is not true-in-Lm ’. The revenger’s charge, then, amounts to this: that the Kripkean model language fails to be enough like our real language to explain at least one of the target phenomena, namely, truth’s consistency. Our metalanguage is part of our ‘real language’, and we can deﬁne {β : β is not true-in-Lm } in our metalanguage. As the Kripkean language ¹⁴ See Appendix for a sketch of the Kripkean ‘partial predicates’ approach. ¹⁵ The point applies to any of the given languages, but the K3 -construction (Strong Kleene) is probably most familiar.

10 / JC Beall cannot similarly deﬁne {β : β is not true-in-Lm }, the Kripkean model language is inadequate: it fails to illuminate truth’s target features. A revenger engages in ‘too easy revenge’ if the revenger only points to such a result without establishing its relevance.¹⁶ The relevance of such a result is not obvious. After all, the given notion is a classically constructed notion; it is a ‘model-dependent’ notion—a notion that makes no sense apart from the given (classically constructed) models—deﬁned entirely in a classical metalanguage. As such, the given notion, presumably, is not one of the target (model-independent, or ‘absolute’) notions in L that Lm is intended to model. The question, then, isn’t whether there’s some notion X (e.g., ‘not true-in-Lm ’) that is inexpressible—or, at least, not consistently expressible—in Lm . The question is the relevance of such a result. One might think that the relevance is plain. One might, for example, think that the semantics for Lm is intended to reﬂect the semantics of L, our real language. Since the semantics of the former essentially involves, for example, not true-in-Lm , the semantics of our real language must involve something similar—at least if Lm is an adequate model of our real language. But, now, since not true-in-Lm is (provably) inexpressible in Lm , we should conclude that Lm is an inadequate model of our real language L, since our real language can express its own semantic notions—i.e. the notions required for giving the semantics of our language. Such an argument might serve to turn otherwise ‘too easy revenge’ into a plainly relevant and powerful objection; however, the argument itself relies on various assumptions that involve quite complex issues. For example, one conspicuous assumption is that the ‘semantics’ of Lm is intended to reﬂect the semantics of our real language L. This needn’t be the case. For example, suppose that one rejects that semantics—the semantics of our real language—is a matter of giving ‘truth conditions’ or otherwise involves some explanatory notion of truth. In the face of Liars (or other paradoxes), one still faces questions about one’s truth predicate, and in particular its logical behavior. By way of answering such questions, one might proceed as above: construct a model language that purports to illuminate how one’s real truth predicate enjoys its relevant features (e.g. Capture and Release) without collapsing from paradox. In constructing and, in turn, describing one’s ‘model language’, one might give ‘truth-conditional-like semantics’ for the model language by giving ‘truth-in-a-model conditions’ for the language. If so, it is plain that the ‘semantics’ of the model language are not intended to reﬂect the ‘real semantics’ of one’s real language; they may, in the end, be only tools used for illuminating the logic of our real language, versus illuminating the ‘real semantics’ of our real language. So, a critical assumption in the argument above—the argument towards the relevance of the given inexpressibility results—requires argument. Likewise, the assumption that our ¹⁶ Thanks to Lionel Shapiro for very useful discussion on ‘too easy revenge’.

Prolegomenon to Future Revenge / 11 real language can ‘express its own semantic notions’, the notions involved in ‘giving the semantics’ of our language, requires argument, argument that may turn, as with the ﬁrst assumption, on difﬁcult issues concerning the very ‘nature of semantics’.¹⁷ A would-be revenger, involved in too easy revenge, would have it easy but too easy. What is (generally) easy is showing that some classically constructed notion is inexpressible—or, at least, not consistently expressible—in a (classically constructed) non-classical ‘model language’. What is too easy is the thought that showing as much is sufﬁcient to undermine the adequacy of the given model language. The hard part is clearly establishing the relevance of such inexpressibility results, that is, clearly substantiating the alleged inadequacy. The difﬁculty, as above, is that the alleged inadequacy often relies on very complicated issues—the ‘nature of semantics’, the role of given model-dependent notions, and more.

1.4.2 Revenger’s recipes, in general Towards clarifying the burden involved in launching revenger’s revenge, it might be useful to lay out a few common recipes for revenge qua objection. For simplicity, let Lm be a given formal model language for L, where L is our target, real language—the language features of which Lm is intended to illuminate. Let M(Lm ) be the metalanguage for Lm , and assume, as is typical, that M(Lm ) is a fragment of L. Then various (related) recipes for revenge run roughly as follows.¹⁸ Rv1. Recipe One. • Find some semantic notion X that is used in M(Lm ) to classify various Lm sentences (usually, paradoxical sentences). • Show, in M(Lm ), that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). Rv2. Recipe Two. • Find some semantic notion X that, irrespective of whether it is explicitly used to classify Lm -sentences, is expressible in M(Lm ). • Show, in M(Lm ), that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). ¹⁷ Some of the chapters in this volume discuss this assumption, an assumption that often goes under the heading ‘semantic self-sufﬁciency’. For arguments against such an assumption, see [7, 8]. ¹⁸ This is not in any way an exhaustive list of recipes!

12 / JC Beall Rv3. Recipe Three. • Find some semantic notion X that is (allegedly) in L. (Argue that X is in L.) • Argue that X is not expressible in Lm lest Lm be inconsistent (or trivial). • Conclude that Lm is explanatorily inadequate: it fails to explain how L, with its semantic notion X, enjoys consistency (or, more broadly, non-triviality). As above, a revenger is one who charges ‘revenge’ against a formal theory of truth, usually along one of the recipes above. The charge is that the model language fails to achieve its explanatory goals. In general, the revenger aims to show that there’s some sentence in our real language L that ought to be expressible in Lm if Lm is to achieve explanatory adequacy. The question is: how ought one reply to revengers? The answer, of course, depends on the details of the given theories and the given charge of revenge. For present purposes, without going into such details, a few general remarks can be made. The weight of Rv1 or Rv2 depends on the sort of X at issue. As in §1.3.2.2 and §1.4.1, if X is a classical, model-dependent notion constructed in a proper fragment of L, then the charge of inadequacy is not easy to substantiate, even if the inexpressibility of X in Lm is easy to substantiate. In particular, if classical logic extends that of Lm , then there is a clear sense in which you may ‘properly’ rely on a classical metalanguage in constructing Lm and, in particular, truth-in-Lm . In familiar non-classical proposals, for example, you endorse that L, the real, target language, is non-classical but enjoys classical logic as a (proper) extension, in which case, notwithstanding particular details, there is nothing prima-facie suspect about relying on an entirely classical fragment of L to construct your model language and, in particular, classical model-dependent Xs. But, then, in such a context, it is hardly surprising that X, being an entirely classical notion, would bring about inconsistency or, worse, triviality, in the (classically constructed) non-classical Lm .¹⁹ Because classical logic is typically an extension of the logic of Lm , the point above is often sufﬁcient to blunt, if not undermine, a revenger’s charge, at least if the given recipe is Rv1 and Rv2. As in §1.4.1, the revenger must establish more than the unsurprising result that a (usually classically constructed) model-dependent X is expressible in M(Lm ) but not in Lm ; she must show the relevance of such a result, which might well involve showing that some non-model-dependent notion—some relevant ‘absolute’ notion—is expressible in L but, on pain of inconsistency (or non-triviality), inexpressible in Lm . And this task brings us to Rv3. Recipe Rv3 is perhaps what most revengers are following. In this case, the idea is to locate a relevant non-model-dependent notion in L and show that Lm cannot, on pain of inconsistency (or triviality), express such a notion. The dialectic along these lines is delicate. ¹⁹ For closely related discussion, see Field’s chapter (Chapter 4) and also [4].

Prolegomenon to Future Revenge / 13 Suppose that Theorist proposes some formal theory of truth, and Revenger, following Rv3, adverts to some ‘absolute’ notion X that, allegedly, is expressible in L. If, as I’m now assuming, Theorist neither explicitly nor implicitly invokes X for purposes of classifying sentences, then Revenger has a formidable task in front of her. In particular, without begging questions, Revenger must show that X really is an intelligible notion of L. For example, recall, from §1.3.2.1, the discussion of ‘incoherent operators’, and assume that Theorist proposes a theory that has features F1 and F2 (and, for simplicity, is otherwise normal with respect to extensional connectives). Let any operator that satisﬁes E1 and E2 be an EE device (for ‘exclusive and exhaustive’). Against a typical paracomplete (or paraconsistent) proposal,²⁰ an Rv3-type revenger might maintain that L, our real language, enjoys an EE device. If the revenger is correct, then standard paracomplete and paraconsistent proposals are inadequate, to say the least. But the issue is: why think that the revenger is correct? Argument is required, but the situation is delicate. What makes the matter delicate is that many arguments are likely to beg the question at hand. After all, according to (for example) paracomplete and paraconsistent theorists, what the Liar teaches us is that, in short, there is no EE device in our language! Accordingly, the given revenger cannot simply point to normal evidence for such a device and take that to be sufﬁcient, since such ‘evidence’ itself might beg questions against such proposals. On the other hand, if the given theorist cannot otherwise explain—or, perhaps, explain away—normal evidence for the (alleged) device, then the revenger may make progress. But the situation, as said, is delicate. The difﬁculty in successfully launching Rv3 might be put, in short, as follows. Theorist advances Lm as a model of (relevant features of) L, our real language. Rv3 Revenger alleges that X exists in L, and shows that, on pain of triviality, X is inexpressible in Lm . The difﬁculty in adjudicating the matter is that, as in §1.3.2.1, Theorist may reasonably conclude that X is incoherent (given the features of our language that Theorist advances). Of course, if Revenger could establish that we need to recognize X, perhaps for some theoretical work or otherwise, then the debate might be settled; however, such arguments are not easy to come by. The burden, of course, lies not only on the Rv3 Revenger; it also lies with the given theorist. For example, typical paracomplete and paraconsistent theorists must reject the intelligibility of any EE device in our language. Inasmuch as such a notion is independently plausible—or, at least, independently intelligible—such theorists carry the burden of explaining why such a notion appears to be intelligible, despite its ultimate unintelligibility. Along these lines, the theorist might argue that we are making a common, reasonable, but ultimately fallacious generalization from ‘normal ²⁰ A paracomplete proposal rejects LEM, and a paraconsistent proposal rejects ‘Explosion’ (i.e., α, ¬α ⇒ β, in both Rule and Conditional form). (See Appendix for the former type of approach.)

14 / JC Beall cases’ to all cases, or some such mistake. (E.g. some connective, if restricted to a proper fragment of our language, behaves in the EE way.) Alternatively, such theorists might argue that, contrary to initial appearances, the allegedly intelligible notion only appears to be a clear notion but, in fact, is rather unclear; once clariﬁed, the alleged EE device (or whatever) is clearly not such a device. (E.g. one might argue that the alleged notion is a conﬂation of various notions, each one of which is intelligible but not one of which behaves in the alleged, problematic way.) Whatever the response, theorists do owe something to Rv3 revengers: an explanation as to why the given (and otherwise problematic) notion is unintelligible.

1.5 Some Closing Remarks I have hardly scratched the surface of revenge in the foregoing remarks. The phenomenon (or, perhaps more accurately, family of phenomena) has in many respects been the fuel behind formal theories of truth, at least in the contemporary period. Despite such a role, a clear understanding of revenge is a pressing and open matter. What, exactly, is revenge? How, if at all, is it a serious problem? Is the problem logical? Is the problem philosophical? And relative to what end, exactly, is the alleged problem a problem? Answers to some of the given questions, I hope, are clear enough in foregoing remarks, but answers—clear answers—to many of the questions remain to be found. Until then, full evaluation of current theories of truth remains out of reach. The hope, however, is that the papers in this volume move matters forward.

Chapter Summaries What follows are brief synopses of the chapters, ordered alphabetically in terms of author(s). The synopses are intended to help the reader ﬁnd chapters of particular interest, rather than serve as discussion of the chapters. Cook. Call a concept C indeﬁnitely extensible just if there’s a rule r such that, when applied to any ‘deﬁnite collection’ of objects falling under C, r yields a new object falling under C. In his ‘Embracing revenge: on the indeﬁnite extendibility of language’, Roy Cook argues that the revenge phenomenon is reason to think that our concept of language, and the associated concept of truth value (or semantic value), is indeﬁnitely extensible. In the end, the revenge phenomenon is a witness to the indeﬁnite extensibility of our language, and, in particular, its ‘semantic values’.

Prolegomenon to Future Revenge / 15 Eklund. In his ‘The liar paradox, expressibility, possible languages’, Matti Eklund focuses on general theses that are standardly tied to the Liar phenomenon. On one hand, there are two related lessons that are sometimes drawn from the Liar’s revenge: namely, radical inexpressibility and (the weaker) inexpressibility. On the other hand, there are two related principles that often make for frustration in the face of revenge: namely, semantic self-sufﬁciency and (what Eklund calls) weak universality. Eklund elucidates the given theses, and focuses attention on inexpressibility and weak universality. Eklund argues that common approaches to such theses may confront difﬁculties from facts governing the space of possible languages, an issue at the heart of Eklund’s essay. Field. In his ‘Solving the paradoxes, escaping revenge’, Hartry Field advances a (paracomplete) theory of truth that, he argues, undermines the ‘received wisdom’ about revenge, where such wisdom, as Field puts it, maintains that ‘any intuitively natural and consistent resolution of a class of semantic paradoxes immediately leads to other paradoxes just as bad as the ﬁrst’. After presenting his own theory of truth (which extends the Kripke approach with a suitable conditional), Field argues that, pace ‘received wisdom’, it is revenge-free. The overall theory and arguments for its revenge-free status have provoked discussion in other chapters (see especially Leitgeb, Priest, Rayo–Welch).²¹ Hofweber. Validity is often thought to be truth-preserving: an inference rule is valid just if truth-preserving.²² Thomas Hofweber, in his ‘Validity, paradox, and the ideal of deductive logic’, argues that two senses of ‘an inference rule is valid just if truthpreserving’ are important to distinguish. One sense is the ‘strict reading’, according to which each and every instance of the given rule is truth-preserving. The other reading is the ‘generic reading’, which, in some sense, is analogous to the claim that bears are dangerous, a claim that is true even though not true of all bears. This distinction, which Hofweber discusses, holds the key to resolving the revenge phenomenon. In ²¹ One issue not discussed is the ideal of ‘exhaustive characterization’, according to which we can truly say (something equivalent to) that all sentences are either True, False, or Whathaveyou (where ‘Whathaveyou’ is a stand in for the predicates used to classify Liars or the like), and do as much in our own language. One might wonder whether the ‘received wisdom’ counts as ‘natural’ only those theories that afford exhaustive characterization, in which case, Field’s argument against ‘received wisdom’ might miss the mark. (Without further clariﬁcation of ‘exhaustive characterization’, I do not intend these remarks as a serious objection, but rather only something for the reader to consider.) ²² I should note that if, as is usual, ‘truth-preserving’ is understood via a conditional, so that α, β is ‘truth-preserving’ just if α → β is true (for some suitable conditional in the language), then many standard theories of transparent truth (i.e., fully intersubstitutable truth) will not have it that valid arguments are truth-preserving. See [2] for some discussion, but also [5] for broader, philosophical issues. This issue, regrettably, is not discussed much in this volume, but it is highly important. Restall’s chapter (Chapter 12) has some direct relevance for the issue, as does Field’s (Chapter 4). Hofweber brieﬂy mentions the issue as it arises for Field’s theory.

16 / JC Beall particular, the Liar’s revenge teaches us that we should abandon the traditional ideal of deductive logic, which requires that our theories be underwritten by rules that are valid in the ‘strict sense’. On the positive side, the Liar’s revenge teaches us that we should embrace the ‘generic’ ideal of deductive logic, which requires only that our rules be ‘generically valid’. Leitgeb. In his ‘On the metatheory of Field’s Solving the Paradoxes, Escaping Revenge’, Hannes Leitgeb argues that whether, in the end, Field’s proposed theory escapes revenge turns on the details of its metatheory. Leitgeb argues that without a clear, explicitly formulated metatheory, the intended interpretation of Field’s proposed truth theory—and, hence, the proposed resolution of paradox—remains unclear. What is ultimately required, Leitgeb argues, is a metatheory that includes a nonclassical set theory for which the logic is the logic of Field’s truth theory.²³ Towards moving matters forward, Leitgeb sketches two target metatheories, a classical and a non-classical one. Leitgeb conjectures that, for reasons he discusses, revenge may emerge for Field’s proposal once a full metatheory is in place. Maudlin. In his ‘Reducing revenge to discomfort’, Tim Maudlin argues that the revenge phenomenon ultimately teaches us something about our normative principles of assertion. As in §1.3 (above), invoking a new category for Liars—say, bugger—seems inevitably to lead to new Liars (e.g. the starred sentence in §1.3 above). While Maudlin maintains that we do need three semantic categories (viz., truth, falsity, and ungroundedness), he argues that we need no more than three. In particular, we may—and should—assert that the ticked sentence in §1.2 above is not true; it’s just that we’ll be bucking the traditional principle according to which only truths are properly assertible. The problem, Maudlin argues, is not with principles of truth (e.g. Release and Capture); the problem is with the traditional principle of assertion.²⁴ On the other hand, Maudlin admits that the revenge phenomenon returns even for his revised principle of assertion (e.g. ‘I am not properly assertible according to Maudlin’s revised principles’). Maudlin argues that this is revenge, but that it is at most a discomfort; it is far from threatening the coherence of truth. Patterson. In his ‘Understanding the liar’, Douglas Patterson advances an ‘inconsistency view’ of the semantic paradoxes in English; however, his view is not a dialetheic ²³ Actually, Leitgeb’s claim needn’t be that the metatheory include a non-classical set theory, but rather that it include a non-classical theory of objects that play the relevant role that sets typically play—e.g. serving as a ‘model’ or etc. ²⁴ I should ﬂag one potential confusion here. Maudlin claims, at least in his fuller work (see references in Chapter 7), that Rule Capture is valid, in the sense that, necessarily, if α is true, then so too is Tr(α). At the same time, the logic governing assertibility is closer to KF, where α ⇒ Tr(α) fails in Rule form (and Conditional form).

Prolegomenon to Future Revenge / 17 view (according to which English is inconsistent, in the sense that some true English sentence has a true negation). Patterson argues that such a view is not that natural languages are inconsistent, but rather that competent speakers of natural languages process such languages in accord with an inconsistent theory. One of Patterson’s principal aims is to show that, perhaps contrary to common thinking, understanding a language can be—and, in the case of English, is—a relation to a false theory. Patterson argues that such an ‘inconsistency view’ is the most promising lesson to draw from the revenge phenomenon. Priest. In his ‘Revenge, Field, and ZF’, Graham Priest does three things. First, Priest characterizes the Liar’s revenge, and carves up three options for dealing with it. Second, Priest directs the discussion towards Field’s chapter (see Chapter 4), and argues that Field’s proposal is not revenge-free, contrary to Field; in particular, it faces an expected problem with the notion of having value 1. (Priest anticipates the immediate thought that, as sketched in §1.4.1 above, he is merely launching a form of ‘too-easy revenge’, conﬂating model-dependent and ‘real’ notions. Priest argues that unless ‘having value 1’ is a real notion, Field has given no reason to think that Field’s proposed logic has anything to do with real validity—i.e. validity in our real language.) Third, Priest argues that the (alleged) troubles facing Field’s proposal are a symptom of deeper revenge in the background theory of ZF, which theory, Priest argues, itself faces a serious revenge-like situation involving V (the cumulative hierarchy): the logic deﬁned by the theory (in terms of models) does not apply to the theory itself, thereby leaving us ‘bereft of a justiﬁcation for reasoning about sets’, as Priest puts it. Rayo and Welch. In their ‘Field on revenge’, Agust´ın Rayo and Philip Welch argue that Field’s allegedly revenge-free truth theory (see Chapter 4) is not really revengefree—or, at least, that its prospects for being revenge-free crucially depend on the outcome of current debates over higher-order languages. Rayo and Welch argue that, just as ‘received wisdom’ maintains, Field’s proposed theory enjoys consistency only in virtue of expressive limitations. In particular, by invoking the appropriate higher-order language, we can explicitly characterize a key semantic notion involved in Field’s proposal: viz., an intended interpretation of L+ , where L+ is the language of Field’s theory (a language enjoying transparent truth and a suitable conditional). Such a notion, as Rayo and Welch argue, plays the Liar’s revenge role: it would generate inconsistency were it expressible in Field’s proposed language.²⁵ Read. In his ‘Bradwardine’s revenge’, Stephen Read discusses a theory of truth proposed by Thomas Bradwardine (who was principally a physicist and theologian in ²⁵ I should be slightly more precise and note that Field (Chapter 4) considers a class of languages (or theories) that enjoy the desiderata of transparent truth and a suitable conditional, and Rayo and Welch direct their remarks against the relevant class.

18 / JC Beall the 1300s). Read shows that Bradwardine’s theory, according to which Liars are not true (because they’d have to be true and not true, which is impossible), is a subtler theory than the later Buridan-like theories that, in effect, reject unrestricted Capture for truth (see §1.1 above). Moreover, the theory, on the surface, as Read argues, seems to promise a revenge-free approach to a whole host of semantic paradoxes. The key for Bradwardine is to distinguish between the claim that the Liar is false from the Liar itself. The propositions appear to be indistinguishable, but they are not. According to Bradwardine, any proposition that ‘says’ of itself that it is false, also ‘says’ of itself that it is true. (As Read points out, this is a subtler thesis than the later Buridanian claim that every claim ‘says’ of itself that it is true.) Restall. In his ‘Curry’s revenge: the costs of non-classical solutions to the paradoxes of self-reference’, Greg Restall discusses the challenges posed by Curry’s paradox to those (non-classical) theories that attempt to preserve Capture and Release, in both Rule and Conditional forms, for truth and related semantic (or logical) notions—e.g. ‘semantical properties’, which serve as the ‘extensions’ of predicates in na¨ıve semantics.²⁶ Restall argues that a Curry conditional is fairly easy to construct unless the language has fairly narrow limits. In particular, a theory that avoids Curry paradox must either reject ‘large disjunctions’, various (otherwise natural) forms of distribution, or the transitivity of entailment. As Restall notes, whatever option is rejected, sound philosophical motivation must accompany the rejection. Scharp. In his wide-ranging ‘Aletheic vengeance’, Kevin Scharp argues that the Liar’s revenge teaches us, among other things, that truth is an inconsistent concept the best theory of which implies that typical truth rules are ‘constitutive’ of truth but nonetheless invalid. Scharp argues that the best (inconsistency) theory of truth takes truth to be a confused concept (in a technical sense), but is a theory that does not use our concept of truth at all. Indeed, Scharp proposes that the proper approach to truth is one that ﬁnds other—non-confused—notions to play the truth role(s). Shapiro. In his ‘Burali-Forti’s revenge’, Stewart Shapiro turns the focus from the Liar paradox to the Burali-Forti paradox, which, he argues, has its own revenge issues. (Using the later von Neumann account, which came after Burali-Forti, the paradox, in short, is that the set of all ordinals satisﬁes all that’s required to be an ordinal, in which case, the successor of , namely + 1, is strictly greater than . But, being ²⁶ Restall doesn’t use the term ‘semantical properties’, but he clearly has this under discussion. (Some philosophers refer to the target entities as ‘na¨ıve sets’, but sets ultimately have little to do with the matter. If we let mathematicians tell us the ‘nature’ of sets—and they’ll likely do so by axiomatizing away Russell problems—we still have to ﬁnd a theory of ‘semantical properties’, the entities that play the familiar role in semantics, namely, those objects ‘expressed’ by any meaningful predicate and ‘exempliﬁed’ by an object just if the given predicate is ‘true of ’ the object.)

Prolegomenon to Future Revenge / 19 itself an ordinal, + 1 must be in , giving the result that < + 1 ≤ , which is impossible.) Shapiro presents the paradox and a variety of ways of dealing with it. He argues that each option faces severe problems, leaving the matter open. Simmons. In his ‘Revenge and context’, Keith Simmons ﬁrst distinguishes between (what he calls) direct revenge and second-order revenge. The former variety is the (what one might call ‘ﬁrst-order’) variety: we already have a stock of semantic terms, and they generate paradox. In particular, as with the ticked sentence in §1.3, one is naturally inclined to classify it as ‘neither true nor false’, but this (at least prima facie) implies untruth, and the paradox remains. One is stuck in direct revenge: an inability to classify the sentence as one thinks it ought to be classiﬁed—but cannot be so classiﬁed, on pain of inconsistency. But, now, one introduces new, technical machinery to deal with the direct revenge problem: one calls the Liar a bugger, or unstable, or whatnot. Second-order revenge emerges with this new machinery, and one is, again, unable to classify the (new) Liars as they ‘ought’ to be (in some sense). Simmons argues that, while there is still work to be done, his ‘singularity theory’ of semantic notions deals not only with direct revenge in a natural way; it also holds the promise of resolving second-order revenge.

Appendix Since many of the chapters in this volume presuppose familiarity with so-called ﬁxedpoint languages, and, in particular, paracomplete languages (see below), this appendix is intended as a user-friendly sketch of the (or a) basic background picture. In particular, I sketch a basic Kripkean picture [11], although I take liberties in the setting up.²⁷ I focus on the non-classical interpretation of Kripke’s (least ﬁxed point) account. My aim is only to give a basic philosophical picture and a sketch of the formal model. I focus on the semantic picture. Philosophical picture One conception of truth has it that truth is entirely transparent, that is, a truth predicate Tr(x) in (and for) our language such that Tr(α) and α are intersubstitutable in all (non-opaque) contexts, for all α in the language. This conception comes with a guiding metaphor, according to which ‘true’ is introduced only for purposes of generalization. Prior to introducing the device, we spoke only the ‘true’-free fragment. ²⁷ This appendix is a very slightly altered version of a section from the much larger [2], which provides more references.

20 / JC Beall (Similarly for other semantic notions/devices, e.g., ‘denotes’, ‘satisﬁes’, ‘true of’, etc.) For simplicity, let us assume that the given ‘semantic-free’ fragment (hence, ‘true’-free fragment) is such that LEM holds.²⁸ Letting L0 be our ‘semantic-free fragment’, we suppose that α ∨ ¬α is true for all α in L0 .²⁹ Indeed, we may suppose that classical semantics—and logic, generally—is entirely appropriate for the fragment L0 . But now we want our generalization-device. How do we want this to work? As above, we want Tr(α) and α to be intersubstitutable for all α. The trouble, of course, is that once ‘is true’ is introduced into the language, various unintended—and, given the role of the device, paradoxical—sentences emerge (e.g. the ticked sentence in §1.2 above).³⁰ The paracomplete idea, of which Kripke’s is the best known, is (in effect) to allow some instances of α ∨ ¬α to ‘fail’.³¹ In particular, if α itself fails to ‘ground out’ in L0 , fails to ‘ﬁnd a value’ by being ultimately equivalent to a sentence in L0 , then the α-instance of LEM should fail. (This is the so-called least ﬁxed point picture.) Kripke illustrated the idea in terms of a learning or teaching process. The guiding principle is that Tr(α) is to be asserted exactly when α is to be asserted. Consider an L0 -sentence that you’re prepared to assert—say, ‘1 + 1 = 2’ or ‘Max is a cat’ or whatever. Heeding the guiding principle, you may then assert that ‘1 + 1 = 2’ and ‘Max is a cat’ are true. In turn, since you are now prepared to assert (1)

‘Max is a cat’ is true

the guiding principle instructs that you may also assert (2)

‘ ‘Max is a cat’ is true’ is true.

And so on. More generally, your learning can be seen as a process of achieving further and further truth-attributions to sentences that ‘ground out’ in L0 . (Similarly for falsity, which is just truth of negation.) Eventually, your competence reﬂects precisely

²⁸ This assumption sets aside the issue of vagueness (and related sorites puzzles). I am setting this aside only for simplicity. The issue of vagueness—or, as some say, ‘indeterminacy’, in general—is quite relevant to some paracomplete approaches to truth. See [4], [13], [21]. ²⁹ This assumption is not essential to Kripke’s account; however, it makes the basic picture much easier to see. ³⁰ With respect to formal languages, the inevitability of such sentences is enshrined in G¨odel’s so-called diagonal lemma. (Even though the result is itself quite signiﬁcant, it is standardly called a lemma because of its role in establishing G¨odel–Tarski indeﬁnability theorems. For user-friendly discussion of the limitative results, and for primary sources, see [18]. For a general discussion of diagonalization, see [19].) ³¹ NB: The sense in which instances of α ∨ ¬α ‘fails’ is modeled by such instances being undesignated (in the formal model). (See ‘Formal model’ below.) How, if at all, such ‘failure’ is expressed in the given language is relevant to ‘revenge’, but I will leave chapters of this volume to discuss that.

Prolegomenon to Future Revenge / 21 the deﬁning intersubstitutivity—and transparency—of truth: that Tr(α) and α are intersubstitutable for all α of the language. But your competence also reﬂects something else: namely, the failure to assert either α or ¬α, for some α in the language. To see the point, think of the above process of ‘further and further truth-attributions’ as a process of writing two (very, very big) books—one, The Truth, the other The False. Think of each stage in the process as completing a ‘chapter’, with chapter zero of each book being empty—this indicating that at the beginning nothing is explicitly recorded as true (or, derivately, false). Concentrate just on the process of recording atomics in The Truth. When you were ﬁrst learning, you scanned L0 (semantic-free fragment) for the true (atomic) sentences, the sentences you were prepared to assert. Chapter one of The Truth comprises the results of your search—sentences such as ‘Max is a cat’ and the like. In other words, letting ‘I(t)’ abbreviate the denotation of t, chapter one of The Truth contains all of those atomics α(t) such that I(t) exempliﬁes α, a ‘fact’ that would’ve been recorded in chapter zero had chapter zero recorded the true semantic-free sentences. (For simplicity, if α(t) is an L0 -atomic such that I(t) exempliﬁes α, then we’ll say that I(t) exempliﬁes α according to chapter zero. In the case of ‘Max is a cat’, chapter zero has it that Max exempliﬁes cathood, even though neither ‘Max is a cat’ nor anything else appears in chapter zero.) In the other book, The False, chapter zero is similarly empty; however, like chapter zero of The Truth, the sentences that would go into The False’s chapter zero are those (atomic) L0 -sentences that, according to the world (as it were), are false—e.g., ‘1 + 1 = 3’, ‘Max is a dog’, or the like.³² If α(t) is a false L0 -atomic, we’ll say that according to chapter zero, I(t) exempliﬁes ¬α (even though, as above, chapter zero explicitly records nothing at all). In turn, chapter one of The False contains all of those atomics α(t) such that, according to chapter zero, I(t) exempliﬁes ¬α (i.e., the L0 -atomics that are false, even though you wouldn’t say as much at this stage). And now the writing (of atomics) continues: chapter two of The Truth comprises ‘ﬁrst-degree truth-attributions’ and atomics α(t) such that, as above, I(t) exempliﬁes α according to chapter one, sentences like (1) and ‘Max is a cat’. In turn, chapter three of The Truth comprises ‘second-degree’ attributions, such as (2), and atomics α(t) such that (as it were) t is α according to chapter two. And so on, and similarly for The False. In general, your writing-project exhibits a pattern. Where Ii (Tr) is chapter i of The Truth, the pattern runs thus: Ii+1 (Tr) = Ii (Tr) ∪ {α(t) : α(t)is an atomic and I(t) exempliﬁes α according to Ii (Tr)} ³² For convenience, we’ll also put non-sentences into The False. Putting non-sentences into The False is not essential to Kripke’s construction, but it makes things easier. Obviously, one can’t write a cat but, for present purposes, one can think of The False as a special book that comes equipped with attached nets (wherein non-sentences go), a net for each chapter.

22 / JC Beall Let S comprise all sentences of the language. With respect to The False book, the pattern of your writing (with respect to atomics) looks thus: Ii+1 (F) = Ii (F) ∪ {α(t) : α(t) is an atomic and I(t) ∈ /S or I(t) exempliﬁes ¬α according to Ii (Tr)} So goes the basic process for atomics. But what about compound (molecular) sentences? The details are sketched below (see ‘Formal model’), but for now the basic idea is as follows (here skipping the relativizing to chapters). With respect to negations, ¬α goes into The True just when α goes into The False. (Otherwise, neither α nor ¬α ﬁnds a place in either book.) With respect to conjunctions, α ∧ β goes into The False if either α or β goes into The False, and it goes into The True just if both α and β go into The True. (Otherwise, α ∧ β ﬁnds a place in neither book.) The case of disjunctions is dual, and the quantiﬁers may be treated as ‘generalized conjunction’ (universal) and ‘generalized disjunction’ (existential). This approach to compound sentences reﬂects the so-called Strong Kleene scheme, which is given below (see ‘Formal model’). Does every sentence eventually ﬁnd a place in one book or other? No. Consider an atomic sentence λ, like the ticked sentence in §1.2, equivalent to ¬Tr(λ). In order to get λ into The True book, there’d have to be some chapter in which it appears. λ doesn’t appear in chapter zero, since nothing does. Moreover, λ doesn’t exemplify anything ‘according to chapter zero’, since chapter zero concerns only the L0 -sentences (and λ isn’t one of those). What about chapter one? In order for λ to appear in chapter one, λ would have to be in chapter zero or be such that λ exempliﬁes ¬Tr(x) according to chapter zero. But for reasons just given, λ satisﬁes neither disjunct, and so doesn’t appear in chapter one. The same is evident for chapter two, chapter three, and so on. Moreover, the same reasoning indicates that λ doesn’t appear in The False book. In general, Liar-like sentences such as the ticked sentence in §1.2 will ﬁnd a place in one of our books only if it ﬁnds a place in one of the chapters Ii (Tr) or Ii (F). But the ticked sentence will ﬁnd a place in Ii (Tr) or Ii (F) only if it ﬁnds a place in Ii−1 (Tr) or Ii−1 (F). But, again, the ticked sentence will ﬁnd a place in Ii−1 (Tr) or Ii−1 (F) only if it ﬁnds a place in Ii−2 (Tr) or Ii−2 (F). And so on. But, then, since I0 (Tr) and I0 (F) are both empty, and since—by our stipulation—something exempliﬁes a property according to I0 (Tr) only if the property is a non-semantic one (the predicate is in L0 ), the ticked sentence (or the like) fails to ﬁnd a place in either book. Such a sentence, according to Kripke, is not only ungrounded, since it ﬁnds a place in neither book, but also paradoxical—it couldn’t ﬁnd a place in either book.³³ ³³ The force of couldn’t here is made precise by the full semantics, but for present purposes one can think of couldn’t along the lines of on pain of (negation-) inconsistency or, for that matter, on pain of being in both books (something impossible, on the current framework).

Prolegomenon to Future Revenge / 23 So goes the basic philosophical picture. What was wanted was an account of how, despite the existence of Liars, we could have a fully transparent truth predicate in the language—and do so without triviality (or, in Kripke’s case, inconsistency). The foregoing picture suggests an answer, at least if we eventually have a chapter Ii (Tr) such that Tr(α) is in Ii (Tr) if and only if α is in Ii (Tr), and similarly a chapter for The False. What Kripke (and, independently, Martin–Woodruff) showed is that, provided our ‘writing process’ follows the right sort of scheme (in effect, a logic weaker than classical), our books will contain such target chapters, and in that respect our language can enjoy a (non-trivial, indeed consistent) transparent truth predicate. Making the philosophical picture more precise is the job of formal, philosophical modeling, to which I now brieﬂy (and somewhat informally) turn.

Formal model For present purposes, I focus on what is known as Kripke’s ‘least ﬁxed point’ model (with empty ground model). I leave proofs to cited works (all of which are readily available), and try to say just enough to see how the formal picture goes. Following standard practice, we can think of an interpreted language L as a triple L, M, σ , where L is the syntax (the relevant syntactical information), M is an ‘interpretation’ or ‘model’ that provides interpretations to the non-logical constants (names, function-symbols, predicates), and σ is a ‘semantic scheme’ or ‘valuation scheme’ that, in effect, provides interpretations (semantic values) to compound sentences.³⁴ Consider, for example, familiar classical languages, where the set V of ‘semantic values’ is {1, 0}. In classical languages, M = D, I , with D our (non-empty) domain and I an ‘interpretation-function’ that assigns to each name an element of D (the denotation of the name), assigns to each n-ary function-symbol an element of Dn −→ D, that is, an n-ary function from Dn into D, and assigns to each n-ary predicate an element of Dn −→ V, a function—sometimes thought of as the intension of the predicate—taking n-tuples of D and yielding a ‘semantic value’ (a ‘truth value’). The extension of an n-ary predicate F (intuitively, the set of things of which F is true) contains all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 1. The classical valuation scheme τ (for Tarski) is the familiar one according to which a negation is true (in a given model) exactly when its negatum is false (in the given model), a disjunction is true (in a model) iff one of the disjuncts is true (in the model), and existential sentences are treated as generalized disjunctions.³⁵ ³⁴ For present purposes, a semantic scheme or valuation scheme σ is simply some general deﬁnition of truth (falsity) in a model. For more involved discussion of semantic schemes, see [8]. ³⁵ I assume familiarity with the basic classical picture, including ‘true in L’ and so on. To make things easier, I will sometimes assume that we’ve moved to models in which everything in the domain has a name, and otherwise I’ll assume familiarity with standard accounts of ‘satisﬁes α(x) in L’.

24 / JC Beall Classical languages (with suitably resourceful L) cannot have their own transparent truth predicate. Paracomplete languages reject the ‘exhaustive’ feature implicit in classical languages: namely, that a sentence or its negation is true, for all sentences. The standard way of formalizing paracomplete languages expands the interpretation of predicates. Recall that in your ‘writing process’ some sentences (e.g. Liars) found a place in neither book. We need to make room for such sentences, and we can expand our semantic values V to do so; we can let V = {1, 12 , 0}, letting the middle value represent (for ‘modeling’ purposes) the status of sentences that found a place in neither book. Generalizing (but, now, straining) the metaphor, we can think of all n-ary predicates as tied to two such ‘big books’, one recording the objects of which the predicate is true, the other the objects of which it is false. On this picture, the extension of a predicate F remains as per the classical (containing all n-tuples of which the predicate is true), but we now also acknowledge an antiextension, this comprising all n-tuples of which the predicate is false. This broader picture of predicates enjoys the classical picture as a special case: namely, where we stipulate that, for any predicate, the extension and antiextension are jointly exhaustive (the union of the two equals the domain) and, of course, exclusive (the intersection of the two is empty). Concentrating on the so-called Strong Kleene account [36],³⁶ the formal story runs as follows. We expand V, as above, to be {1, 12 , 0}, and so our language Lκ = L, M, κ is now a so-called three-valued language (because it uses three semantic values).³⁷ Our designated values—intuitively, the values in terms of which validity or consequence is deﬁned—are a subset of our semantic values; in the Strong Kleene case, there is exactly one designated element, namely 1. A (Strong Kleene) model M = D, I is much as before, with I doing exactly what it did in the classical case except that I now assigns to n-ary predicates elements of Dn −→ {1, 12 , 0}, since V = {1, 12 , 0}. Accordingly, the ‘intensions’ of our paracomplete (Strong Kleene) predicates have three options: 1, 12 , and 0. What about extensions? As above, we want to treat predicates not just in terms of extensions (as in the ³⁶ This is one of the paracomplete languages for which Kripke proved his deﬁnability result. Martin–Woodruff proved a special case of Kripke’s general ‘ﬁxed point’ result, namely, the case for so-called ‘maximal ﬁxed points’ of the Weak Kleene scheme, or Weak Kleene languages. ³⁷ Kripke [11] made much of emphasizing that ‘the third value’ is not to be understood as a third truth value or anything else other than ‘undeﬁned’ (along the lines of Kleene’s original work [10] ). I will not make much of this here, although what to make of semantic values that appear in one’s formal account is an important, philosophical issue. (Note that if one wants to avoid a three-valued language, one can let V = {1, 0} and proceed to construct a Kleene-language by using partial functions (hence, the standard terminology ‘partial predicates’) for interpretations. I think that this is ultimately merely terminological, but I won’t dwell on the matter here.

Prolegomenon to Future Revenge / 25 classical languages) but also antiextensions. The extension of an n-ary predicate F, just as before, comprises all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 1. (Again, intuitively, this remains the set of objects of which F is true.) The antiextension, in turn, comprises all n-tuples a1 , . . . , an of D such that I(F)(a1 , . . . , an ) = 0. (Again, intuitively, this is the set of objects of which F is false.) Of course, as intended, an interpretation might fail to put x in either the extension or antiextension of F. In that case, we say (in our ‘metalanguage’) that, relative to the model, F is undeﬁned for x.³⁸ Letting F+ and F− be the extension and antiextension of F, respectively, it is easy to see that, as noted above, classical languages are a special case of (Strong Kleene) paracomplete languages. Paracomplete languages typically eschew inconsistency, and so typically demand that F+ ∩ F− = ∅, in other words, that nothing is in both the extension and antiextension of any predicate. In this way, paracomplete languages typically agree with classical languages. The difference, of course, is that paracomplete languages do not demand that F+ ∪ F− = D for all predicates F. But paracomplete languages allow for such ‘exhaustive constraints’, and in that respect can enjoy classical languages as a special case. To see the close relation between classical languages and Strong Kleene, notice that κ, the Strong Kleene valuation-scheme, runs as follows (here treating only ¬, ∨, and ∃). Where VM (α) is the semantic value of α in M (and, for simplicity, letting each object in the domain name itself), and, for purposes of specifying scheme κ, treating V as standardly (linearly) ordered: K1. VM (¬α) = 1 − VM (α). K2. VM (α ∨ β) = max(VM (α), VM (β)). K3. VM (∃x α(x)) = max{VM (α(t/x)) : for all t ∈ D}. The extent to which classical logic is an extension of a given paracomplete logic depends on the semantic scheme of the language.³⁹ Since κ, as above, is entirely in keeping with the classical scheme except for ‘adding an extra possibility’, it is clear that every classical interpretation is a Strong Kleene-interpretation (but not vice versa).⁴⁰ ³⁸ A common way of speaking is to say that, for example, F(t) is ‘gappy’ with respect to I(t). This terminology is appropriate if one is clear on the relation between one’s formal model and the target notions that the model is intended to serve (in one respect or other), but the terminology can also be confusing, since, e.g., in the current Strong Kleene language, one cannot truly assert of any α that α is ‘gappy’, i.e. ¬Tr(α) ∧ ¬Tr(¬α). (This issue arises in various chapters in the current volume.) ³⁹ Here, perhaps not altogether appropriately, I am privileging model theory over proof theory, thinking of ‘logic’ as the semantic consequence relation that falls out of the semantics. This is in keeping with the elementary aims of the essay, even though (admittedly) it blurs over a lot of philosophical and logical issues. ⁴⁰ Note that in classical languages, VM (A) ∈ {1, 0} for any A, and the familiar classical clauses on connectives are simply (K1)–(K3).

26 / JC Beall Let us say that an interpretation veriﬁes a sentence α iff α is designated (in this case, assigned 1) on that interpretation, and that an interpretation veriﬁes a set of sentences iff it veriﬁes every element of . We deﬁne semantic consequence in familiar terms: α is a consequence of iff every interpretation that veriﬁes also veriﬁes α. I will use ‘SK ’ for the Strong Kleene consequence relation, so understood. Let us say that a sentence α is logically true in Lκ exactly if ∅ Sκ α, that is, iff α is designated (assigned 1) in every model. A remarkable feature of Lκ is that there are no logical truths. To see this, just consider an interpretation that assigns 12 to every atomic, in which case, as an induction will show, every sentence is assigned 12 on that interpretation. Hence, there’s some interpretation in which no sentence is designated, and hence no sentence designated on all interpretations. A fortiori, LEM fails in Strong Kleene languages.⁴¹ And now an answer to one guiding question becomes apparent. What we want is a model of how our language can be non-trivial (indeed, consistent) while containing both a transparent truth predicate and Liar-like sentences. In large part, the answer is that our language is (in relevant respects) along Strong Kleene lines, that the logic is weaker than classical logic. Such a language, as Kripke showed, can contain its own (transparent) truth predicate. The construction runs (in effect) along the lines of the ‘big books’ picture. For simplicity, let Lκ be a classical (but nonetheless Strong Kleene) language such that L (the basic syntax, etc.) is free of semantic terms but has the resources to describe its given syntax—including, among other things, having a name α for each sentence α. (In other words, I assigns to each n-ary predicate an element of Dn −→ {1, 0}, even though the values V of Lκ also contain 12 .) What we want to do is move to a richer language the syntax Lt of which contains Tr(x), a unary predicate intended to be a transparent truth predicate for the enriched language. For simplicity, assume that the domain D of Lκ contains all sentences of Lt .⁴² Think, brieﬂy, about the ‘big books’ picture. One can think of each successive ‘chapter’ as a language that expands one’s ofﬁcial record of what is true (false). More formally, one can think of each such ‘chapter’ of both books as the extension and antiextension of ‘true’, with each such chapter expanding the interpretation of ‘true’. Intuitively (with slight qualiﬁcations about chapters zero), one can think of Ii+1 (Tr) as explicitly recording what is true according to chapter Ii (Tr). The goal, of course, is to ﬁnd a ‘chapter’ at which we have Ii+1 (Tr) = Ii (Tr), a ‘ﬁxed point’ at which anything ⁴¹ This is not to say, of course, that one can’t have a Strong Kleene—or, in general, paracomplete—language some proper fragment of which is such that α ∨ ¬α holds for all α in the proper fragment. (One might, e.g., stipulate that arithmetic is such that α ∨ ¬α holds.) ⁴² This is usually put (more precisely) as that the domain contains the G¨odel-codes of all such sentences, but for present purposes I will skip over the mathematical details.

Prolegomenon to Future Revenge / 27 true in the language is fully recorded in the given chapter—one needn’t go further. Thinking of the various ‘chapters’ as languages, each with a richer interpretation of ‘true’, one can think of the ‘ﬁxed chapter’ as a language that, ﬁnally, has a transparent truth predicate for itself. Returning to the construction at hand, we have our Strong Kleene (but classical) ‘ground language’ Lκ that we now expand to Ltκ , the syntax of which includes that of Lκ but also has Tr(x) (and the resulting sentences formable therefrom). We want the new language to ‘expand’ the ground language, and we want the former to have a model that differs from the latter only in that it assigns an interpretation to Tr(x). For present purposes, we let It , the interpretation function in Ltκ , assign (∅, ∅) to Tr(x), where (∅, ∅) is the function that assigns 12 to each element of Dt . (Hence, the extension and antiextension of Tr(x) in Ltκ are both empty.) This is the formal analogue of ‘chapter zero’. The crucial question, of course, concerns further expansion. How do we expand the interpretation of Tr(x)? How do we move to ‘other chapters’? How, in short, do we eventually reach a ‘chapter’ or language in which we have a transparent truth predicate for the whole given language? This is the role of Kripke’s ‘jump operator’. What we want, of course, are ‘increasingly informative’ interpretations (Ti + , Ti − ) of Tr(x), but interpretations that not only ‘expand’ the previous interpretations but also preserve what has already been interpreted. If α is true according to chapter i, then we want as much preserved: that α remain true according to chapter i + 1. This is the role of the ‘jump operator’, a role that is achievable given the so-called monotonicity of Strong Kleene valuation scheme κ.⁴³ The role of the jump operator is to eventually ‘jump’ through successive interpretations (chapters, languages) Ii (Tr) and land on one that serves the role of transparent truth—serves as an interpretation of ‘is true’. As above, letting Ii (Tr) be a function (Ti + , Ti − ) yielding ‘both chapters i’, the goal is to eventually ‘jump’ upon an interpretation (Ti + , Ti − ) such that + − (Ti + , Ti − ) = (Ti+1 , Ti+1 ). Focusing on the ‘least such point’ in the Strong Kleene setting, Kripke’s construction proceeds as above. We begin at stage 0 at which Tr(x) is interpreted as (∅, ∅), and we ⁴³ Monotonicity is the crucial ingredient in Kripke’s (similarly, Martin–Woodruff ’s) general result. Let M and M be paracomplete (partial) models for (uninterpreted) L. Let FM+ be the extension of F in M, and similarly FM+ for M . (Similarly for antiextension.) Then M extends M iff the models have the same domain, agree on interpretations of names and function signs, and FM+ ⊆ FM+ and FM− ⊆ FM− for all predicates F that M and M interpret. (In other words, M doesn’t change M’s interpretation; it simply interprets whatever, if anything, M left uninterpreted.) Monotonicity property: A semantic (valuation) scheme σ is monotone iff for any α that is interpreted by both models, α’s being designated in M implies its being designated in M whenever M extends M. So, the montonicity property of a scheme ensures that it ‘preserves truth (falsity)’ of ‘prior interpretations’ in the desired fashion.

28 / JC Beall + − , Ti+1 ) deﬁne a ‘jump operator’ on such interpretations:⁴⁴ Tr(x) is interpreted as (Ti+1 + + − at stage i + 1 if interpreted as (Ti , Ti ) at the preceding stage i, where, note well, Ti+1 comprises the sentences that are true (designated) at the preceding stage (chapter, − language) i, and Ti+1 the false sentences (and, for simplicity, non-sentences) at i. Accordingly, we deﬁne the ‘jump operator’ JSK thus:⁴⁵ + − , Ti+1 ) JSK (Ti + , Ti − ) = (Ti+1

The jump operator yields a sequence of richer and richer interpretations that ‘preserve prior information’ (given monotonicity), a process that can be extended into the transﬁnite to yield a sequence (T0+ , T0− ), (T1+ , T1− ), . . . , (Tγ+ , Tγ− ), . . . deﬁned (via transﬁnite recursion) thus:⁴⁶ Jb. Base. (T0+ , T0− ) = (∅, ∅). Js. Successor. (Tγ++1 , Tγ−+1 ) = JSK ((Tγ+ , Tγ− )). Jl. Limit. For limit stages, we collect up by unionizing the prior stages: + − + − Tε , Tε (Tλ , Tλ ) = ε 1, ‘Pn (x)’ is the nth ‘pathogical’ predicate.

40 / Roy T. Cook the newest truth-value, and we also add a new conditional. Each conditional is a better approximation of the ‘ideal’ but unattainable conditional which satisﬁes all of the standard axioms and inference rules, including modus ponens and conditional proof.¹³ We assume that G¨odel coding is carried out such that, for any sentence in Lα , its G¨odel code (written < >) relative to Lα , if it has one, is the same as its G¨odel code relative to Lβ for any β > α. Note that there will be languages Lβ such that some sentences of Lβ will not receive G¨odel codes (since if β is large enough, Lβ is uncountable, yet we only have a countable inﬁnity of numerals to serve as codes). We construct a model for each language as follows: The model of L0 is just any classical model of set theory < D, I > where D is the domain and I is an interpretation function mapping sentences onto {t, f}. For each Lα (α ≥ 0) we construct an n+3-valued Kripke-style ﬁxed point model < D, Iα > recursively as follows. The truth-values (i.e. the range of Iα ) are: {t, f, n} ∪ {pβ : β ≤ α}

where t is the value true, f is false, pβ is the β th ‘pathological’ value, and n is a placeholder value given to sentences involving the application of a semantic predicate to a numeral that either is not the G¨odel code of any sentence or is the G¨odel code of a sentence which has not yet been added to the language (i.e. n is the value sentences get when they have not yet received a ‘real’ value). In other words, we do not assign a legitimate truth-value to the sentence ‘5 is true’ (or its negation) if 5 is not the G¨odel code of any sentence (in any of our languages), nor does the sentence ‘57568489 is true’ obtain a legitimate truth-value in L3 if the sentence coded by 57568489 is a sentence which contains the predicate ‘P3 (x)’. Given an assignment of semantic values to the atomic formulas in a language Lα , the interpretation of the logical connectives is determined as follows: I( ∧ )

= min{I(), I()} relative to the ordering:¹⁴ n < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < f < t

I( ∨ )

= max{I(), I()} relative to the ordering: f < t < p1 < p2 < . . . < pβ < pβ+1 < . . . < pα < n

¹³ The fact that one cannot have a wholly satisfactory conditional and a wholly satisfactory truth predicate in the same language is well known, and, like Field (2003), we opt here for retaining the na¨ıve notion of truth, and settling for an account of the conditional (and thus the biconditional) which does not provide everything that we might wish for. In the ﬁnal section of the chapter the present approach, which posits a never-ending series of conditionals that provide better and better approximations of the ‘ideal’ conditional, is contrasted with Field’s approach, where a single, rather complicated, conditional is constructed in the hopes of getting the closest approximation possible. ¹⁴ The clauses for the binary connectives are just a generalization of the weak Kleene scheme (Kleene (1952)). Interestingly, the strong Kleene scheme will not work in this framework.

Embracing Revenge / 41 I(¬) =

For β < α: I( →β )

f if I() = t t if I() = f I() otherwise. = n

if I() = n or I() = n

max{I(), I()} relative to the ordering: f < t < p1 < p2 < . . . < pβ < pβ+1 < . . . < pα < n if I() = n and I() = n and I() = pδ or I() = pδ where δ > β t

if I() = n and I() = n and for any δ > β, I() = pδ and I() = pδ and I() ≤ I() relative to the ordering: f < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < t

f

otherwise.

The universal and existential quantiﬁers are treated as generalized versions of conjunction and disjunction respectively. We now construct a sequence of models < D, Iα β > via transﬁnite recursion (the ﬁnal model < D, Iα > for each language Lα will be the least ﬁxed point in this sequence). Given the clauses above, we need merely to specify the interpretation of atomic sentences in each model in the series.: For any atomic sentence of Lα+1 : Base: Iα+1 0 ()

=

n Iα () pα+1

if = T(n) or = F(n) or = Pi (n) where n is not the G¨odel code of a sentence in Lα+1 if is a sentence of Lα . otherwise.

Successor: If = T(< >) then: Iα+1 β+1 ()

=

Iα+1 β ()

If = F(< >) then: Iα+1 β+1 () = t = f = Iα+1 β ()

if Iα+1 β () ifIα+1 β () otherwise

= =

f t

= =

pδ t

If = Pδ (< >) for some δ < α, then: Iα+1 β+1 ()

= = =

t f Iα+1 β ()

if Iα+1 β () ifIα+1 β () otherwise

42 / Roy T. Cook If = T(< >) and = F(< >) and = Pδ (< >) then: Iα+1 β+1 ()

=

Iα+1 β ()

Limit: = max{Iα+1 β (): β < γ } relative to the partial ordering: For any δ : n < pγ , pγ < t, pγ < f, and pγ < pδ iff δ < γ ¹⁵

Iα+1 γ ()

All operators, quantiﬁers, etc. in the above scheme are monotonic with respect to the partial ordering of truth-values utilized in the limit case above. Thus, the sequence of models described above will have a ﬁxed point—that is—a model < D, Iα+1 β > such that: < D, Iα+1 β > = < D, Iα+1 β+1 > < D, Iα+1 >, our model of Lα+1 , is just the minimal such ﬁxed point (for the mathematical details of such ﬁxed point constructions see Kripke (1972) or Fitting (1986)). We extend the series of models into the transﬁnite by adding the additional clause handling atomic sentences in the base case for Lγ where γ a limit: Base: Iγ 0 ()

=

max{Iβ (): β < γ } relative to the partial ordering: For any δ : n < pγ , pγ < t, pγ < f, and pγ < pδ iff δ < γ

Applying the same clauses as before for successor and limit superscripts, we are again guaranteed the existence of a ﬁxed point. The minimal such ﬁxed point is our model < D, Iγ >, of Lγ . A few observations regarding the formal theory are in order (proofs have been omitted in order to keep this chapter concise and accessible): [1] In any language Lα where α ≥ 1, T(< >) and are interchangeable salve valutate (i.e. the truth predicate is ‘transparent’, that is, T(< >) and always receive the same semantic value in < D, Iα >). [2] For any languages Lα and Lα+β , and any sentence in Lα (and thus in Lα+β ), Iα () = Iα+β (). In other words, once a sentence is expressible in one of our languages, and it receives a genuine truth-value (i.e. one other than n) in the model of that language, its truth-value does not change in models of extensions of that language.¹⁶ ¹⁵ In addition to guaranteeing the existence of a ﬁxed point, the monotonicity of our valuation scheme relative to this ordering guarantees that the maximum referred to in this clause exists. ¹⁶ A critic might be tempted to object to the present view on the ground that, as we pass from one language to another, the meaning of the truth predicate changes (which might, in turn, push us towards a view more like that sketched in Williamson (1998)). Fact [2] goes some ways towards assuaging such worries. The extension of the truth predicate does change from one language to the next, but only by adding new instances—no sentence can ‘change’ from true to some other value. Such changes in extension do not imply a change in meaning, however, any more than the production of the present volume changed the meaning of ‘published anthology’.

Embracing Revenge / 43 [3] No model < D, Iα > makes true all instances of the T-schema: T(< >) ↔ on any deﬁnition of the biconditional.¹⁷ Given [1], however, this is clearly not due to a fault in the truth predicate, but a failure to express a suitable conditional (and thus biconditional). The seriousness of this omission is lessened by the fact that each of languages makes true all instances of the T-schema for sentences of earlier languages. In other words, letting: ↔α

=df

( →α ) ∧ ( →α )

Lα+1 makes true all instances of the T-schema: T(< >) ↔α where is a sentence of Lα . [4] Given [3], it is not surprising that we can express, within Lα+1 , the complete semantic theory for Lα .¹⁸ For example, letting ‘Sentα (x)’ abbreviate the unary arithmetic predicate that holds of n if and only if n is the G¨odel number of a sentence in Lα , and letting ‘Conj(x, y, z) abbreviate the ternary arithmetic predicate expressing ‘x is the code of the conjunction of the sentences that y and z code’, we can, within L3 , formulate the clause for conjunction in L2 as follows. First, we need to deﬁne slightly altered versions of our truth and pathological predicates: T(α) (< >) F(α) (< >) P(α,β) (< >)

=df =df =df

T(< >) ↔α (T(< >) ↔α T(< >))¹⁹ F(< >) ↔α (T(< >) ↔α T(< >)) Pβ (< >) ↔α (T(< >) ↔α T(< >))

Intuitively, P(α,β) () gets the value t if gets the α th pathological value, gets the value pδ if δ > α and gets the value pδ , gets the value n if get the value n, and gets f otherwise (similarly for T(α) (< >) and F(α) (< >)). We can then formulate the clause for conjunction as: (∀x)(∀y)(∀z)((Sent2 (x) ∧ Conj(x, y, z)) →2 ((T(2) (x) ↔2 (T(2) (y) ∧ T(2) (z)))∧ (P(2,2) (x) ↔2 (P(2,2) (y) ∨ P(2,2) (z)))∧ ¹⁷ This is because none of our conditionals validates every instance of ( → ). For any sentence in Lα , however, ( →α+1 ) is a theorem (in Lα+1 ). ¹⁸ In order to describe the semantics for languages Lα , where α is inﬁnite, we make us of the fact that Lα+1 contains a satisfaction predicate Sat(x, y) expressing the relation ‘x is the code of a predicate satisﬁed by y’ and a predicate Pth (x,, y) expressing the relation ‘x is the code of the yth pathological predicate’. ¹⁹ T(α) is the α-level ‘strong’ truth predicate, which, when applied to a sentence , receives true if is true, receives the α + 1th pathological value if does, and is false otherwise. Notice that these predicates, unlike the ofﬁcial truth predicate, are not transparent. If is the simple Liar (and so receives the ﬁrst pathological value), then T(< >) also recieves the ﬁrst pathological value, while T(α+1) (< >) is false.

44 / Roy T. Cook (P(2,1) (x) ↔2 ((P(2,1) (y) ∨ P(2,1) (z)) ∧ (¬P(2,2) (y) ∧ ¬P(2,2) (z))))∧ (F(2) (x) ↔2 ((F(2) (y) ∧ F(2) (z)) ∧ (¬P(2,1) (y) ∧ ¬P(2,1) (z) ∧ ¬P(2,2) (y) ∧ ¬P(2,2) (z))))))

While the formal theory, as sketched, accurately models the philosophical picture as described in previous sections, there are a number of ways in which we could modify the details. Two are worth mentioning here. The ﬁrst way is to take, instead of the minimal ﬁxed point, the maximal intrinsic ﬁxed point (see Kripke (1972)). While studying the properties of various sorts of ﬁxed points in this iterated version of Kripke’s construction would no doubt provide us with a better understanding of the general framework as a whole, none of [1] through [4] above depend taking the minimal ﬁxed point (we only need the weaker claim that we have taken some ﬁxed point). So the choice between the minimal ﬁxed point and the maximal intrinsic ﬁxed point (or between these and some other ﬁxed point) will depend on one’s attitude towards ungrounded but determinate sentences such as: D:

D is either true or false.

The second way in which we might alter the present account is by adding more than one pathological truth-value at each extension of the language. Motivation for this idea is not hard to ﬁnd. Recall that in the ﬁrst section we saw that the Liar required a third truth-value because it fell into the category of sentences where ‘what is says is the case if and only if what it says is not the case.’ Given that truth and falsity are no longer being treated as exhaustive, this status is distinct from truth and falsity themselves. But it is not obvious that the status of the Liar, looked at in this way, is the same as that of other problematic ungrounded sentences. For example, the Truth Teller ‘This sentence is true’ falls into the ‘what it says is the case if and only if what it says is the case’ category, and the determinate sentence D above falls into the ‘what it says is the case if and only if either what it says is the case or what it says is not the case’. Given that truth and falsity are not exhaustive (and thus, we cannot assume that, for every sentence, either what it says is the case, or what it says is not the case), it is not obvious that these two categories are identical to the status of the Liar sentence.²⁰ Of course, the last paragraph is a bit rough and loose. Nevertheless, it does indicate that there might be good reasons for exploring the idea that extending our language by adding a new semantic predicate might introduce more than one new truth-value. Although adding such additional truth-values at each stage in the above construction ²⁰ In future work I intend on ﬂeshing out this idea by treating the different truth-values as intimately connected to (and thus, within the semantics, representable by) the directed graphs associated with the different patterns of referential dependency exhibited by these different ‘sorts’ of pathological sentence. For an initial stab in this direction, one can consult Cook (2004).

Embracing Revenge / 45 greatly complicates an already complicated picture, the framework could be extended in this manner without greatly affecting the ﬁnal shape of the account (in particular, versions of [1] through [4] will still hold). At this point, however, we will rest content with merely having pointed out the possibility of such extensions of the basic picture.

2.4 Dodging Revenge On the account just sketched, the Revenge Problem is not a problem, but instead provides the crucial insight motivating the account: Given a language L, if we can completely describe the semantics of L, then we have (knowingly or not) extended our language to a new language L’ (where such an extension involves not only adding to the set of wffs, but adding an additional truth-value that those wffs can receive). The semantic theory of L, however, as expressed in L’ is not sufﬁent for L’ itself. In order to describe its semantics, we must extend the language again (and as a result extend the collection of truth-values as well). And so on. In hacker-speak, the Revenge Problem is no longer a bug—it is now a feature, exemplifying the indeﬁnite extensibility of the concepts language, statement, and truth-value. One advantage of this view is that there are no real limitations on what can be expressed. At any stage in the series of languages, we are free to extend the language we presently speak in order to describe all the facts (semantic and otherwise), although in doing so we might introduce additional truth-values (and thus enable ourselves to ‘access’ more facts). But further extensions will allow us to speak of those as well, and so on. Of course, given this lack of expressive restriction, a critic might be forgiven for thinking that Revenge is likely to reappear. After all, if we can, at any stage in the game, extend our semantic resources in order to describe all of the (currently accessible) facts, then what is to stop us from extending the language so as to contain all possible semantic predicates at once?²¹ In fact, doesn’t the language used in the previous section, in describing the formal theory, amount to a language that does just that? But once we have allowed ourselves to extend the language in this way, there is nothing to stop us from forming the Super-Strengthened Liar: SUP: This sentence is either false or has one of the pathological values. ²¹ The comments of the next few paragraphs also explain why we cannot at any point add operators such as ‘exclusion’ negation to our language (i.e. a connective ∗ such that ∗ (P) is false if P is true, and true otherwise). Thanks go to Stewart Shapiro for pointing this out.

46 / Roy T. Cook It is not difﬁcult to see that such a sentence would be inconsistent in our formal theory, were it expressible. Fortunately for our account, however, there are good reasons for thinking that, contrary to appearances, SUP is not expressible in any language.²² Actually, there are two separate claims involved in this global version of the Revenge Problem which need to be distinguished, so each can be dealt with in turn. The ﬁrst is that there is nothing to stop us from extending our language so that we can talk about all possible truth-values at once. The second claim is that we have already done this, in the previous section when formulating our formal account of the semantics. Regarding the ﬁrst claim, the initial answer is easy: there are reasons why we cannot formulate a language that contains every possible truth-value. Anytime we attempt to add a predicate ‘is a pathological truth-value’ to a language Lα , we end up extending the language, forming a new language Lα+β . The semantics of Lα+β will require at least one more truth-value than that of Lα , and our new predicate will only be satisﬁed (i.e. receive the value t) by (G¨odel codes of) sentences which receive values p1 , p2 , . . . pδ for some δ < α + β. This is the very lesson that the Revenge Problem teaches us: That any attempt to construct a language that allows us to talk about ‘all’ semantic values (in the sense of containing predicates for each value) brings a new pathological value into the picture, one which is not described by the language in question. In other words, the indeﬁnite extensibility of the concept statement prevents us from every being able to say things about all statements (or, derivatively, about all truth-values). At this point, however, we run into the second aspect of the objection. How can we claim that we can never talk about all truth-values at once, so the criticism goes, when we obviously quantiﬁed over all of them in the formal account given in the previous section? The easy answer to this question is that it misrepresents what exactly the formal model is doing. In particular, this objection confuses describing a language and using that very same language. The formal semantics presented in §2.3 is a description (i.e. a model, in the intuitive sense of model) of a sequence of possible language extensions. No semantic predicates are used in describing this mathematical structure—the account is (or, can be reformulated) within ﬁrst-order set theory. As a result, the formal model (can) occur in our (actual) base language corresponding to L0 (and, in fact, this is the proper place ²² Actually, this is a bit of a sloppy way of putting it, since we have expressed SUP, in English, just a few lines previously. More precisely, SUP is expressible (we just expressed it), but it does not say what we think it does. There are intricate questions looming here, connected to whether or not we are, in any sense, ‘allowed’ to extend our language in inconsistent ways. If, however, we assume that such a predicate as that used in SUP is added to our language, and that the language remains consistent, then the predicate does not express what we (in some sense) intended it to express. The situation is exactly analogous to the fact that we cannot add the term ‘the set of all sets’ to our language and expect it to consistently have its ‘intended’ meaning.

Embracing Revenge / 47 for such theorizing). Thus, the account of the formal semantics does not occur within a language that uses all of the semantic notions which it describes as occurring in the hierarchy.²³ Nevertheless, the critic might continue, if our base language L0 contains the semantic theory as described in the previous section, could we not extend the language by adding the predicate: is a sentence which receives one of the pathological values described in L0 . Furthermore, once we have added such a predicate to our language, what is to stop us from using this predicate to formulate a version of the Super Strengthened Liar? The answer to this two-part question is of course a two-part answer. First, and most easily, we can grant to the critic that he is free to add such a predicate to his language (thus in a sense ‘skipping’ the step by step individual extensions of the language as described in the formal theory). The answer to the second question, however, is more complicated, and requires our getting a bit clearer on what we mean by the phrase ‘pathological value described in L0 ’. Remember that L0 (plus its interpretation) contains a set theory at least as strong as ﬁrst-order ZFC. In addition, it describes a series of languages Lα for every ordinal α. Thus, there will be a least ordinal, call it π, such that there is a model of L0 with π set theoretic ranks, and every model of π has at least π ranks. As a result, the theory of L0 only guarantees the existence of all ordinals less than π (although it will have models with more ordinals as well). As a result, L0 will only guarantee the existence of π distinct languages, with a corresponding collection of π distinct truth-values. Since a theory only ‘describes’ those objects that are guaranteed to exist according to that theory (or, at least, we assuming as much here), we should interpret the phrase ‘pathological value described in L0 ’ as being satisﬁed only by pathological values pβ where β < π. As a result, adding the suggested predicate to the language extends the language to a sub-language of Lπ which requires the same truth-values for its interpretation as Lπ itself (the language in question will be a sub-theory of Lπ since the language we obtain—a natural language—is countable while Lπ is not). Lπ (and the new semantic value pπ ) are not described in our (present) base theory.²⁴ ²³ Something similar to this approach is found in Field (2003). Although our accounts differ signiﬁcantly, I owe much to the careful study of this chapter. ²⁴ Actually, it is this very phenomenon that justiﬁes our carrying out the construction of the previous section into the transﬁnite. We did not do so because it is likely, or even possible, that we might some day master a language with an uncountably inﬁnite vocabulary such as those high up in the hierarchy. Rather, the thought is that we might, through tricks such as the one considered here, extend our language in such a way as to require the same collection of truth-values as is required by one of the uncountable languages in the hierarchy (even if the actual language we are using remains countable).

48 / Roy T. Cook So, given a particular set theory in the base language L0 , we can extend our language past all of the languages described in the formal semantics developed in L0 by adding this predicate. This pushes us ‘up’ to a language beyond any described in our L0 semantics. But of course we wish to be able to provide a semantics for this language as well. The solution is to strengthen the set theory at the base level. If we add an additional axiom guaranteeing the existence of the ordinal π, then our new base theory will allow us to formulate a new account of the formal semantics which is identical to the original one other than the fact that it implies the existence of more languages (and more pathological values to accompany them). We can of course repeat the process, adding other semantic predicates, and obtaining even stronger languages as a result. In principle, there is no limit to such extensions (other than those imposed by our ﬁnite lifespans, etc.). Thus, in a sense we can never provide a complete account of all of possible extensions of our base language, since for any such account (formulated in a particular set theory T ), we can add a predicate: is a sentence which receives one of the pathological values described in T. which extends the language past what can be described by T. This is not a ﬂaw in our formal semantics, however, but instead reﬂects a well-known feature of set theory. Our formal semantics entails that there is a language Lα for each ordinal α. But for any consistent set theory T, there is a stronger consistent set theory T’ such that T’ implies the existence of more ordinals than does T. Since we can never formulate a set theory which implies the existence of all possible ordinals, we can never formulate a formal semantics for our account which implies the existence of all possible extensions of our language (and corresponding truth-values) in some absolute sense of the word ‘all’. While no single set theory implies the existence of all ordinals, however, there seems to be no reason to doubt that, for any ordinal, there is a set theory that implies its existence. As a result, for any possible extension of our language, we can formulate a semantics for it (by utilizing a suitably strong set theory in the base theory). Earlier we drew an analogy between the indeﬁnite extensibility of the concept ordinal and the indeﬁnite extensibility of the concept language (and the corresponding indeﬁnite extensibility of truth-value). The previous few paragraphs suggest, however, that there is more to this than just an analogy—in fact, the indeﬁnite extensibility of our language just is the indeﬁnite extensibility of the ordinals. This insight promises fruitful connections between the semantic and set-theoretic paradoxes.²⁵ ²⁵ One such connection involves the fact that certain very powerful versions of the Strengthened Liar, such as SUP above, seem (on the present account) to entail (at least indirectly) large cardinal axioms. These connections will be explored in future work.

Embracing Revenge / 49 Thus, the lesson to learn from the Revenge Problem is just this: What we can say (and the semantic values that what we say can receive) is indeﬁnitely extensible in exactly the way the ordinals are. This implies that there is no language in which we can say everything. It does not imply, however, that there is something (coherent) which we cannot say in any language.

2.5 Truth, the Conditional, and Field Something should be said, at this point, regarding the conditional, or, on the present view, the conditionals, plural. The formal theory presented in section 2.3 (and the informal account motivating it) provides, not a single conditional, but a new conditional for each extension of the language. The reason for this move, odd looking though it might be at ﬁrst glance, is the need to avoid Curry’s Paradox. The arithmetic version of Curry’s construction can be carried out as follows: Given an arbitrary sentence , we can diagonalize to obtain a sentence X where: X ↔ (T < X >→ ) If (a) our truth predicate is transparent (i.e. for any , and T < >are intersubstitutable in all non-opaque contexts), (b) our conditional satisﬁes the inference rule modus ponens, and (c) our conditional validates the contraction axiom ((A → (A → B)) → (A → B)), then we can prove : [1] [2] [3] [4] [5] [6] [7]

X → (T < X >→ ) X → (X → ) (X → (X → )) → (X → ) (X → ) (T < X >) → ) X

Diagonalization [1], Transparency of Truth Contraction [2], [3], Modus Ponens [4], Transparency of Truth [5], Diagonalization [4], [6] Modus Ponens

Note that if we deﬁne negation in terms of the conditional and a primitive absurd sentence ⊥: ¬ =df → ⊥ then the Liar paradox is merely a special case of Curry’s paradox.²⁶ ²⁶ The fact that the Liar Paradox can be seen as a special instance of Curry’s Paradox, and thus in some sense the problems with semantic paradox in general hinge crucially on the conditional, is vastly underappreciated. Greg Restall’s contribution to this volume, however, is a notable exception to this.

50 / Roy T. Cook Thus, the real problem highlighted by the semantic paradoxes is that we cannot (on pain of triviality) have both: (a) (b)

A transparent truth predicate. A conditional that uniformly satisﬁes the standard axioms and rules of inference for the conditional.

Since a transparent truth predicate is one of the main motivations for the present account, it follows that we cannot have a single conditional uniformly satisfying the rules usually attributed to the conditional. Denying the validity of basic principles such as contraction and modus ponens is a bitter pill to swallow, however, so the obvious move is to give up as ‘little’ of the standard rules for the conditional as possible. There seem to be two ways in which this can be accomplished. The ﬁrst way is to attempt to provide a single conditional that comes as close as possible to the standard classical account. Hartry Field’s important (2003) (and its successors, including his contribution to this volume) represents such an approach. Field’s theory contains a transparent truth predicate, and in addition it contains a conditional that is well behaved when applied to non-pathological sentences. The main drawback with this approach, however, is that the conditional cannot satisfy all instances of the standard axioms and rules for the conditional.²⁷ In particular, contraction can fail when the subformulas involved fail to receive classical values (as happens with the instance of contraction relevant to Curry’s paradox). On the other hand, however, we can give up on the idea of a single conditional, and instead accept that the concept conditional is itself indeﬁnitely extensible. This is the approach taken above. In each extension of our language, we obtain (or, at least, we can obtain) a new conditional that is a better approximation to our intuitive ideas regarding ‘if. . .then. . .’. Each conditional satisﬁes all the axioms and rules we would expect, at least when applied to sentences of earlier languages. In particular, every instance of both modus ponens and contraction are validated. This last claim, of course, needs some explanation, given that if we have both modus ponens and contraction without any restrictions whatsoever, then we can (as we have already seen) reconstruct Curry’s paradox. The point is this: all instances of the inference rule modus ponens are valid, since (assuming we deﬁne validity in the standard way as truth-preservation in models) no matter which instance of the conditional one takes, if it is true, and its antecedent is true, it follows that its consequent must be true. In addition, given any sentences A and B, there will be some ²⁷ This is not to say that his account does not do an admirable job, from a technical perspective, of validating as many of the standard rules as is possible.

Embracing Revenge / 51 ordinal α such that ((A → (A → B)) → (A → B)) is a theorem.²⁸ Curry’s paradox is blocked, however, since the conditional →α validating the relevant instance of contraction occurs later in the hierarchy than does the conditional occurring in the Curry sentence we obtain through diagonalization. This would seem to be the main advantage of the present view over Field’s approach, at least if one of our motivations (secondary, perhaps, to a transparent truth predicate) is to retain as much of the traditional account of the conditional as is possible. On the present account we need not give up any instance of any standard rule or axiom for the conditional (we just need to remember that not all instances of all rules or axioms are valid for all conditionals!).²⁹ Despite this difference, both views (Field’s and the present one) represent instances of a family of views which take as their primary motivation the salvation of a single, uniﬁed truth predicate that is transparent (at least in non-opaque contexts). Both views accept the fact that an account achieving this (and containing some ²⁸ Proof sketch: Let α be the least ordinal such that A and B both occur in Lα . Consider the Lα+1 sentence: ((A →α (A →α B)) → (A →α B)) Since A and B receive, as truth-values, either t, f, or pβ for some β ≤ α, the relevant portion of the satisfaction clause for this conditional is: I( →α )

=

t f

I() ≤ I() relative to the ordering : f < pα < . . . < pβ+1 < pβ < . . . < p2 < p1 < t otherwise.

Assume, for reductio, that this formula fails to be true, that is: I(A →α (A →α B)) > I(A →α B) Then: I(A →α (A →α B)) = t And: I(A →α B) = f The former implies that: I(A) ≤ I(A →α B) And the latter implies: I(B) < I(A) Since partial orders are transitive, we obtain: I(B) < I(A →α B) = f But this is impossible, since f is the minimal element of the ordering. ²⁹ In the end, however, the choice of one of these accounts over the other, or of some third view over both, should depend, not on technical merits, but on philosophical motivation.

52 / Roy T. Cook sort of reasonable conditional) will require stratiﬁcation of some sort (this is what Tarski had right), at least if we wish out ﬁnal account to be able, in some sense, to ‘completely’ characterize all semantically problematic sentences. Field chooses to ﬁnd the stratiﬁcation in a hierarchy of stronger and stronger ‘deﬁnite’ truth predicates, while on the present view the same role is played by a hierarchy of stronger and stronger conditionals (plus the various semantic predicates that accompany them). Nevertheless, despite the vastly different details, the general philosophical viewpoint seems roughly the same.³⁰

References Cook, R. (2004). ‘Patterns of Paradox’, Journal of Symbolic Logic 69: 767–74 Dummett, M. (1993). The Seas of Language, Oxford,: Clarendon Press Field, H. (2003). ‘A revenge-immune solution to the semantic paradoxes’, Journal of Philosophical Logic 32: 132–177 Fitting, M. (1986). ‘Notes on the mathematical aspects of Kripke’s theory of truth’, Notre Dame Journal of Formal Logic 27: 75–88 Godel, K., (1992). On Formally Undecidable Propositions. New York,: Dover Kleene, S. (1952). Introduction to Metamathematics, Amsterdam,: North Holland Kripke, S. (1975). ‘Outline of a theory of truth’, Journal of Philosophy 72 (1975), 690-716; reprinted in Robert L. Martin (ed.), Recent Essays on Truth and the Liar Paradox, Oxford, Clarendon Press (1984), pp. 53–81 Russell, B. (1906). ‘On some difﬁculties in the theory of transﬁnite numbers and order types’, Proceedings of the London Mathematical Society 4, 29–53 Shapiro, S., and Wright, C. (2006). ‘All things indeﬁnitely extensible’, in A. Rayo and G. Uzquiano (eds.), Absolute Generality, Oxford: Oxford University Press (2006) Tarski, A. (1933). ‘The concept of truth in the languages of the deductive sciences’, Prace Towarzystwa Naukowego Warszawskiego, Wydzial III Nauk Matematyczno-Fizycznych 34, Warsaw; English translation in Alfred Tarski, Logic, Semantics, Metamathematics, Papers from 1923 to 1938, John Corcoran (ed.), Indianapolis,: Hackett Publishing Company, (1983), pp. 152–278 Williamson, T. (1998). ‘Indeﬁnite extensibility’, Grazer Philosophische Studien 55: 1–24 ³⁰ Earlier versions of this chapter were presented at The Ohio State University and Arche: The AHRC Centre for the Philosophy of Logic, Language, Mathematics, and Mind, and the current version has beneﬁted much from the comments, criticisms, and congenial atmosphere found there. Thanks are also due to JC Beall, Stewart Shapiro, Timothy Williamson, and an anonymous referee for additional comments or guidance. A special debt is owed to the students in my Spring 2006 undergraduate Leibniz seminar at Villanova University, who were forced to listen to early versions of these ideas before and after class, and without whose prodding I might never have started, or ﬁnished, the chapter.

3 The Liar Paradox, Expressibility, Possible Languages Matti Eklund

3.1 Introduction Here is the liar paradox. We have a sentence, (L), which somehow says of itself that it is false. Suppose (L) is true. Then things are as (L) says they are. (For it would appear to be a mere platitude that if a sentence is true, then things are as the sentence says they are.) (L) says that (L) is false. So, (L) is false. Since the supposition that (L) is true leads to contradiction, we can assert that (L) is false. But since this is just what (L) says, (L) is then true. (For it would appear to be a mere platitude that if things are as a given sentence says they are, the sentence is true.) So (L) is true. So (L) is both true and false. Contradiction. In the literature there is a bewildering variety of purported solutions to the liar paradox. I will not discuss any of these purported solutions in any detail. Instead I will further problematize the question of what a solution should achieve. I will bring up a somewhat neglected cluster of problems connected with the liar paradox. These problems remain even if one of the solutions to the liar paradox currently on offer would succeed perfectly in solving the problem it is designed to solve. They arise when we consider what possible languages there are.¹ Often in discussions of the liar paradox, truth ¹ (1) I will throughout conceive of languages as necessarily existing abstract objects. If this assumption should be rejected, my conclusions about this will just have to be reformulated as claims about what

54 / Matti Eklund predicates are treated as if they were applicable only to sentences (in contexts), and then only to sentences of one language. Clearly this is a simpliﬁcation. The English language predicate ‘true’ can be applied to sentences of all sorts of languages, to utterances of all sorts of languages, and, perhaps primarily, to propositions. As the arguments here will show, the simpliﬁcation is not innocent, but is an important distortion. I will approach the problem of possible languages via an issue that has attracted some attention in the liar literature: that of the liar paradox and expressibility. As is well known, a consequence of many of the most popular and most widely discussed solutions to the liar paradox is that some properties are deemed inexpressible in natural language. Theorists disagree about just which notions, if any, are inexpressible in natural language; how problematic or not such consequences are; and how best to conceive of the inexpressibility. In section 3.2, I will outline the main views on the liar paradox and expressibility. Sections 3.3 and 3.4 will be devoted to critical discussion of the most important of these views. Section 3.5 will focus on the issue of what possible languages there are. In section 3.6, I will discuss how serious revenge problems really are.

3.2 Inexpressibility and Revenge Problems Whatever is in the end the correct account of the liar paradox, the liar reasoning undoubtedly establishes what is sometimes called Tarski’s theorem: in no language whose logic is classical and which can talk about itself to a sufﬁcient extent can there be a predicate that satisﬁes the T-schema, where for a predicate to satisfy the T-schema is for a valid schema to result when this predicate is substituted for the ‘T’ in s is T iff p, (where instances of this schema are obtained by putting sentences for p and names of the corresponding sentences for s). A natural thought is that Tarski’s theorem entails that natural language is expressively limited. The reasoning would be that natural language satisﬁes the there could be. (2) There are independent problems concerning the notion of all languages. Focus, more narrowly on a problem concerning talk of all possible predicates. (If there is a problem concerning all possible predicates, there is a problem concerning all possible languages.) If, no matter what objects the xs are, there is a predicate true of exactly the xs, there are more possible predicates than there are objects—a contradiction, if predicates are objects. We must either think of quantiﬁcation over predicates and languages as restricted, or else take the space of languages and predicates as smaller. Serious though these problems are, I will set them aside as orthogonal to the problems that I will mainly focus on.

The Liar Paradox, Expressibility, Possible Languages / 55 conditions stated in Tarski’s theorem: so the property which satisﬁes the T-schema is not expressible in natural language, where a property satisﬁes the T-schema iff it is such that any predicate which expresses it satisﬁes the T-schema. (Notice that here and throughout I will use ‘expresses’ for the relation between predicates and the corresponding properties.) There are two ways to resist this reasoning. One is to deny that natural language does satisfy the conditions of Tarski’s theorem (the relevant type of sentence cannot be meaningfully formed or the logic of natural language is not classical).² The other is to deny that there is a property which satisﬁes the T-schema. The thought in the latter case would be that the property simply doesn’t exist. However, if there can be a richer metalanguage with a predicate which behaves in the right way semantically—speciﬁcally, which has the right semantic features to express the property—then it is hard to maintain that the property does not exist. (Here, and throughout, I will assume an ‘abundant’ ontology of properties.)³ Someone who takes Tarski’s theorem to show that there is a property which cannot be expressed in natural language might further hold that any predicate expressing the property of being true would have to satisfy the T-schema: so what would be established is that truth cannot be expressed in natural language.⁴ Sometimes, e.g. in Robert Martin’s (1984a), this reasoning is attributed to Tarski himself. This suggested conclusion is on the face of it bizarre. The claim is that truth cannot be expressed in natural language. But is truth not expressed by the predicate ‘true’, undoubtedly a natural language predicate? The threat that the liar reasoning entails that natural language is expressively limited comes up even if we deny that natural language satisﬁes the conditions of Tarski’s theorem, as in fact many theorists writing about the liar do. Most popular theories of the liar paradox seem to have the conclusion that certain properties—including properties which are expressed by predicates employed in these theories—cannot, on pain of consistency, be expressed in natural language but only in a richer metalanguage. Take ﬁrst theories of the liar which attempt to solve the paradox by saying that the liar sentence has a semantic status somehow intermediate between truth and falsity. The most famous theory of this kind is Saul Kripke’s (1975). Sometimes this intermediate status is conceived of as a third truth-value; sometimes, as in Kripke, it is rather conceived of as ‘unsettled’, or the absence of a truth-value. I will call sentences with this intermediate status neuter; and I will for simplicity talk about this ² There are non-trivial problems concerning going back and forth between what holds for formal theories and natural languages. ³ Below I will present a justiﬁcation of this choice and explain why it does not beg any important questions. ⁴ Or, focusing on language-speciﬁc truth predicates as the literature does, that truth in L cannot be expressed in L, for L a natural language.

56 / Matti Eklund as a truth-value, even if this begs otherwise important questions concerning how the intermediate status is best conceived. A liar sentence of the kind we started talking about can consistently be said to be neuter. But a purported solution of this general kind immediately invites the strengthened liar. Consider a new liar sentence which says of itself that it is untrue. A sentence that is neuter would appear to be, among other things, untrue. But then we are obviously back in paradox. There are different ways around this problem. But one popular way out is to say that in a three-valued setting, both ‘not’ and ‘true’ are ‘weak’: that they take neuter into neuter.⁵ If a sentence S is neuter, so is its negation and so is a sentence which says of S that it is true. If the negation sign and the object language truth predicate work this way, then a sentence which says of itself that it is not true is after all not paradoxical: it can consistently be ascribed the value neuter. Note it is not enough for paradox to be avoided that ‘not’ and ‘true’ be in this way weak: so long as there is any construction at all in the object language that takes both neuter and false into truth and takes truth into falsity—so long as strong negation, as it is often called, can be expressed at all—we land in paradox. The position must be that no construction in the object language can express that a sentence has some semantic status other than truth: as I will put it, that a sentence is untrue. Kripke also holds that untruth can be expressed only in a richer metalanguage. Second, consider the revision theory of truth, defended most prominently in Anil Gupta and Nuel Belnap (1993).⁶ The revison theorist retains classical logic but holds that sentences like the liar sentence cannot stably be assigned any truth-value. This theorist cannot, it seems, allow that ‘not stably true’ of the object language expresses what we would na¨ıvely take this predicate to express. For then consider a liar sentence that says of itself that it is not stably true. Like Kripke, Gupta and Belnap hold that some semantic notions needed in a semantic theory for a natural language can be expressed only in a richer metalanguage: that natural language is not semantically self-sufﬁcient. Gupta (1997) argues at some length that the ideal of avoiding this consequence may well be unattainable.⁷ In general, the situation is this. A standard kind of revenge problem for purported solutions to the liar paradox is the problem that given the expressive resources used to solve the solution to the liar in its simple form, a new paradox can be formed. The standard form of the revenge problem is this: the expressive resources of our language allow us to exhaustively and exclusively divide sentences into the true ones and the rest. If our language has sufﬁcient expressive resources to state ⁵ See e.g. Kripke (1975), Soames (1999), and Maudlin (2004). ⁶ Earlier presentations of the revision theory are Belnap (1982), Gupta (1982), and Herzberger (1982 and 1982a). ⁷ Gupta (1997), pp. 439 ff.

The Liar Paradox, Expressibility, Possible Languages / 57 an exhaustive and exclusive division of all sentences into the true ones and the rest, paradox can be reinstated. Just let our new liar sentence say of itself that it belongs to the rest.⁸ One common way to avoid a threatening revenge problem is to deny semantic self-sufﬁciency. The revenge problem arises only if it is assumed that the semantic theory can be formulated in the object language. It is only then that a new liar sentence can be formulated in the object language. But it is often regarded as an embarrassment for a theory of the liar paradox if it is forced to, so to speak, push some predicates into such a metalanguage. That a theory is so forced can often seem suspect already on intuitive grounds: it does not seem as if I expand my language when I use a construction expressing strong negation or when I use an expression expressing the property of not being stably true. There are also some arguments in the literature to the effect that any theory which is so forced simply must be unacceptable: natural language must be regarded as semantically self-sufﬁcient. One argument is suggested by Vann McGee (1994 and 1997): human language lies within the reach of human understanding and hence it must be possible to state a correct theory of human language in a human language.⁹ However, even if we agree on the assumption here, the conclusion that we must be able to give a semantic theory for a natural language L in L does not follow. Whereas the assumption might entail that there is some possible natural language where we can give a semantics for English, and generally that for every natural language there is some natural language where we can give a semantic theory for it, this does not mean that English possesses those resources. The assumption at most entails that for every natural language L there is some language we are capable of employing in which a semantics for L can be given. Graham Priest (1990) presents a different argument for semantic self-sufﬁciency. He says, ‘The whole point of solutions to the liar paradox (as opposed to reformist suggestions as to how to change our language) is to show that our semantic discourse (about truth, etc.) is, appearances notwithstanding, consistent. An attempt to show this which produces more and more discourse, not in its own scope, therefore fails.’¹⁰ This is not a convincing argument either. First, even if Priest is entirely right, one might think that the proper conclusion to draw might be that a solution of the kind described is unattainable. (Compare again Gupta’s stance.) Second, there would be a problem if the further discourse ‘produced’ could not be plausibly believed to be ⁸ Priest (1987), p. 29, makes this point well. ⁹ See McGee (1994), p. 628, and (1997), p. 402. f; compare too Gupta (1997), pp. 440. ff, who criticizes the argument. McGee comes closest to actually endorsing the argument discussed in the main text in his (1994); the discussion in his (1997) is considerably more guarded. ¹⁰ Priest (1990), p. 202.

58 / Matti Eklund consistent. But there is no immediate reason why the hierarchy of metalanguages that the Kripkean is saddled with should fail to be consistent.¹¹, ¹² In the above discussion I have, inter alia, suggested a number of different views on the implications of the liar paradox regarding the expressive limitations of natural languages. Let me now be more systematic: RADICAL INEXPRESSIBILITY (RI). Some seeming ordinary semantic properties, like truth, are in fact not expressible in ordinary natural languages, but only in a richer metalanguage. INEXPRESSIBILITY (I). Some semantic properties—including semantic properties we need to be able to express in an adequate semantic theory of natural language—are expressible only in a metalanguage. SEMANTIC SELF-SUFFICIENCY (SS). All semantic properties we need to be able to express in an adequate semantic theory of natural language are expressible already in natural language. WEAK UNIVERSALITY (WU). All properties are expressible in natural language; or at any rate: the liar paradox casts no doubt on this claim. (It is the qualiﬁcation that warrants the ‘weak’.)

(RI) entails (I) but not vice versa. (WU) entails (SS) but not vice versa. Martin ascribes view (RI) to Tarski. View (I) has the best claim to being orthodoxy: it is subscribed to by both Kripke (1975) and Gupta and Belnap (1993). McGee’s and Priest’s arguments are arguments for (SS). McGee is explicitly doubtful about (WU). Although Priest does not outright endorse (WU), some of the moves he makes are explicable only on the assumption that he holds this stronger view. For example, having shown to his satisfaction that Boolean negation (or, to take this in terms of properties, the property that a sentence has when its Boolean negation is true) need not be expressed in a semantic theory for natural language, Priest anyway sees it as incumbent upon himself to provide an argument for why Boolean negation just is not there to be expressed, and that it is not an expressive limitation of English that English does not possess the resources to express it. Hartry Field should also be mentioned along with McGee and Priest as someone defending a view of type (SS) or (WU). Field’s theory of truth is designed to meet the requirement of self-sufﬁciency, and Field takes this to be a supremely important consideration in its favor.¹³ ¹¹ It is of course unclear what it is for a ‘discourse’ or a ‘language’ to be consistent or inconsistent. But this unclarity is a problem for Priest, not for me. ¹² Similar remarks as apply to McGee and to Priest seem to me to apply to the argument for self-sufﬁciency given in Scharp (manuscript). Scharp gives an argument for why a theory of truth for a language needs to be internalizable: roughly, that even if it cannot be given in our actual natural language L there must be some expanded version of natural language, L+, where it can be given and such that the theory also applies to L+. The argument is that if this internalizability requirement is not met the theory cannot be both ‘descriptively correct’ and ‘descriptively complete’. Even if Scharp is right about this, the conclusion may well be that a semantic theory of the kind described simply cannot be had. ¹³ See Field (2003a), (2003b), (2005a), (2005b), (this volume).

The Liar Paradox, Expressibility, Possible Languages / 59 I will in my discussion focus primarily on (I) and (WU). Given that the arguments for semantic self-sufﬁciency are unpersuasive, I fail to see (SS) as a principled position. As for (RI), I happen to think that a brief dismissal of it as absurd is too quick—more on this at the very end of the Chapter—but I will anyway not focus much on it. The problems I will discuss regarding the more popular view (I) apply in a similar way to (RI). In section 3.3, I will discuss (I). In section 3.4, I will discuss the viability of (WU).

3.3 Universality I will approach (I) via consideration of the question of the universality of natural language.¹⁴ The notion of universality was introduced by Tarski: A characteristic feature of colloquial language . . . is its universality. It would not be in harmony with the spirit of this language if in some other language a word occurred which could not be translated into it; it could be claimed that ‘if we can speak meaningfully about anything at all, we can also speak about it in colloquial language’.¹⁵

What does ‘universality’ mean exactly? Instead of considering what Tarski and other theorists might have meant by it let me introduce my own—admittedly rough—characterization (which however seems to be in the spirit of Tarski): A language L is universal iff for every property there is a predicate of L that expresses it.¹⁶ A few remarks on this characterization are in order: First, for simplicity I talk only about the expressibility of properties. In a properly general characterization of the expressive power of natural languages we should consider not only properties, but also logical operations, objects, etc. Second, given the focus on properties, metaphysical questions concerning the nature and abundance of properties become relevant. On a conception of properties given which there are only very few properties (a sufﬁciently ‘sparse’ conception of properties, in common jargon) the claim of universality can be (relatively) trivial, and not the interesting claim it is intended to be. (Consider for instance a view on which the only properties there are, are properties corresponding ¹⁴ The classic discussion of universality is Tarski (1935/83); see also Fitch (1946 and 1964), Gupta (1997), Herzberger (1970), Martin (1976), McGee (1991 and 1997), Priest (1984 and 1987), and Simmons (1993). ¹⁵ Tarski (1935/83), p. 164. It is because Tarski says these things that (RI) cannot unproblematically be attributed to him. ¹⁶ The characterization of universality employs unrestricted quantiﬁcation over properties. Such quantiﬁcation is obviously problematic. Although the problems concerning the possibility of unrestricted quantiﬁcation over properties are not unrelated to the liar paradox, I will not here attempt to directly address those problems.

60 / Matti Eklund to the predicates employed in microphysics.) As already mentioned, I will throughout assume an abundant conception of properties, given which, roughly speaking, there are maximally many properties. A justiﬁcation for this strategy is that the talk of properties can be regarded as a convenient way of talking about what objects fall under a predicate. Third, there are questions about how to individuate languages. In logic, languages are individuated very ﬁnely: add another constant and you have a new language. Outside logic, however, natural languages are individuated more liberally: a language can undergo signiﬁcant changes in its vocabulary and its syntax and still we happily speak of it as the same language after these changes. I will here individuate natural languages ﬁnely. The justiﬁcation for this is straightforward. For ‘English is universal’ to express the substantive claim it is meant to express, it must not be sufﬁcient for its truth that for every property, English as it currently exists could be expanded with a predicate which expresses this property, while still we would call it the same language. Fourth, there are some questions regarding just how to understand ‘expressed’. These questions I will soon get to. Now, despite what unclarities may remain in the given characterization of universality, the characterization ought to be clear enough that it should seem intuitively implausible that all natural languages are universal. Consider, say, eighteenthcentury English. Do we really want to say that eighteenth-century English contained the resources for expressing the property of being a quark, or the property of being an inaccessible cardinal? Moreover, is it not perfectly possible that there are properties that simply are determinately outside our cognitive reach, in the way that (I would suppose) the property of being an inaccessible cardinal is outside the cognitive reach of a gorilla? But there is a complication. Stick with the example of eighteenth-century English. Intuitively, it is in some respects limited in expressive resources. It cannot express the property of being an inaccessible cardinal. So there is some property that can be expressed in some natural language that cannot be expressed in eighteenth-century English. But suppose the tallest man on Earth at midnight 1 July 2016 employs the predicate ‘is an inaccessible cardinal’. Then consider the predicate of eighteenthcentury English ‘falls under the predicate employed by the tallest man on Earth at midnight 1 July 2016’. This predicate is true of exactly the inaccessible cardinals: and so, arguably, expresses the property of being an inaccessible cardinal. Hence it appears that our verdict on eighteenth-century English must be revised: it is after all universal.¹⁷ There is an immediate objection to what has just been argued. The objection is that what is shown is at most that for every property φ there is a predicate of ¹⁷ Objection: did eighteenth-century English really contain linguistic vocabulary like ‘falls under’ and ‘predicate’? Reply: The details are irrelevant. What I take for granted is that eighteenth-century English contained some linguistic vocabulary that could be employed for the same purpose.

The Liar Paradox, Expressibility, Possible Languages / 61 eighteenth-century English which, given that the circumstances are propitious, as used on a particular occasion is true of exactly what has φ. But this is not enough for the predicate to express φ. The predicate ‘falls under the predicate employed by the tallest man on Earth at midnight 1 July 2016’ does not express untruth, even if the predicate the tallest man on Earth then used did express this property; rather, it expresses the property of falling under the predicate employed by the tallest man on Earth at midnight 1 July 2016. That is different. However, while there is a distinction here, it is actually for present purposes possible to disregard it. Let us say that a property φ is indirectly expressible in a language L iff there is a predicate F of L such that for some context c, an utterance of F by a speaker of L is such that for all x, ‘F(x)’ and ‘x has φ’ have the same truth-value. It would be wrong, for the reason just noted, to identify indirect expressibility as expressibility.¹⁸ But all that the present discussion needs is indirect expressibility of the relevant properties. The reason is that provided that given that a certain theory of truth demands that property not be expressible in English, it also demands that this property not be indirectly expressible in English. Let me take a few more examples to illustrate further what is at issue, before returning to the lessons of the liar paradox. Suppose, as earlier suggested, that there are properties beyond our cognitive reach: they are too complex, too fantastic, too divine, what have you. These too are properties we can express in our ordinary language, if what I have suggested with respect to the previous examples is correct. Consider predicates of the form ‘falls under the predicate employed at t by the most F creature in the universe’ (Where F might be intelligent, or old, or morally praiseworthy, or what have you.) Some of these predicates might be true of exactly what has one of these properties which supposedly are beyond our ken. It is unlikely, perhaps, that for every cognitively unreachable property (as we might call these properties) there will in fact be a predicate of our language which is in fact true of whatever has that property, as things stand. (Some cognitively unreachable properties need not be thought of at any time by any creature in the universe, etc.) But for all that, each of these properties may still be indirectly expressible. Compare too an objection Keith Simmons presses against taking natural languages to be universal: for every set in the ZF hierarchy there is a distinct concept (the concept of being a member of that set), and ‘[g]iven certain assumptions about natural languages (in particular, about upper limits on the size of vocabularies and the length of sentences), these concepts would outrun the expressive capacity of any natural ¹⁸ Compare the notion of ‘loose speaker expressibility’ characterized in Hofweber (2006): a property φ is loosely speaker expressible in L iff there is a predicate F of L and a context C such that that an utterance of F (by a speaker of L) in C expresses φ.

62 / Matti Eklund language’.¹⁹ This is a convincing argument against the possibility of the expressibility of all properties in natural language. However, Simmons’ reasoning leaves open that these concepts should all be indirectly expressible in natural language.²⁰ Return now to the liar paradox. The present considerations can appear to have the consequence that any theory of the liar paradox that relies on kicking some predicates up into the metalanguage can be easily refuted. Let U be the relevant metalanguage predicate. Then in the object language we will have predicates of the form ‘falls under the predicate U of the language spoken by so-and-so at time so-and-so’. Whatever is expressed by U will be indirectly expressible in the object language. But then we are back in paradox: let P, in a given context, indirectly express what U expresses, and consider a sentence which says of itself that it is P. (Call this kind of version of the liar paradox an interlanguage paradox.) However, it would be much too hasty to conclude from this reasoning that no view of type (I) can be maintained. For the reasoning relies on the assumption that for any predicate F, F and ‘falls under F’ are coextensive—as I will put it, that ‘falls under’ is transparent (generally, that satisfaction predicates like ‘falls under’, ‘is true of’, ‘satisﬁes’, etc. are transparent). While it is natural to hold that ‘falls under’ is transparent, liar-like reasoning—speciﬁcally, Grelling’s paradox²¹—calls this into doubt. More speciﬁcally, the assumption that ‘falls under’ is transparent is in the same category as the assumption that a sentence S and the corresponding sentence ‘S is true’ always have the same truth-value (that ‘true’ is transparent). Both assumptions are called into doubt by liar reasoning. But although there is no immediate refutation of views of type (I) in the ofﬁng, there is a nearby puzzle that deserves stressing. Although some take the liar reasoning to show that the truth and satisfaction predicates of English are not transparent, for example Kripke (1975) and others following up on his work have shown that there can be languages which are like natural languages whose truth and satisfaction predicates are transparent. One way the truth and satisfaction predicates of a language L can be transparent is if L lacks an untruth predicate. This is for instance the way that Kripkean theories of truth achieve transparency. The languages he describes ¹⁹ Simmons (1993), p. 15. Simmons talks about the expressibility of concepts; I prefer to think of the issue as about the expressibility of properties. ²⁰ Once the notion of indirect expressibility is introduced, one can introduce a corresponding notion of indirect universality: a language L is indirectly universal iff for every property there is a predicate of L that indirectly expresses it. What I have argued is that some immediate reasons for doubt about whether English is universal do not readily show that English fails to be indirectly universal. There may, for all I am concerned to argue, be reasons to doubt whether English really is indirectly universal. All I am concerned with is the indirect expressibility of those properties some theorists say for liar-related reasons can only be expressed in a metalanguage. ²¹ Grelling’s paradox: Let a predicate be heterological just in case it is not true of itself. Is ‘heterological’ true of itself or not?

The Liar Paradox, Expressibility, Possible Languages / 63 can contain transparent truth and satisfaction predicates on pain of not containing untruth predicates, as earlier discussed. If they did, we would be back in paradox, as is well known. But here we must be careful. Can there be a language L which contains transparent truth and satisfaction predicates on pain of not containing an untruth predicate?²² There is a difﬁculty here. There cannot be such a language L, if (i) there is some other language L∗ with an untruth predicate for languages of the kind to which L, if it exists, belongs (say, natural languages), and (ii) L’s truth and satisfaction predicates apply to sentences and predicates of languages other than L; speciﬁcally, of L∗ (if it exists). Now, if we think of L as something like a natural language, L would appear to have to satisfy (ii). It then comes down to condition (i). We are in a curious position. For L can exist if L∗ doesn’t and L∗ can exist if L doesn’t. Which of these two languages exists? I will return below to the signiﬁcance of this kind of puzzle.²³ For now let me just note the following. The puzzle does not have to do with what we should say about actual natural languages. The question is about a hypothetical natural language L, like actual ones, except for the possible difference that its truth and satisfaction predicates are transparent and on pain of consistency it does not contain an untruth predicate. One might well think that L, whether actually used or not, certainly exists. But now it turns out that L does not exist, if there is another language L∗ , with an untruth predicate for languages like L. The sort of puzzle I have called attention to does not refute view (I) on semantic self-sufﬁciency and universality. All it shows is that such a view can be hostage to facts about what possible languages there are.

3.4 Priest on Boolean Negation Turn now to views of type (WU): views according to which the liar paradox does not have the consequence that some notions are inexpressible. The discussion will be ²² Note the formulation. The question is not about the possibility of transparent truth and satisfaction predicates, simpliciter. It is about one particular way of achieving transparency. ²³ It deserves noting that while the puzzle presents problems for Kripkean theories of truth, it presents no problem for more old-fashioned, Tarskian theories, denying that truth in L can be expressed in L. Two points deserve noting. (i) A consistent Tarskian ought to say the same thing about ‘true of ’, ‘satisﬁes’, and ‘falls under’ as about ‘true’. (ii) A Tarskian obviously wouldn’t hold that L can express truth in other languages. (After all, truth in a natural language can for the Tarskian be expressed in a metalanguage. But a Tarskian will obviously want to hold that there cannot be pairs of languages L and L such that L can express truth in L and L can express truth in L. It should be clear that no interlanguage paradox can force the Tarskian to accept that truth and satisfaction in L can at least be indirectly expressed in L.

64 / Matti Eklund focused on Graham Priest’s defense of a view of this type. Priest is a dialetheist, and this informs his discussion. But even though dialetheism is a radical view, it should be clear that the points I make regarding Priest generalize. A dialetheist holds that there are true contradictions. The liar paradox presents one of the main arguments for dialetheism. The idea is that the solution to the liar paradox is to recognize true contradictions (the liar sentence is both true and not true). For the view that there are true contradictions to get off the ground, negation must not work in such a way that from a contradiction everything follows (it must not satisfy ex falso quodlibet); negation must, in other words, be paraconsistent. For otherwise the dialetheist would have to hold that all propositions are true. One of Priest’s main arguments for a dialetheist solution to the liar paradox is this. The liar reasoning forces us to make a choice between embracing expressive incompleteness (hence to adopt view (RI) or (I)) or embracing inconsistency. Nondialetheist theories of the liar commit their proponents to embracing the former alternative. But that is untenable. Hence the latter alternative, dialetheism, is forced upon us.²⁴ There are a few different lines of resistance to an argument like this. One is to deny that non-dialetheic theories have the consequence that natural languages are expressively incomplete. A second is to deny that this consequence is particularly damaging. A third, which I will focus on in this section, is to argue that the dialetheism does not in fact avoid expressive incompleteness. Such an argument can take different forms. It can be argued that the dialetheist cannot allow a predicate which expresses that a sentence is only untrue. For consider then a liar sentence which says of itself that it is only untrue. The standard dialetheist strategy, of saying that liar sentences are both true and untrue, does not seem to be applicable here. For if this sentence is both true and untrue, then it must be only untrue. But if it is only untrue, then it is true after all.²⁵ Another argument, which is the one I will focus on, since it is the one that Priest (1990) is about, is to the effect that the dialetheist cannot allow the expressibility of Boolean negation. As Priest himself states the problem, ‘If [Boolean negation] is allowed then, using the T-scheme and self-reference in the usual way, we can produce a sentence equivalent to its own Boolean negation, and hence deduce a Boolean contradiction, whence everything follows by Boolean EFQ [ex falso quodlibet]’.²⁶ Talk about the inexpressibility of Boolean negation is on its face different from talk about inexpressibility of properties, since a negation sign does not express a property. ²⁴ See e.g. Priest (1990), p. 202. ²⁵ The point of the argument (which has been presented by others, see e.g. Parsons (1990)) is just that dialetheism faces its own expressibility problem, prima facie as serious as the expressibility problems that arise for other theories. Naturally, dialetheists have responses to these problems; see e.g. Priest (1995). ²⁶ Priest (1990), p. 203.

The Liar Paradox, Expressibility, Possible Languages / 65 But I will slide over this difference, since the same remarks that apply to the question of the possibility of having a logical operator that expresses Boolean negation apply to the question of the possibility of having a predicate that expresses the corresponding property (the property a sentence has when its Boolean negation is true). Priest’s way out is to deny that there is such a thing as Boolean negation. It just is not there to be expressed. So the supposed fact that no expression of English expresses it indicates no expressive limitation in English. His argument for this perhaps surprising conclusion proceeds by way of criticizing possible attempts to show that there must be such a thing as Boolean negation. Let me brieﬂy summarize how Priest argues. ²⁷ Suppose ﬁrst that someone says that we can state rules of inference characterizing Boolean negation, and therefore Boolean negation exists. To this the reply is that we can state rules of inference for Prior’s ‘tonk’—the disjunction introduction rule and the conjunction elimination rule—and this does not mean that there is an associated operation expressed by the connective.²⁸ A more sophisticated version of the appeal to rules of inference would have it that the rules of inference for Boolean negation satisfy the condition for successful introduction of a connective while those for ‘tonk’ do not. But a standard condition is conservativeness: and a connective satisfying the rules of inference for Boolean negation will not satisfy this condition, if it is introduced into a dialetheist language with a truth predicate satisfying the T-schema (precisely because it makes the language trivial).²⁹ A different suggestion is to say that the rules of inference manage to characterize an operation if they are demonstrably sound. Concerning this, Priest argues that any argument to the effect that the relevant rules of inference are sound must itself employ Boolean negation, and so is question-begging.³⁰ As Priest is quite clear about, the force of his argument against the existence of Boolean negation relies on there being an independent case for thinking that dialetheism is true of English.³¹ Consider the point about how the rules of inference for Boolean negation fail to satisfy the conservativeness criterion. That point relies on assumptions about what is in the language prior to the introduction of Boolean negation. Priest argues that this fact about the dialectic is no cause for concern. ²⁷ My discussion of this argument will follow Priest (1990). The argument is also given in Priest (2005), ch. 5, which however focuses less on the liar paradox. Another type of argument for why some things we thought were expressible at least in some possible language do not in fact exist is that given in Maudlin (2004), passim. Maudlin goes from the assumption that ‘truth and falsity are always rooted in the state of the world’ to the claim that ‘if a sentence is true or false, then either it is a boundary sentence, made true or false by the world of non-semantic facts, or it is semantically connected to at least one boundary sentence, from which its truth value can be traced’, (p. 49), and from this Maudlin concludes that, e.g., there cannot be such a thing as strong negation. Not only is our negation not strong; in no language can negation behave that way. I cannot adequately discuss Maudlin’s stance here. Sufﬁce it to say that I agree with the observation of Gupta (2006) that Maudlin’s claim does not follow from the initial assumption. ²⁸ Priest (1990), p. 204. ²⁹ Ibid., pp. 204 f. ³⁰ Ibid., pp. 205–8. ³¹ Ibid., p. 209.

66 / Matti Eklund I have two remarks to make about Priest’s argument. First, note that if Priest’s argument for why a dialetheist need not appeal to inexpressibility is successful, then a non-dialetheist can make use of essentially the same argument, provided she does not need for the statability of her view that the property deemed inexpressible in the object language is expressible in the metalanguage. Field’s theory is arguably a non-dialetheist theory which meets this condition.³² Second, more importantly, if we widen our perspective to look at possible languages, I think we can see that, despite Priest’s claims to the contrary, there is a problem with the structure of Priest’s argument. Let me explain. Consider a hypothetical linguistic community such that the speakers of this community have implicitly made a semantic decision to use an expression to mean Boolean negation (where by their having made an ‘implicit semantic decision’ I simply mean that they have come to use, and come to conceive of, their negation sign as obeying the principles governing Boolean negation). Surely such a linguistic community is conceivable. And considering such a community begs no question against Priest. His argument is to the effect that Boolean negation does not exist. All that follows from this with respect to any actual or hypothetical linguistic community is that any attempt to express Boolean negation will fail. Insofar as we are convinced by Priest’s arguments that dialetheism is the correct story about our language—and suppose for argument’s sake that we are—then from our perspective, it seems that our negation is not Boolean. Given this, an argument along Priest-style lines can be given that Boolean negation does not exist. But from the perspective of this hypothetical community, it appears that their negation is Boolean and an argument analogous to that Priest gives can be given for taking something we take to exist (perhaps truth taken as satisfying the T-schema) not to exist. What is the truth of the matter? Here are some straightforward suggestions. (a) We are simply right and they are simply wrong (or, for that matter, they are simply right and we are simply wrong). (b) It is somehow objectively indeterminate what exists and what does not exist. (c) Both types of negation really do exist: it is just that they are expressible in different languages. (d) The underlying metaphysical assumptions ought to be questioned. ³² Field (this volume) argues that a fully general property of determinacy—such as otherwise would cause revenge problems for his theory—in fact does not exist. The reason, brieﬂy, is that not only does Field’s object language not contain the resources to deﬁne the relevant notion of determinacy: more generally, one cannot extrapolate from the resources of the object language to make intelligible such a notion of determinacy. In the main text I focus on Priest’s argument rather than Field’s because Priest’s argument seems more theoretically interesting. Why should Field’s reasoning make us the least inclined to conclude that the relevant notion does not exist? (Scharp (this volume) discusses this problem with Field’s argument at greater length. Compare too Yablo (2003), pp. 328 f.)

The Liar Paradox, Expressibility, Possible Languages / 67 The suggestions are all very problematic. Consider ﬁrst some problems with (a), which is the suggestion Priest would be forced to embrace. First, the situation seems clearly to be symmetric. Second, the consequence that whole linguistic communities might in this way be wrong about meaning is not easy to swallow. It is not hard to imagine cases where whole communities have mistaken views about what the expressions of their language mean. But a case like this would be different. Here a community would not just embrace the theory that, say, their negation sign expresses Boolean negation. Their whole use of this predicate would also point this way—they actually employ the relevant inference rules, etc.—and yet they could be wrong. Or that is what embracing (a) would have us accept. Third, setting these questions of knowledge and ignorance aside, one might think that it would be odd if meaning facts were hostage to metaphysics in the way suggested by this alternative: metaphysical facts about what propositions there are, not readily knowable to ordinary speakers, are held to determine what speakers mean. Let me elaborate on the second and third objections. We are, to be sure, well accustomed to the idea of semantic and psychological externalism, theses to the effect that the meanings of a speaker’s linguistic expressions and the contents of her mental states can be in part determined by factors external to the speaker. The arguments from Kripke, Hilary Putnam, and Tyler Burge are very familiar and taken to be persuasive.³³ Moreover, David Lewis has made popular the idea that some entities in the world are more natural, and hence more eligible to be meant and referred to, than are others.³⁴ This is another route to externalism. It might be thought that in light of these points, the third objection should not be very serious. But Kripkean and Putnamean appeals to the causal environment are beside the point when we are talking about expressions like ‘true’ and ‘not’. And the appeal to the social environment in Putnam’s ‘elm’/‘beech’ case and in Burge’s arthritis case is beside the point when we are talking about whole linguistic communities being wrong. The way in which the external world would matter to the meanings of speakers’ expressions under hypothesis (a) is most similar to the way that Lewis takes the world to matter. A Lewisian appeal to naturalness can certainly be made also with respect to expressions that purport to refer to abstract objects. But there is a crucial difference between the kind of dependence on the external world that obtains if Lewis is right and the dependence that (a) would saddle us with. For Lewis stresses that he does not claim that nothing could mean grue or quus. As he notes, such a claim would on the face of it be absurd. All he claims is that if the facts about a speaker’s dispositions with respect to her use of ‘+’ do not determine whether this sign as the speaker uses it means plus or quus, then the greater naturalness of the former decides in its favor. By contrast, what we are now asked to accept is that it would be absolutely impossible for anyone to mean Boolean negation by a sign. ³³ See Kripke (1980), Putnam (1975), and Burge (1979).

³⁴ Lewis (1983) and (1984).

68 / Matti Eklund Someone might defend (a) from the objections pressed by attempting to argue that our intuitions about logical matters not only tell us what we do and do not mean by various logical signs, but also about what operations there are and are not in logical reality. So for instance, in so far as our intuitions are best respected by a semantics for our language given which negation is paraconsistent instead of classical, this can be taken to tell us something not only about our language but about what logical operations there are, and hence what possible languages there can be. If something like this is right, then the objection to suggestion (a) that the envisaged situation is perfectly symmetric is completely beside the point. This appeal to what members of a hypothetical linguistic community would think and intuit is as misbegotten as an objection to reliance on what our senses deliver which appeals to what hypothetical creatures whose senses would deliver contrary things would be. It is of course correct that if intuition is viewed in the way here described, the objections I have presented to suggestion (a) are beside the point. But how plausible is this view on logical intuition? (Further remarks relevant to this issue follow in the next section.) So suggestion (a), which Priest is forced to embrace, faces serious problems. However, it is not evident that this means that Priest has been shown wrong. For the other suggestions likewise face serious problems. Turn ﬁrst to suggestion (b). This is appropriately fair when it comes to the question of who is right and who is wrong. But how make sense of the postulated indeterminacy? The indeterminacy cannot be semantic, given the way semantic indeterminacy is normally understood. For in order for an expression to be semantically indeterminate as between two different meanings (the way that, for instance, a vague expression is sometimes held to be semantically indeterminate as between different precisiﬁcations) the different meanings must exist. And the problem we are currently dealing with is that the respective meanings cannot possibly coexist. So the indeterminacy in question must be ontological. But ontological indeterminacy is widely regarded with suspicion: what does it even mean for the world to be in and of itself indeterminate?³⁵ As for (c), taking this route would mean abandoning Priest’s universality view for a view of type (I). Moreover, given that the dialetheist’s reasons for excluding Boolean negation apply equally in the case of a predicate expressing the property corresponding to Boolean negation, the concerns above stressed with respect to (I) would apply here too. These negative reﬂections may make us attracted to (d): we may think that it is the associated metaphysical picture that leads us wrong in the ﬁrst place. But rejecting the associated metaphysical picture—of there being these things, properties ³⁵ Notice too that here we are talking about a very special kind of purported ontological indeterminacy: indeterminacy in what logical operations there are. Postulating indeterminacy here is quite different from, say, suggesting that clouds and mountains are ontologically indeterminate.

The Liar Paradox, Expressibility, Possible Languages / 69 and operations, which our predicates and logical connectives express—does not get around the problem. For what Priest is fundamentally concerned with when arguing that there is no such thing as Boolean negation is that there can be no language with an expression that functions in thus-and-such a way. This question too may be metaphysical—it concerns what languages there can be—but if it is metaphysical, it is unavoidably so. It does not presuppose a potentially objectionable ontology of entities supposedly expressed by predicates and by logical operators. The situation is curious. Alternatives (a)–(d) appear to be exhaustive. But they all seem objectionable. This is a paradox. There is no clean objection here to what Priest says about Boolean negation, for although the view he is committed to, (a), seems bad, it is not clearly a worse view than the others. But Priest’s position is far from unproblematic.

3.5 Possible Languages I have discussed views (I) and (WU) on the liar paradox and expressibility, and somewhat tentatively presented some problems with both views. Let me now broaden the perspective a bit, before in the next section returning to the so-called revenge problem often raised in discussions of the liar paradox. Consider two different questions raised by the liar reasoning: The actual-language question: what is the correct semantic theory of ‘true’, ‘not’ and other key expressions in the liar reasoning? The possible-language question: what possible languages are there?

The actual-language question is more often discussed. The reason for the focus on this question is clear. The liar reasoning is carried out in ordinary natural language. It presents a puzzle about how ordinary language works. Hence it is only to be expected that those theorists who seek to solve the liar paradox are concerned to get natural language right. The formal theories that are developed are meant to be accurate models of natural language. But the liar paradox also has signiﬁcant implications for the possible-language question. For instance, Tarski’s theorem can be regarded as a theorem about what possible languages there are. In this section I will focus on the possible-language question. For one thing, this is very much an additional question, both difﬁcult and philosophically signiﬁcant, and it is not clear that it is solved even by an otherwise perfectly adequate solution to the liar paradox. For another, consideration of the possible-language question is of consequence for the actual-language question. What if the way I take English actually to be does not correspond to the way that any language could possibly be?

70 / Matti Eklund What possible languages are there? Many are happy to embrace a principle of plenitude for abstract objects, according to which (to put things intuitively) all the pure abstracta that coherently can exist also do exist. In the case of mathematical entities, the consequences of such a principle of plenitude are relatively straightforward. But in the case of languages, it is harder to see what the consequences of a principle of plenitude are.³⁶ Take a semantic theory according to which the semantics of English is somehow fuzzy-valued (degree-theoretic). Does this semantic theory describe a possible language at all? One kind of reason for doubt is this. Even if, given a principle of plenitude, there will of course be an abstract object some of whose constitutive parts or elements—the would-be sentences—are somehow mapped onto real numbers, the question is whether this is a possible language. It is a possible language only if these real numbers can adequately be regarded as, in some sense, truth-values (or, generally, semantic values of sentences). This it is possible to deny, consistently with ﬁrm adherence to a principle of plenitude. A classic discussion of this is Michael Dummett’s (1959), where it is argued that for general reasons having to do with the speech act of assertions, assertions can only be either correct or incorrect, and since truth and falsity corresponds to correctness and incorrectness in assertions, any appeal to ‘multivalence’—there being truth-values distinct from truth and falsity—must be denied. On what we may call a liberal view on what languages there are, there are languages corresponding to quite different classical and non-classical semantics. On a restrictive view, only languages corresponding to a very restricted class of semantics exist. Even if a very restrictive view is correct, there are further problems with respect to what possible languages there are. For it may be—as brought up above in connection with the interlanguage paradox—that whether a given, considered by itself seemingly perfectly possible, language really exists depends on what other languages there are. For future reference, let us say that there are two possible types of restrictions on what possible languages there are: Dummettian restrictions, having to do with e.g. what can intelligibly be said about the truth-values of sentences (I call them ‘Dummettian’ since Dummett is the theorist who has done the most to put them in the foreground—but I want to include under this general label also considerations Dummett himself would refuse to endorse) and liar-related restrictions, having to do with what liar-type reasoning entails. The obvious methodology for dealing with the actual-language question posed by the liar paradox is to compare proposed accounts of the paradox with intuitions we have. ‘Intuitions’ is used in many senses in the philosophical literature. But here it ³⁶ As stated in footnote 1, I here presuppose that languages are pure abstracta. If this is denied, the relevant question is instead about what languages there can be. The same considerations are still relevant.

The Liar Paradox, Expressibility, Possible Languages / 71 would appear that we are dealing with semantic intuitions: e.g. intuitions about the truth-values of various sentences containing ‘true’ and other semantic predicates. The correct answer to the actual-language question posed by the liar would standardly be assumed to be: that which best respects our semantic intuitions.³⁷ It is certainly possible that our semantic intuitions are inconsistent. In that case, our language is arguably the language that, so to speak, is the language ‘nearest’ our intuitions.³⁸ Attention to the possible-language question complicates this picture. First, even if our semantic intuitions were to be best respected by a theory that postulates for English a many-valued or fuzzy semantics it can be that theoretical considerations show that there are Dummettian restrictions on what languages there are, so that no language has such a semantics. Second, there are liar-related restrictions, such as illustrated by the interlanguage paradox. Can there be a language L, otherwise like a natural language, with transparent truth and satisfaction predicates?—Only if this language is not what we may call seriously limited: only if there is no predicate U of some possible language such that the property expressed by U cannot, on pain of paradox, be expressed in L. There can consistently be a language L of the kind described. But L cannot exist if there is another language as described. And such a language can also consistently exist. Focus now on liar-related restrictions. Suppose that, as earlier outlined, there are two different languages L and L∗ both of which can consistently exist but which are incompatible. How can we know which of L and L∗ exists? One need not be in general skeptical of the idea of reliable substantive metaphysical intuitions to feel the force of the problem here. For even if there are reliable substantive metaphysical intuitions, problems remain. First, such intuitions do not seem to be what we actually draw on when evaluating particular proposed solutions to the liar paradox: we seem rather to rely on ordinary semantic intuitions. Second, more to the point, when it comes to choices like that between a language like English except for the possible difference that its truth predicate is transparent and a language where untruth can be expressed, is it really plausible that we have any intuitions that speak to the issue of which language it is that exists? We can perhaps have strong intuitions that speak to the issue of which ³⁷ I have elsewhere argued that the liar paradox shows that natural language is inconsistent, in the sense that the paradox shows that principles such that it is part of semantic competence to accept them are jointly inconsistent (See especially my (2002).) What is mentioned in the text as a possibility is something considerably weaker. There is a distinction between on the one hand semantic intuitions, in the sense merely of intuitions we actually have about semantic matters, and on the other hand what competence requires us to accept. ³⁸ There are problems in spelling out what ‘nearness’ here amounts to. But there is no way of getting around this issue. It seems rather obvious that our semantic intuitions are inconsistent—that is why the liar paradox is a paradox—and so long as that does not mean that we have failed to endow our expressions with meaning we are forced into the ‘nearness’ talk.

72 / Matti Eklund of these languages it is that we speak, provided both exists, but that is a different matter. Compare the case of other abstract objects; say, mathematical entities. To make this case analogous with the language case, suppose that no plenitude principle, according to which some pure mathematical entities exist if they can consistently exist, is true. Arguably, we have intuitions that promise to speak to the issue of which mathematical entities then exist. If so, then presumably what we think is that the entities quantiﬁed over in relatively natural mathematical theories like ZF and PA exist but the entities quantiﬁed over in seemingly unnatural theories do not, where perhaps Quine’s NF is the prime example of an unnatural theory. Of course serious questions can be raised about the reliability of our intuitions about these matters. My point is just that in this case it can at least be reasonably urged that we have the relevant intuitions. By contrast, do we even have intuitions about what type of negation really exists? If there are sufﬁcient Dummettian restrictions, then maybe the liar-related restrictions do not present this sort of problem. For maybe there are not then two otherwise possible but incompatible languages. But in lieu of actual arguments to the effect that there are sufﬁcient Dummettian restrictions, this is only a pious hope.³⁹

3.6 Revenge Problems Let us now return to revenge problems. Such problems are often assumed to be lethal to the purported solutions that face them. Are they? This question is obviously closely tied to the questions that the discussion thus far has focused on, about universality and about whether certain properties simply fail to exist. There are two main ways to deal with purported revenge problems. One obvious way to deal with a revenge problem is to deny semantic self-sufﬁciency. This is the way of dealing with a revenge problem that corresponds to views (RI) and (I) on the ³⁹ Let me also stress an independent reason to be concerned with the interlanguage paradoxes. If indeed our semantic intuitions are inconsistent, as suggested above, and the correct semantic theory of (the relevant fragment of ) English is just the theory that does the best job of capturing these inconsistent intuitions, then the following possibility should easily suggest itself: it is simply semantically indeterminate what is the semantic value of ‘true’ (and of other key expressions in the liar reasoning and its variants). There are different semantic theories assigning different semantic values to ‘true’, such that nothing about whatever determines the semantic value of ‘true’ determines between these theories. The point is naturally put in terms of possible languages. Nothing determines, among these possible languages, which one it is that we speak. There is a range of languages such that, for all that is determined about what is our actual language, any one of them could be our natural language. Call these possible languages ‘candidate-languages’. The relevance of the interlanguage paradoxes is that they show that we cannot take for granted that these candidate-languages all exist; perhaps at most one does, in which case the semantic values of the relevant expressions can be determinate after all.

The Liar Paradox, Expressibility, Possible Languages / 73 expressive power of natural language. One problem with such a strategy is that it can seem that the relevant property clearly is expressible already in English. (Consider a view according to which truth is not expressible in English, or according to which it is not expressible in English that a sentence is unsettled or false.) Call this type of problem incredulity. A second problem is that of whether, for instance because of the reasoning of the interlanguage paradox discussed earlier, it just is not coherent to maintain that some property is expressible only in another language: for the relevant property may anyway be indirectly expressible, and that is sufﬁcient for problems to arise. Call this second problem the interlanguage problem. The interlanguage problem arises for some, but by no means all, views of this general kind. Given that general arguments for semantic self-sufﬁciency like those of McGee and Priest are ﬂawed, I do not see what other problem besides these two could threaten to arise generally for this type of way of dealing with the revenge problem. In fact, I suspect that even those discussions of revenge problems that on the face of it focus on self-sufﬁciency are motivated by what I called the incredulity problem: the problem is not ascent into a richer metalanguage per se but rather that the property said to be expressible only in this other language is one that we seem already to be able to express. Those who want to avoid having to kick up properties into the metalanguage because of the liar paradox—that is, those who do not accept views (RI) or (I)—seek to show that no revenge problems arise for their theories. Normally this is done by an attempt to demonstrate that the expressive resources employed in stating the solution can be allowed into the object language with no untoward consequences: that our language can be taken to be self-sufﬁcient. Suppose this is successfully done for some given theory of the liar. This does not necessarily avoid either the incredulity problem or the interlanguage problem. Take ﬁrst the incredulity problem. Even if none of the expressions employed in a particular account of the liar has to be conceived of as belonging to a richer metalanguage, the theory can still have the consequence that some property that intuitively seems expressible in English in fact is not expressible. So the incredulity problem can still arise.⁴⁰ What is more, even if the theory’s account of the truth and satisfaction predicates is consistent and even if no untoward consequences follow from allowing the expressive resources employed in stating the solution into the object language, the resources for an interlanguage liar paradox may still be available, so long as (i) there is some language with a predicate expressing a property which on pain of contradiction cannot be expressed by any predicate in the object language, and (ii) such a property is thereby indirectly expressible already in the object language. Let me close by making a few brief remarks on the incredulity problem. One might think that the incredulity problem is anyway the really hard one: that all otherwise ⁴⁰ See again the discussion of Field’s theory of truth in Scharp (this volume).

74 / Matti Eklund acceptable theories of the liar paradox end up deeming inexpressible in a given natural language properties that clearly are expressible in that language. There are two ways to respond to this suspicion. One is of course to attempt to devise a theory which does not deem any intuitively expressible properties inexpressible. A second, more original and theoretically more involved suggestion involves being more careful about expressibility. Consider the kind of view on the liar paradox I have myself elsewhere defended—an inconsistency view, I will here call it.⁴¹ According to this view, the liar paradox arises because meaning-constitutive principles for expressions of our language are inconsistent, where a meaning-constitutive principle is something it is part of semantic competence with the relevant expressions to be disposed to accept.⁴² I will not here attempt to defend this sort of view. But what I wish to stress is that this sort of view provides the materials for a promising response to the incredulity problem. On an inconsistency view one can say that the meaning-constitutive principles for the logical connectives are that they satisfy the inference rules characteristic of classical logic and a meaning-constitutive principle for the truth predicate is that it satisfy the T-schema. Tarski’s theorem then shows that the semantic values of the relevant expressions cannot possibly make true all the relevant meaning-constitutive principles. The semantic values of the relevant expressions are then what come closest to satisfying the relevant meaning-constitutive principles, possibly given other constraints. Suppose—I am only concerned to give a ‘model’ here—that the truth predicate fails to satisfy the T-schema. It is then fairly natural to say, absurd though it may sound, that truth fails to be expressible in English. There is no predicate of English that expresses what the truth predicate aims to express; no predicate that satisﬁes the meaning-constitutive principles for the truth predicate. And even if saying that this means that truth fails to be expressible in English is unnecessarily paradoxical, that does not affect the main point. The inconsistency view provides the materials for a way of getting around incredulity problems. We can respect the idea that truth obviously is expressible in English by noting that there is a predicate of English which aims to express it (or, to cash this out, which is governed by the right meaning-constitutive principles). We can respect the result that truth does not seem to be expressible in English by noting that (on the assumptions mentioned) there is no predicate of English whose semantic value satisﬁes the meaning-constitutive principles for ‘true’. Let an expression misﬁre iff its semantic value fails to satisfy the associated meaningconstitutive principles. The point is then that on an inconsistency view some expressions will misﬁre. Speciﬁcally, one can think that there can be properties such ⁴¹ See Eklund (2002). ⁴² There are many important matters concerning formulation that I here slide over. For instance, it is far from clear that this talk of dispositions to accept is what best describes what semantic competence involves.

The Liar Paradox, Expressibility, Possible Languages / 75 that any expression of a given language that aim to express this property will misﬁre. Such a property is in the most straightforward sense not expressible in the language in question. But while the property is inexpressible there may be predicates that aim to express it. Notice that if this is how the incredulity problem is dealt with, it is far from clear why a view of type (I) should be preferable to a view of type (RI). For the reason that a type (I) view would be preferable is that it is more palatable that a seemingly ‘technical’ property should be expressible only in a richer language than that an ‘ordinary’ property which ordinary speakers seem clearly able to express should be expressible only in a richer language. But given the route just outlined, a property deemed inexpressible is one that at the same time is claimed that we aim to express. And it is far clearer that we aim to express an ‘ordinary’ property such as that of being true than that we actually succeed in this aim.⁴³

References Barwise, Jon, and Etchemendy, John (1987). The Liar: An Essay on Truth and Circularity, Oxford University Press, Oxford Beall, JC (ed.) (2003). Liars and Heaps: New Essays on Paradox, Clarendon Press, Oxford Beall, JC, and Armour-Garb, Bradley (eds.) (2005). Deﬂationism and Paradox, Oxford University Press, Oxford Belnap, Nuel (1982). ‘Gupta’s rule of revision theory of truth’, Journal of Philosophical Logic 11: 103–16 Burge, Tyler (1979). ‘Individualism and the mental’. In Peter French, Theodore Uehling, and Howard Wettstein (eds.), Midwest Studies in Philosophy IV, University of Minnesota Press, Minneapolis, pp. 73–121 Eklund, Matti (2002). ‘Inconsistent languages’, Philosophy and Phenomenological Research 64: 251–75 Field, Hartry (2003a). ‘A revenge-immune solution to the semantic paradoxes’, Journal of Philosophical Logic 31: 1–27 (2003b). ‘The semantic paradoxes and the paradoxes of vagueness’, in Beall (2003) (2005a). ‘Is the liar sentence both true and false?’, in Beall and Armour-Garb (2005) ⁴³ Note that if the view sketched in the last few paragraphs is accepted, and view of type (RI) is acceptable, then many of the statements in this chapter will have to be taken with a ‘grain of salt’ (compare Frege (1892), p. 192): for some property that I aim to express will not in fact be expressible. But in contrast with other theorists who ask not to be begrudged a pinch of salt, I have a theoretical explanation of what is going on when my expressions do not in fact express what they are meant to express. An earlier version of this chapter was presented as a paper at a conference on paradoxes at the University of Latvia, November 2005. Thanks to the conference participants for good discussion. I also wish to thank Dan Korman, Agust´ın Rayo, and Kevin Scharp for helpful comments on earlier versions of this chapter.

76 / Matti Eklund Field, Hartry (2005b). ‘Variations on a theme by Yablo’. In Beall and Armour-Garb (2005) this volume, ‘Solving the paradoxes, escaping revenge’ Fitch, Frederic (1946). ‘Self-reference in philosophy’, Mind 55: 64–73 (1964). ‘Universal metalanguages for philosophy’, Review of Metaphysics 17: 396–402 Frege, Gottlob (1892). ‘On concept and object’. In Michael Beaney (ed.), The Frege Reader, Blackwell, Oxford (1997), pp. 181–93 Gupta, Anil (1982). ‘The liar paradox’, Journal of Philosophical Logic 11: 1–60. Reprinted in Martin (1984) (1997). ‘Deﬁnition and revision: a response to McGee and Martin’. In Villanueva (1997), pp. 419–43 (2006). ‘Truth and Paradox: Solving the Riddles, by Tim Maudlin’, Mind 115: 163–5 Gupta, Anil, and Belnap, Nuel (1993). The Revision Theory of Truth, MIT Press, Cambridge, Mass. Herzberger, Hans (1970). ‘Paradoxes of grounding in semantics’, Journal of Philosophy 67: 145–67 (1982). ‘Notes on naive semantics’, Journal of Philosophical Logic 11: 61–102. Reprinted in Martin (1984) (1982a). ‘Naive semantics and the liar paradox’, Journal of Philosophy 79: 479–97 Hofweber, Thomas (2006). ‘Inexpressible properties and propositions’. In Dean Zimmerman (ed.), Oxford Handbook of Metaphysics, vol. 2 Kripke, Saul (1975). ‘Outline of a theory of truth’, Journal of Philosophy 72: 690–716. Reprinted in Martin (1984) (1980). Naming and Necessity, Harvard University Press, Cambridge, Massa. Lewis, David (1983). ‘New work for a theory of universals’, Australasian Journal of Philosophy 61: 343–77 (1984). ‘Putnam’s paradox’, Australasian Journal of Philosophy 62: 221–36 Martin, Robert L (1976). ‘Are natural languages universal?’, Synth`ese 32: 271–91 (1984). Recent Essays on Truth and the Liar Paradox, Clarendon Press, Oxford (1984a). ‘Introduction’. In Martin (1984), pp. 1–8 Maudlin, Tim (2004). Truth and Paradox: Solving the Riddles, Clarendon Press, Oxford McGee, Vann (1991). Truth, Vagueness and Paradox, Hackett Publishing Company, Indianapolis (1994). ‘Afterword: truth and paradox’. In Robert M. Harnish (ed.), Basic Topics in the Philosophy of Language, Prentice Hall, Englewood Cliffs, NJ, pp. 615–33 (1997). ‘Revision’, in Villanueva (1997), pp. 387–406 Parsons, Terence (1990). ‘True contradictions’, Canadian Journal of Philosophy 20: 335–53 Priest, Graham (1984). ‘Semantic closure’, Studia Logica 43: 117–29 (1987). In Contradiction, Kluwer Academic Publishers, Dordrecht (1990). ‘Boolean negation and all that’, Journal of Philosophical Logic 19: 201–15 (1995). ‘Beyond gaps and gluts: reply to Parsons’, Canadian Journal of Philosophy 25: 57–66 (2005). Doubt Truth to Be a Liar, Oxford University Press, Oxford Putnam, Hilary (1975). ‘The meaning of ‘‘meaning’’ ’. In Keith Gunderson (ed.), Language, Mind, and Knowledge, University of Minnesota Press, Minneapolis, pp. 131–93 Scharp, Kevin: this volume, ‘Aletheic vengeance’ manuscript, ‘Truth and internalizability’

The Liar Paradox, Expressibility, Possible Languages / 77 Simmons, Keith (1993). Universality and the Liar, Cambridge University Press, Cambridge Soames, Scott (1999). Understanding Truth, Oxford University Press, Oxford Tarski, Alfred (1935/83). ‘The concept of truth in formalized languages’. In John Corcoran (ed.), Logic, Semantics, Metamathematics: Papers from 1923 to 1938, 2nd edn, Indianapolis: Hackett Publishing Company. English translation by J. H. Woodger of ‘Der Wahrheitsbegriff in Formalisierten Sprachen’, Studia Philosophica 1 (1935) Villanueva, Enrique (ed.) (1997). Truth, Ridgeview, Atascadero, Calif. Yablo, Stephen (2003). ‘New grounds for naive truth theory’. In Beall (2003), pp. 312–30

4 Solving the Paradoxes, Escaping Revenge Hartry Field

It is ‘the received wisdom’ that any intuitively natural and consistent resolution of a class of semantic paradoxes immediately leads to other paradoxes just as bad as the ﬁrst. This is often called the ‘revenge problem’. Some proponents of the received wisdom draw the conclusion that there is no hope of any natural treatment that puts all the paradoxes to rest: we must either live with the existence of paradoxes that we are unable to treat, or adopt artiﬁcial and ad hoc means to avoid them. Others (‘dialetheists’ ) argue that we can put the paradoxes to rest, but only by licensing the acceptance of some contradictions (presumably in a paraconsistent logic that prevents the contradictions from spreading everywhere).¹ I think the received wisdom is incorrect. In my effort to rebut it, I will focus on a certain type of solution to the paradoxes. This type of solution has the advantage of keeping the full Tarski truth schema (T) True(A ) ↔ A ¹ This latter view is only reasonable if ‘revenge’ is less of a worry for inconsistent solutions to the paradoxes than for consistent ones. I think myself that advocates of inconsistent solutions face a prima-facie revenge problem, and doubt that they can escape it without employing the devices I suggest in this chapter on behalf of certain consistent solutions. But that is a matter for another occasion.

Solving the Paradoxes, Escaping Revenge / 79 (and more generally, a full satisfaction schema). This has a price, namely that we must restrict both the law of excluded middle and the law connecting A → B with ¬A ∨ B, but we can carve the restrictions narrowly enough so that ordinary reasoning (e.g. in mathematics and physics) is unaffected.² I’ll call solutions of this type G-solutions. (If you want to think of the ‘G’ as standing for ‘good’ I won’t stop you.) The literature contains several demonstratively consistent solutions of this sort; for purposes of this chapter there is no need to choose between them. (I will give an informal introduction to this type of solution in sections 1, 3, and 4, and a formal account in section 5.) It will turn out that any such solution generates certain never-ending hierarchies of sentences that may seem ‘increasingly paradoxical’ (roughly speaking, it is harder to ﬁnd a theory that satisfactorily treats later members of the hierarchy than to ﬁnd one that satisfactorily treats earlier members); but the G-solution gives a consistent treatment of each member of each such hierarchy. The existence of these hierarchies prevents certain kinds of revenge problems from arising: certain attempts to state revenge problems simply involve going up a level in a hierarchy all levels of which have been given a non-paradoxical treatment. Still, there are certain ‘vindictive strategies’ (strategies for trying to ‘get revenge’ against solutions to the paradoxes) that G-solutions may seem to be subject to. I’ll argue that the most popular such strategy is based on a misunderstanding of the signiﬁcance of model-theoretic semantics. But there is a far more interesting strategy for which this is not so. As mentioned, a G-solution generates certain never-ending hierarchies of apparently paradoxical sentences which however are each successfully treated by the account. But shouldn’t it be possible to ‘break out of the hierarchies’ to get paradoxes that are not resolved by the account? Or to put it another way: If we can’t ‘break out of the hierarchies’ within the language that our solution to the paradoxes treats, isn’t that simply due to an expressive limitation in that language? That, I think, is the most difﬁcult revenge worry for G-solutions to deal with. Some of the worries about ‘breaking out of the hierarchies’ turn out to be intimately connected to K¨onig’s paradox of the least undeﬁnable ordinal, a paradox in the same ballpark as those of Berry and Richard. The G-solutions provide a consistent treatment of these deﬁnability paradoxes. The treatment of K¨onig’s paradox will be an important element in my argument that we are unable to ‘break outside the hierarchies’, but that this does not reﬂect an expressive limitation of the language. ² Also, the equivalence between A → B and ¬A ∨ B will hold on the assumption of excluded middle for A and for B.

80 / Hartry Field

Part One Introductory Discussion 1 The paradoxes and excluded middle Imagine that we speak a ﬁrst-order language L: it has the usual connectives and quantiﬁers, and it contains no ambiguous terms and no indexicals. It is to be a very rich language, powerful enough to express all our mathematics, including the richest set theory we currently know how to develop. It should be able to talk about its own expressions and their syntax; though we needn’t actually make this a separate requirement, since as G¨odel showed we can use arithmetical surrogates. If we like we can also assume that L can express all our current claims about the physical world too, though this will not really matter to the problem to be discussed. Finally, L should contain terms like ‘true’ and ‘true of’. For present purposes we needn’t worry about how such terms apply to sentences and formulas in languages other than our own, so we may as well assume that they have been restricted to apply only to the sentences and formulas of L. These assumptions about our language L are enough to generate paradoxes (or rather, apparent paradoxes). Some of them, like the Liar paradox, arise from the fact that by any of a number of well-known routes we can construct self-referential sentences: sentences that attribute to themselves any property you like. For instance, the Liar paradox arises from any sentence that directly or indirectly asserts its own untruth; let Q be some such sentence, and Q its standard name.³ Since Q asserts its own untruth, it certainly seems that Q ↔ ¬True(Q ) had better be part of our overall theory. In addition, it seems that our theory of truth ought to include every ‘Tarski biconditional’ , i.e. every instance of the schema (T) mentioned earlier; hence in particular, True(Q ) ↔ Q. ³ There is a familiar distinction between contingent and non-contingent Liar sentences. If the sentence ‘Nothing written on the ﬁrst blackboard manufactured in 2005 is true’ is written on a blackboard that (perhaps unbeknownst to the writer) was the ﬁrst to be manufactured in 2005, it is a contingent Liar: given the contingent facts about blackboard-manufacture it in effect asserts its own untruth. Non-contingent Liar sentences assert their own untruth independent of such empirical facts. Some of the formulations below are only strictly correct for non-contingent Liars, but could easily be generalized to apply to contingent Liars too. (For instance, in the next sentence of the text, replace ‘be part of our overall theory’ by ‘follow from our overall theory together with the empirical facts’.) The distinction between the two kinds of Liar sentences will make no important difference.

Solving the Paradoxes, Escaping Revenge / 81 But then if the conditional, and the biconditional deﬁned from it in the obvious way, are at all reasonable, we can infer (∗ )

True(Q ) ↔ ¬True(Q ).

And being of form B ↔ ¬B, this leads to contradiction in classical logic. There are similar paradoxes that don’t require the construction of self-referential sentences. For instance, just as our theory of truth ought to include the instances of Schema (T), so our theory of satisfaction ought to include the instances of the following schema: (S) For all x, x satisﬁesA(v) if and only if A(x) (where to say that x satisﬁes A(v) is the same as saying that A(v) is true of x).⁴ In the special case where A(v) is the formula ‘v does not satisfy v’, this yields that for all x, v does not satisfy v satisﬁes x if and only if x does not satisfy x, and hence (∗∗ )

v does not satisfy v satisﬁes itself if and only if it does not satisfy itself. ⁵

Again, (∗∗ ) is of form B ↔ ¬B and hence leads to contradiction in classical logic. The ‘G-solutions’ that I’ll be considering accept these derivations of (∗ ) and (∗∗ ). But unlike ‘dialetheic’ views (e.g. [16] ), they do not accept contradictions (sentences of form C ∧ ¬C). So they must reject all arguments that would take us (for arbitrary B) from B ↔ ¬B to a sentence of form C ∧ ¬C. I think that the most revealing way of trying to argue from B ↔ ¬B to a contradiction is as follows: (i) Assume both B ↔ ¬B and B. Then by modus ponens, ¬B; so B ∧ ¬B. (ii) Assume both B ↔ ¬B and ¬B. Then by modus ponens, B; so B ∧ ¬B. ⁴ (S) should really be called (S1 ): it is the schema for the satisfaction predicate ‘satisﬁes1 ’ that applies to formulas with exactly one free variable, and there is an analogous schema (Sn ) for each satisfaction predicate ‘satisﬁesn ’ that applies to formulas with exactly n free variables. But (S1 ) can be taken as basic: in any language rich enough to code ﬁnite sequences we can artiﬁcially deﬁne the higher satisfaction predicates in terms of 1-place satisfaction, in a way that guarantees the schemas for the former if we have (S1 ): e.g. to say that ‘v1 is larger than v2 ’ is satisﬁed by o1 and o2 in that order is in effect to say that ‘u is an ordered pair whose ﬁrst member is larger than its second’ is satisﬁed by o1 , o2 . We can similarly reduce truth to satisfaction: to say that ‘Snow is white’ is true is in effect to say that ‘Snow is white and u = u’ is satisﬁed by everything (or equivalently, by something). (T) then falls out of (S1 ), so (S1 ) can be taken as the sole basic schema. But it’s more natural to talk in terms of truth than satisfaction, so I’ll keep on talking about (T). ⁵ A natural abbreviation for ‘v satisﬁes itself’ would be ‘v is onanistic’. But for some reason ‘v is homological’ has caught on instead, with ‘heterological’ for ‘non-onanistic’.

82 / Hartry Field (iii) Since B ∧ ¬B follows both from the assumptions B ↔ ¬B and B and from the assumptions B ↔ ¬B and ¬B, then it follows from the assumptions B ↔ ¬B and B ∨ ¬B. (Reasoning by cases.) (iv) But B ∨ ¬B is a logical truth, so B ∧ ¬B follows from B ↔ ¬B alone. I now further stipulate that G-solutions accept both modus ponens and reasoning by cases (aka disjunction elimination). So they take the reasoning to be valid through step (iii). What G-solutions question is the use of the law of excluded middle in step (iv). Unlike intuitionists, though, G-theorists take excluded middle to be perfectly acceptable within standard mathematics, physics, and so forth; it is only certain reasoning using truth and related concepts that are affected.⁶ There is a verbal issue here about exactly how this point should be put. One way to put it is to say that excluded middle is literally valid in some contexts like mathematics, but invalid outside that domain. But it might be thought that the ‘topic neutrality’ of logic implies that if excluded middle can’t be accepted everywhere then it can’t be taken as literally valid anywhere. Even so, this doesn’t undermine the claim that it is effectively valid⁷ in contexts like mathematics: if one accepts all instances of the schema A ∨ ¬A that don’t contain ‘true’, then even if one doesn’t claim that they are logical truths one can reason from them just as a classical logician reasons in mathematics and physics. So it really makes no difference in which of the two ways we talk. Another argument from B ↔ ¬B to a contradiction runs as follows: after step (i) as above, we conclude that B ↔ ¬B entails ¬B by a reductio rule (that if X and B together entail ¬B, then X alone entails ¬B); that result and (ii) then give the contradiction. But the most obvious argument for that reductio rule is based on the law of excluded middle (together with reasoning by cases). The argument is that if X and B together entail ¬B, then since X and ¬B certainly entail ¬B, it follows that X and B ∨ ¬B entails ¬B; and since B ∨ ¬B is a logical truth, this means that X entails ¬B. So I will ⁶ Actually advocates of G-solutions might want to further restrict excluded middle, e.g. by disallowing its application to certain sentences containing vague concepts; and indeed it is not out of the question to regard certain mathematical concepts such as ‘ordinal’ as having a kind of ‘indeﬁnite extensibility’ that is akin to vagueness. Still, for purposes of this chapter I assume that excluded middle applies unrestrictedly within standard mathematics. Another plausible restriction of excluded middle is to sentences containing normative concepts like ‘appropriate’ or ‘reasonable’; this is relevant to certain ‘doxastic paradoxes’ involving, for instance, sentences asserting that it is not appropriate to believe them. But such paradoxes are outside the scope of this chapter. ⁷ ‘Effectively valid’ means ‘in effect valid’: it has nothing to do with effective procedures. Similarly I’ll use ‘effectively classical’ to mean ‘in effect classical’, i.e. excluded middle holds even if not as a logical law.

Solving the Paradoxes, Escaping Revenge / 83 assume that in giving up (or restricting) excluded middle we give up (or restrict) this reductio rule as well. Admittedly, this reductio rule is valid in intuitionist logic even in absence of excluded middle, so I can’t say that we are compelled to give up the reductio rule if we give up excluded middle. But intuitionist logic does not evade the paradoxes, so we had best not follow its lead.⁸ My point is that there is a natural response to the Liar paradox which sees this kind of reductio reasoning as depending on the law of excluded middle and both as needing restriction; and that is the response that G-solutions adopt.

2 Trying to preserve classical logic Weakening classical logic to deal with the paradoxes is obviously not something to be done lightly, and there are questions about how to understand the proposal, some of which I will address in the next section. But ﬁrst I’d like to brieﬂy survey the options for handling the paradoxes within classical logic. One reason for doing this is to make evident the appeal of the non-classical approach, and another is to facilitate a later discussion of the ‘hierarchies of paradoxical sentences’ that arise within G-solutions. In classical logic, the reasoning of the Liar paradox can easily be turned into a proof of the following disjunction: Either (i) Q is true, but ¬Q or (ii) Q is not true, but Q. At this point, classical theorists have three options. (Of course, there is also the possibility of remaining agnostic between the options, but that is of no particular interest.) The ﬁrst option is to choose disjunct (i). This would seem quite unattractive: doesn’t calling Q true while saying ‘nonetheless, ¬Q’ deprive the notion of truth of signiﬁcance? The second option is to choose disjunct (ii). This seems on its face almost equally unattractive: if one holds that Q is not true, what is one doing holding Q? ⁸ Intuitionists tend to motivate the reductio rule by way of the law ¬(A ∧ ¬A) (sometimes misleadingly called the ‘law of non-contradiction’ ). But to anyone who accepts the deMorgan law ¬(A ∧ B) |= ¬A ∨ ¬B, this version of the ‘law of non-contradiction’ simply amounts to ¬A ∨ ¬¬A, a slightly restricted version of excluded middle that few who reject excluded middle would accept. (That’s why dialetheists who accept excluded middle accept ¬(A ∧ ¬A), making clear that it does not adequately capture the principle that we should reject contradictions.) The intuitionist argument for reductio thus turns on their rejection of the deMorgan law.

84 / Hartry Field The third option is to accept the disjunction of (i) and (ii) while ruling out as absurd the acceptance of either disjunct. (It is because the acceptance of either disjunct is viewed as absurd that this is really a third option, distinct from agnosticism between the ﬁrst two options). This third option takes the acceptance of either (i) or (ii) to be absurd, on the ground that commitment to A requires commitment to A being true and conversely; but it nonetheless allows commitment to A ∨ ¬A. Now, many people think that if one accepts a disjunction of two options each of which would be absurd to accept, one has already accepted an absurdity. Indeed, that principle appears to be built into classical logic: it is the principle of reasoning by cases (or disjunction elimination), to which attention was called above. This third option is based on rejecting that principle, except in restricted form.⁹ So it is probably best thought of as only a semi-classical option: it does accept all the validities of classical logic, but disallows natural applications of disjunction elimination and some of the other standard meta-rules. These three options seem to be the only possibilities for keeping the validities of classical logic without accepting contradictions.¹⁰ Admittedly, one could insist with Tarski that the predicate ‘true’ should be given a hidden subscript, or that its extension ⁹ The restricted form is that if together with A entail C by classical rules, and together with B entail C by classical rules, then together with A ∨ B entail C. The third option can accept that, but cannot accept the generalization to ‘entailment’ by the truth rules (that commitment to A requires commitment to True(A ) and conversely). And these truth rules must have a quasi-logical status on the third option, since it was only by holding acceptance of (i) and of (ii) to be absurd that the view differentiated itself from agnosticism between the ﬁrst two options. ¹⁰ I know of no one who has seriously proposed taking the ﬁrst option. Classical and semi-classical logicians who do technical work on the paradoxes mostly tend to prefer the third option: see [10] and [14]; also [9], where ﬁve of the nine types of theories discussed fall under option three. The option of choice among non-specialists seems to be option two, but some specialists prefer it as well, e.g. [2] and [13]. (If the description of the latter as a classical theory seems surprising, see [8].) What about Kripke’s seminal [11]? That’s more complicated since Kripke offers a model-theoretic semantics with no instructions on how to read the theory off the semantics. But if we interpret him as suggesting that though the extension of ‘True’ is a ﬁxed point, the logic is classical, then his theory also falls under option two. An alternative and I think more attractive interpretation of Kripke is to take the set of acceptable sentences to coincide with the extension of ‘True’: they are both the contents of the same ﬁxed point. But if the ﬁxed points are based on a Kleene semantics, this gives a non-classical logic, and so is not germane to the discussion in this section. (This way of interpreting Kripke has been advocated in [21]—not altogether consistently, in my view, since Soames talks in terms of truth-value gaps, which seems to presuppose the classical logic interpretation. [19] clearly distinguishes the two ways of getting a theory from a Kripkean ﬁxed point, in the distinction between the theories there called KF and KFS.) On the non-classical reading of Kripke, his solution is similar in sprit to the G-solutions under discussion in this chapter; however, the nonclassical logic one obtains from this way of reading Kripke is unsatisfactorily weak, since Kleene semantics has no serious conditional. G-solutions do much better in this regard.

Solving the Paradoxes, Escaping Revenge / 85 vary with context. Still, given classical logic (even in the weak sense that includes only the validities and not the meta-rules), the above three options are the only consistent ones when the subscript and context are held ﬁxed.¹¹ A problem with all of the classical and semi-classical solutions is that they prevent the notion of truth from fulﬁlling its generally accepted role. The standard story about why we need a notion of truth ([18], [12]) is that we need it to make certain kinds of generalizations. For instance, the only way to generalize over (Snow is white) → ¬¬(Snow is white) (Grass is green) → ¬¬(Grass is green)], is to ﬁrst restate them in terms of truth and then generalize using ordinary quantiﬁers: For every sentence, if it is true, so is its double negation. But this says what we want it to say only if we assume the intersubstitutivity of True(A ) with A: that is, the principle Intersubstitutivity: If Y results from X by replacing some occurrences of A with True(A ), then X and Y entail each other. [This needs to be restricted to cases where the substitution is into contexts that aren’t quotational, intentional, etc.; but I’ll take the language L to contain no such contexts.]

This principle entails the truth schema in classical logic, indeed in any logic in which A ↔ A is a logical truth. So the three classical and semi-classical options all reject the intersubstitutivity principle. Thus they fail to satisfy the purpose of the notion. For instance, relative to the assumption that what Jones said was exactly A1 , . . . ,An , we want If everything Jones said is true then to be equivalent to If A1 and . . . and An then

.

This requires that the True(Ai ) be intersubstitutable with the Ai inside the conditional, but that won’t in general be so on any of the classical and semi-classical theories. The semi-classical theory does better than the fully classical ones: it allows for intersubstitutivity of True(A ) with A in more contexts. Indeed the fully classical ones don’t even allow substitutivity for unembedded occurrences: True(A ) and A can’t be mutually entailing in a classical theory that includes disjunction elimination (as we’ll see in the next section). But though the semi-classical theories do better, that ¹¹ I’m putting aside solutions to the Liar paradox based on unmotivated syntactic restrictions that prevent the formation of self-referential sentences. Very strong syntactic restrictions are required for this, and the solutions are of little interest since they do not generalize to the heterologicality paradox.

86 / Hartry Field isn’t good enough. An advantage of G-solutions is that not only do they accept the Tarski schema, they accept the full Intersubstitutivity Principle. I conclude this section with some further remarks on the second classical option; in particular, on a version of the second classical option that invokes a hierarchy of truth predicates. This will play a role later in the chapter, where I will compare it to a hierarchy of strengthenings of a single truth predicate that arises in G-solutions. A common theme among proponents of the second option is that the schema (T) holds for all sentences that ‘express propositions’, where to ‘express a proposition’ is to be either true or false, i.e. to either be true or have a true negation. (On this view, expressing a proposition is much stronger than being meaningful: it would be hard to argue that the ‘contingent Liar sentences’ of note 3 aren’t meaningful, but proponents of this option take them not to express propositions.) So instead of (T) we have (RT) [True(A ) ∨ True(¬A )] → [True(A ) ↔ A]. It is easily seen that this is equivalent to the left-to-right half of (T), i.e. to (LR) True(A ) → A.¹² Obviously, then, a decent theory of truth containing (RT)/(LR) needs to contain vastly more. (It’s compatible with (RT)/(LR) that nothing is true; or that only sentences starting with the letter ‘B’ are true; etc.) The crucial question for such a theory is, how are we to ﬁll it out without leading to paradox? It turns out that however we try to ﬁll it out, we are led to the conclusion that basic principles of the truth theory itself fail to be true. (They also fail to be false, so that they come out as ‘not expressing propositions’ .) Of course any non-contingent Liar sentence is itself an assertion of the theory that the theory asserts not to be true, but it presumably doesn’t count as one of the basic principles of the theory. But what does count as a basic principle of the theory is (RT), or its equivalent (LR). And Montague [15] showed that with very minimal extra assumptions, one can derive from (LR) a conclusion of form ¬True[True(M ) → M ], i.e. that some speciﬁc instance of (LR) isn’t true. Most people regard it as a serious defect of a theory that it declares central parts of itself untrue; and saying that these parts ‘don’t express propositions’ doesn’t appear to help much. This is the point at which the idea of a hierarchy of truth predicates may suggest itself. The idea is that we don’t have a general truth predicate, but only a hierarchy of predicates ‘trueα ’, where the subscripts are notations for ordinal numbers (in a ¹² In proving a given instance of schema (RT) from schema (LR) one uses two instances of the latter, one for A and one for ¬A.

Solving the Paradoxes, Escaping Revenge / 87 suitably large initial segment of the ordinals that has no last member). The Montague theorem then shows that the principles of the theory of truthα aren’t trueα , but the idea is to try to ameliorate this by saying that they’re trueα+1 . Call such a view a stratiﬁed truth theory. Besides their artiﬁciality, stratiﬁed truth theories seriously limit what we can express, in a way that undermines the point of the notion of truth. For instance, suppose we disagree with someone’s overall theory of something, but haven’t decided which part is wrong. The usual way of expressing our disagreement is to say: not all of the claims of his theory are true. Without a general truth predicate, what are we to do? The only obvious idea is to pick some large α, and say ‘Not all of the claims of his theory are trueα ’ . But this is likely to fail its purpose since we needn’t know how large an α we need. (Indeed, there would be strong pressure on each of us to use very high subscripts α even in fairly ordinary circumstances, but however high we make it there is a signiﬁcant risk of it not being high enough to serve our purposes. This was the lesson of the famous discussion of Nixon and Dean in [11]. Nixon and Dean wanted to say that nothing the other said about Watergate was true, and to include those assertions of the other in the scope of their own assertions; but to succeed, each needed to employ a strictly higher subscript than the other.) Actually, the situation is even worse than this. For suppose that we want to express disagreement with a stratiﬁed truth theorist’s overall ‘theory of truth’ (i.e. the theory he expresses with all of his ‘trueα ’ predicates), but that we haven’t decided which part of that theory is wrong. Here the problem isn’t just with knowing how high an α to pick; rather, no α that we pick could serve its purpose. The reason is that it’s already part of the stratiﬁed theory that some of its claims aren’t trueα , namely, the principles about truthα ; that’s why the theorist introduced the notion of truthα+1 . So whatever α we pick, we won’t succeed in expressing our disagreement. The problems just mentioned are really just an important special case of a problem that I’ve argued to infect all classical and semi-classical theories: they can’t give truth its proper role as a device of generalization. Except possibly for dialetheic theories, which I will not consider here, restricting excluded middle seems to be the only way to avoid crippling limitations on our notion of truth.

3 More on rejecting excluded middle It is important to note that in classical logic you don’t need anything like the full strength of the truth schema (T) (or the satisfaction schema (S)) to derive contradictions: indeed, if you allow reasoning by cases as well as the classical validities, all that is required is the two assumptions (T-Elim)

A follows from True(A )

88 / Hartry Field and (T-Introd)

True(A ) follows from A

(or the analogous Elimination and Introduction rules for satisfaction). For using these instead of (T), we can easily recast the derivation (i)-(iv) in Section 1 (with True(Q ) as the B) as follows: (i*) By (T-Elim), True(Q ) implies Q,¹³ which is equivalent to ¬True(Q ); hence True(Q ) implies the contradiction True(Q ) ∧ ¬True(Q ); (ii*) ¬True(Q ) is equivalent to Q, which by (T-Introd) implies True(Q ); hence ¬True(Q ) also implies the contradiction True(Q ) ∧ ¬True(Q ). (iii*) Since True(Q ) ∧ ¬True(Q ) follows both from the assumption True(Q ) and from the assumption ¬True(Q ), then it follows from the assumption True(Q ) ∨ ¬True(Q ). (Reasoning by cases.) (iv*) But True(Q ) ∨ ¬True(Q ) is a logical truth, so we have a derivation of the contradiction True(Q ) ∧ ¬True(Q ). (If we strengthened (T-Elim) to the assumption of the conditional True(A ) → A, we could give a derivation that doesn’t involve reasoning by cases.) In fact, we don’t even need the full strength of (T-Introd); we can make do with the weaker assumption (T-Incoherence) A and ¬True(A ) are jointly inconsistent. Inconsistency proof: True(Q ) implies Q by (T-Elim), and ¬True(Q ) implies Q since it is equivalent to Q, so we derive Q using reasoning by cases plus excluded middle. Using the other half of the equivalence between Q and ¬True(Q ), we get Q ∧ ¬True(Q ), which is inconsistent by (T-Incoherence). The fact that the paradox arises from weaker assumptions than (T) is important for two reasons. First and most obviously, it means that if we insist on keeping full classical logic we must do more than restrict (T), we must restrict the weaker assumptions as well. But the second reason it’s important concerns not classical solutions, but G-solutions: it gives rise to an important moral for what G-solutions have to be like. For even though G-solutions take truth to obey the Tarski schema (T), we’ll see that they recognize other ‘truth-like’ predicates (e.g. ‘determinately true’) that don’t obey the analog of (T) but do obey the analogs of (T-Elim) and (T-Introd) (or at the very least, (T-Elim) and (T-Incoherence)). For each truth-like predicate, there is a Liar-like sentence that asserts that it does not instantiate this predicate. ¹³ That is, Q follows from True(Q ). (One reader took my ‘A implies B’ to mean ‘if A then B’, and on this basis accused me here of illicitly extending (T-Elim) to hypothetical contexts; but that is not what I mean by ‘implies’.)

Solving the Paradoxes, Escaping Revenge / 89 Reasoning as in (i∗ ) and (ii∗ ) is thus validated, and since G-solutions accept reasoning by cases without restriction, paradox can only be avoided by rejecting the application of excluded middle to these Liar-like sentences formed from truth-like predicates. In short, G-solutions are committed to the view that there can be no truth-like predicate for which excluded middle can be assumed. (Since excluded middle is to hold within ordinary mathematics and physics, this means that no truth-like predicates can be constructed within their vocabulary.) As we’ll see, the conviction that there must be truth-like predicates obeying excluded middle is one primary source of revenge worries. I close this section by trying to make clear what is involved in restricting the application of excluded middle to certain sentences, e.g. the Liar sentence, when one accepts the intersubstitutivity of True(Q ) with Q. In particular, what is the appropriate attitude to take to the claim True(Q )? According to the sort of solution to the paradoxes I’ve sketched, one must reject the claim that True(Q ) and also reject the claim that ¬True(Q ), since these claims each imply a contradiction relative to any theory of truth that implies the Tarski biconditionals. (One can take rejection as a primitive state of mind, involving at the very least a refusal to accept; a slightly more informative account of rejection can be found in [4] (Section 3).) We must likewise reject the corresponding instance of excluded middle Z:

True(Q ) ∨ ¬True(Q ),

for it too leads to contradiction. And because we reject Z, our refusal to either accept True(Q ) or accept ¬True(Q ) doesn’t seem appropriately described as ‘agnosticism’ about the truth of Q. We would be agnostic about True(Q ) if we believed Z but were undecided which disjunct to believe; but when we reject Z the very factuality of the claim that True(Q ) is being put into question, so our not believing True(Q ) while also not believing ¬True(Q ) isn’t happily described as ‘agnosticism’ . It should be immediately noted that a solution of this sort does not postulate a ‘truth-value gap’ in Q: it does not say that Q is neither true nor false, i.e. that neither Q nor its negation is true. It also does not say that Q is neither true nor not true. Saying that Q is ‘gappy’ or ‘non-bivalent’ in either of these senses would trivially entail that Q is not true, which (by the Tarski biconditionals and modus ponens) leads to contradiction. Since the claim that Q is ‘gappy’ (non-bivalent) leads to contradiction, we must reject it. That isn’t to say that we should believe that Q is bivalent (or that it is not ‘gappy’; these are the same, assuming the equivalence of ¬¬A to A, as I henceforth shall). The claim that Q is bivalent or non-gappy amounts to Z∗ : True(Q ) ∨ True(¬Q ). This in turn amounts to Z (non-truth and falsity turn out to coincide as applied to sentences in the language), and as we’ve seen, Z must be rejected.

90 / Hartry Field If it seems odd that both the claim that Q is gappy and the claim that it is not gappy lead to contradiction, it shouldn’t: from the fact that Gappy(Q ) and ¬Gappy(Q ) each lead to contradiction, all we can conclude is that [email protected] :

Gappy(Q ) ∨ ¬Gappy(Q )

leads to contradiction; so the proper conclusion is that this instance of excluded middle must also be rejected.¹⁴ In other words, the claim that Q is ‘gappy’ has the same status as Q itself has. In particular, just as it is misleading to declare ourselves ‘agnostic’ about the Liar sentence, it is also misleading to declare ourselves ‘agnostic’ about the claim that the Liar sentence is ‘gappy’ or the claim that it is bivalent: for we don’t recognize that there is a fact to be agnostic about. I think it would be a serious problem if there were no way to assert the ‘defective’ status of Q within the language. As we’ll see, there is a way; but it can’t be done by saying that Q suffers a truth-value gap.

4 The Berry–Richard–K¨onig paradox I think that all of the semantic paradoxes turn on excluded middle, though some of them (especially some of the ones involving conditionals) do so in an indirect and unobvious fashion. I will make this precise in Section 5. There I will introduce a language that contains a ‘quasi-classical conditional’ which obeys many of the classical laws for conditionals even in the absence of excluded middle; moreover it reduces to the material conditional when excluded middle is assumed for antecedent and consequent. I will then state a result (proved elsewhere) according to which every semantic ‘paradox’ that can be formulated in this language has a solution that is compatible with the Tarski biconditionals. The solution may depend on the failure of some of the classical laws for the conditional, but that failure will always be traceable to a breakdown in excluded middle for the antecedent or consequent of one of the conditionals in question. We thus diagnose these apparent paradoxes as only apparent, they depend on illicit applications of excluded middle. Of course, the fact that those apparent paradoxes that can be formulated in the language turn out not be genuinely paradoxical does not settle the revenge issue: settling that issue requires considering the possibility of expanding the language to get new paradoxes. I will have a lot to say toward undermining the idea of revenge in later sections. ¹⁴ Indeed whenever one rejects a given instance A ∨ ¬A of excluded middle, one should also reject the instance (A ∨ ¬A) ∨ ¬(A ∨ ¬A), for they are equivalent by very uncontroversial reasoning; hence one should reject Bivalent(A ) ∨ ¬Bivalent(A ). [Reason for the equivalence: ¬(A ∨ ¬A) implies ¬A, so (A ∨ ¬A) ∨ ¬(A ∨ ¬A) implies (A ∨ ¬A) ∨ ¬A, which implies A ∨ ¬A. The other direction is trivial.]

Solving the Paradoxes, Escaping Revenge / 91 First though I will consider how the paradoxes of deﬁnability fare on this sort of view. There are a number of slightly different paradoxes of deﬁnability, the most famous being Berry’s and Richard’s, but they all have the same underlying idea. Because of its relevance later in the chapter, I will focus attention on a variant of the Berry and Richard paradoxes due to K¨onig. Recall that L is a ﬁrst order language adequate to expressing its own syntax, and that contains a satisfaction predicate. From that predicate we can deﬁne ‘L-deﬁnable’: z is L-deﬁnable if and only if there is at least one formula of L (with exactly one free variable) that is satisﬁed by z and by nothing else. Now, L is assumed to be built from a ﬁnite or countably inﬁnite vocabulary, so it contains only countably many formulas; from which it follows that only countably many things are L-deﬁnable. But there are uncountably many ordinal numbers; indeed, uncountably many countableordinalnumbers.So thereare (countable)ordinalnumbers that are not L -deﬁnable. So there is a smallest ordinal number that is not L-deﬁnable, and it must be unique. But then ‘v is an ordinal number that is not L-deﬁnable but for which all smaller ordinals are L-deﬁnable’ is uniquely satisﬁed by this ordinal, so it is L-deﬁnable after all, which is a contradiction. That is the paradoxical line of argument. Any solution to the paradoxes of satisfaction will implicitly contain a solution to this deﬁnability paradox. On classical logic solutions, if the language L contains the predicate ‘satisﬁes’ then certain instances of schema (S) from Section 1 are refutable; and in particular, if we deﬁne ‘L-deﬁnable’ from ‘satisﬁes’ as above, there will be counterinstances to even the more restricted schema (Sdefn ) For all x, x satisﬁes ‘v is L-deﬁnable’ if and only if x is L-deﬁnable. This gives one possible diagnosis of the error in the argument: that it lies in the inference from σ being the uniquely smallest L-undeﬁnable ordinal to ‘v is the smallest L-undeﬁnable ordinal’ being uniquely satisﬁed by σ . But on the approach that I’ve sketched, we are committed to maintaining all instances of (S), and in particular all instances of (Sdefn ). Where then does the reasoning of the paradox go wrong? Where the reasoning goes wrong, I think, is that it makes an implicit application of excluded middle to a formula involving ‘L-deﬁnable’. Excluded middle can be assumed for certain restricted deﬁnability predicates. For instance, let L0 be obtained from L by deleting ‘satisﬁes’ and terms deﬁned from it (such as ‘deﬁnable’) or closely related to it; then excluded middle holds for formulas that contain ‘L0 -deﬁnable’ (as long as they don’t contain problematic terms in addition). There is no even prima facie problem about the least ordinal that is not L0 -deﬁnable, since the description of it just given is in a part of L that goes beyond L0 . Similarly for expansions of L0 in which the application of ‘satisﬁes’ is somehow restricted in a way that guarantees excluded

92 / Hartry Field middle (e.g. a language L1 in which ‘x satisﬁes y’ can occur only in the context ‘x satisﬁes y and y is an L0 -formula’, or a language L2 in which ‘x satisﬁes y’ can occur only in the context ‘x satisﬁes y and y is an L1 -formula’). Given a well-deﬁned hierarchy of such expansions, each of which includes all the vocabulary of the previous, one gets within L a hierarchy of restricted deﬁnability predicates, each more inclusive than the previous. But for deﬁnability in the full language L, the fact that excluded middle must be rejected for ‘satisﬁes’ suggests that it will almost certainly have to be rejected for the predicate ‘L-deﬁnable’ deﬁned from it; and the paradox shows that indeed it does. The implicit application of excluded middle to a formula involving ‘L-deﬁnable’ occurred in the step from (1)

There are ordinal numbers that are not L-deﬁnable

(2)

There is a smallest ordinal number that is not L-deﬁnable.

to To see that the inference from (1) to (2) depends on excluded middle, consider any speciﬁc ordinal β, and suppose that every ordinal less than β is L-deﬁnable. Given this supposition, (2) says in effect (3)

Either β is not L-deﬁnable, or there is an ordinal α>β such that α is not L-deﬁnable and all its predecessors are L-deﬁnable.

But this entails (4)

β is not L-deﬁnable or β is L-deﬁnable;

and so if we reject (4) we must reject (2).¹⁵ But there is no call to reject (1): there are certainly ordinals that are not L-deﬁnable, for uncountable ones can’t be L-deﬁnable (and there may well be sufﬁciently large countable ones which are deﬁnitely not L-deﬁnable too). So the inference from (1) to (2) relies on excluded middle. This resolution of the paradox may seem to have a high cost. For the inference from ‘There are ordinals α such that F(α)’ to ‘There is a smallest ordinal α such that F(α)’ is absolutely fundamental to ordinary set-theoretic reasoning; doesn’t what I’m saying count as a huge and crippling restriction on ordinary set theory? Not at all: ordinary set theory allows sets to be deﬁned only by ‘effectively classical’ properties, that is, properties F for which the generalized law of excluded middle ∀x[F(x) ∨ ¬F(x)] holds. I’m not suggesting any restriction whatever on the ordinary laws of set theory; what I am saying, and what is independently quite obvious, is that one has to be very careful if one wants to extend set theory by allowing properties (or formulas) that aren’t known to be effectively classical into its axiom schemas. ¹⁵ Which isn’t to say that we should accept the negation of (2): that would require (an existential quantiﬁcation of) a negation of excluded middle, which would lead to contradiction.

Solving the Paradoxes, Escaping Revenge / 93 This point is worth elaboration. Standard set theory (ZFC) contains two axiom schemas (the schemas of Separation and Replacement). On a strict interpretation of the theory, the allowable instances of the schemas are just those instances in the language of set theory; however, the ‘impure’ set theory that most of us accept and employ is more extensive than this, it allows instances of the schemas in which physical vocabulary occurs (e.g. we take the separation schema to allow us to pass from the existence of a set of all non-sets to the existence of a set of all neutrinos). But when the law of excluded middle is not assumed to hold unrestrictedly, there is a question of just how far the extension should go. I think a suitable extension of the schema of separation to be the rule (Extended Separation) (∀x ∈ z)(Ax ∨ ¬Ax) ∃y∀x(x ∈ y ↔ x ∈ z ∧ A(x)) (allowing free parameters in the formula A(x)), where any vocabulary at all, including ‘true’, can appear in A(x). Requiring excluded middle as an assumption of separation seems reasonable. Otherwise, we would license sets for which membership in the set depends on whether the Liar sentence is true; given extensionality, this would lead at the very least to indeterminate identity claims between sets, and it isn’t at all clear that paradox could be avoided even allowing that. But Extended Separation as formulated above avoids such oddities, while allowing such sets as the set of true sentences in the ‘true’-free set-theoretic language; it seems to me as much of an extension of separation to the language containing ‘true’ as we ought to want. It is easy to see that if the formula F(x) is allowed to contain non-classical vocabulary, then Extended Separation (together with the fact that every non-empty set has a member of least rank) justiﬁes reasoning from ‘There is at least one ordinal α such that F(α) and such that for all ordinals β

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close