(to appear in Mind, 2002)

Truthfulness and Relevance

deirdre wilson and dan sperber

Abstract

This paper questions the widespread view that verbal communication is governed by a maxim, norm or convention of truthfulness which applies at the level of what is literally meant, or what is said. Pragmatic frameworks based on this view must explain the frequent occurrence and acceptability of loose and figurative uses of language. We argue against existing explanations of these phenomena and provide an alternative account, based on the assumption that verbal communication is governed not by expectations of truthfulness but by expectations of relevance, raised by literal, loose and figurative uses alike. Sample analyses are provided, and some consequences of this alternative account are explored. In particular, we argue that the notions of ‘literal meaning’ and ‘what is said’ play no useful theoretical role in the study of language use, and that the nature of explicit communication will have to be rethought.

1. Introduction

Here are a couple of apparent platitudes. As speakers, we expect what we say to be accepted as true. As hearers, we expect what is said to us to be true. If it were not for these expectations, if they were not often enough satisfied, there would be little point in communicating at all. David Lewis (who has proposed a convention of truthfulness) and Paul Grice (who has argued for maxims of truthfulness), among others, have explored some of the consequences of these apparent platitudes. We want to take a different line and argue that they are strictly speaking false. Of course hearers expect to be informed and not misled by what is communicated; but what is communicated is not the same as what is said. We will argue that language use is not governed by any convention or maxim of truthfulness in what is said. Whatever genuine facts such a convention or maxim was supposed to explain are better explained by assuming that communication is governed by a principle of relevance.

According to David Lewis (1975), there is a regularity (and a moral obligation) of truthfulness in linguistic behaviour. This is not a convention in Lewis's sense, since there is no alternative regularity which would be preferable as long as everyone conformed to it. However, for any language £ of a population P, Lewis argues that there is a convention of truthfulness and trust in £ (an alternative being a convention of truthfulness and trust in some other language £'):

My proposal is that the convention whereby a population P uses a language £ is a convention of truthfulness and trust in £. To be truthful in £ is to act in a certain way: to try never to utter any sentences of £ that are not true in £. Thus it is to avoid uttering any sentence of £ unless one believes it to be true in £. To be trusting in £ is to form beliefs in a certain way: to impute truthfulness in £ to others, and thus to tend to respond to another's utterance of any sentence of £ by coming to believe that the uttered sentence is true in £. (Lewis 1975, p. 167)

Lewis considers the objection that truthfulness might not be the only factor which needs to be taken into account, and replies as follows:

Objection: Communication cannot be explained by conventions of truthfulness alone. If I utter a sentence s of our language £, you—expecting me to be truthful in £—will conclude that I take s to be true in £. If you think I am well informed, you will also conclude that probably s is true in £. But you will draw other conclusions as well, based on your legitimate assumption that it is for some good reason that I chose to utter s rather than remain silent, and rather than utter any of the other sentences of £ that I also take to be true in £. I can communicate all sorts of misinformation by exploiting your beliefs about my conversational purposes, without ever being untruthful in £. Communication depends on principles of helpfulness and relevance as well as truthfulness.

Reply: All this does not conflict with anything I have said. We do conform to conversational regularities of helpfulness and relevance. But these regularities are not independent conventions of language; they result from our convention of truthfulness and trust in £ together with certain general facts—not dependent on any convention—about our conversational purposes and our beliefs about one another. Since they are by-products of a convention of truthfulness and trust, it is unnecessary to mention them separately in specifying the conditions under which a language is used by a population. (Lewis 1975, p. 185)

However, Lewis does not explain how regularities of relevance might be by-products of a convention of truthfulness. One of our aims will be to show that, on the contrary, expectations of truthfulness—to the extent that they exist—are a by-product of expectations of relevance.

Paul Grice (1967), in his William James Lectures, sketched a theory of utterance interpretation based on a Co-operative Principle and maxims of truthfulness, informativeness, relevance and clarity (Quality, Quantity, Relation and Manner). The Quality maxims went as follows:

(1) Grice's maxims of Quality

Supermaxim: Try to make your contribution one that is true.

(i) Do not say what you believe to be false. [maxim of truthfulness]

(ii) Do not say that for which you lack adequate evidence.

The supermaxim of Quality is concerned with the speaker's overall contribution (what is communicated, either explicitly or implicitly), while the first and second maxims of Quality relate only to what is said (the proposition explicitly expressed or asserted). Grice saw the first maxim of Quality, which we will call the maxim of truthfulness, as the most important of all the maxims. He says in the William James Lectures:

It is obvious that the observance of some of these maxims is a matter of less urgency than is the observance of others; a man who has expressed himself with undue prolixity would, in general, be open to milder comment than would a man who has said something he believes to be false. Indeed, it might be felt that the importance of at least the first maxim of Quality is such that it should not be included in a scheme of the kind I am constructing; other maxims come into operation only on the assumption that this maxim of Quality is satisfied. While this may be correct, so far as the generation of implicatures is concerned, it seems to play a role not totally different from the other maxims, and it will be convenient, for the present at least, to treat it as a member of the list of maxims. (Grice 1967, p. 27)

In his ‘Retrospective Epilogue’, written 20 years later, this view is apparently maintained:

The maxim of Quality, enjoining the provision of contributions which are genuine rather than spurious (truthful rather than mendacious), does not seem to be just one among a number of recipes for producing contributions; it seems rather to spell out the difference between something's being and (strictly speaking) failing to be, any kind of contribution at all. False information is not an inferior kind of information; it just is not information. (Grice 1989a, p. 371)

Notice, though, an interesting shift. While he talks of ‘the maxim of Quality,’ Grice's concern here is with the speaker's contribution as a whole; indeed, there is room for doubt about whether he had the first maxim of Quality or the supermaxim in mind. We believe that this is not a minor detail. One of our aims is to show that the function Grice attributes to the Quality maxims—ensuring the quality of the speaker's overall contribution—can be more effectively achieved in a framework with no maxim of truthfulness at all.

There is a range of apparent counterexamples to the claim that speakers try to tell the truth. These include lies, jokes, fictions, metaphors and ironies. Lewis and Grice are well aware of these cases, and discuss them in some detail. Grice (1967, p. 30), for instance, notes that his maxims may be violated, and lists several categories of violation, each with its characteristic effects. Lies are examples of covert violation, where the hearer is meant to assume that the maxim of truthfulness is still in force and that the speaker believes what she has said. Jokes and fictions might be seen as cases in which the maxim of truthfulness is overtly suspended (the speaker overtly opts out of it); the hearer is meant to notice that it is no longer operative, and is not expected to assume that the speaker believes what she has said. Metaphor, irony and other tropes represent a third category: they are overt violations (floutings) of the maxim of truthfulness, in which the hearer is meant to assume that the maxim of truthfulness is no longer operative, but that the supermaxim of Quality remains in force, so that some true proposition is still conveyed.

We will grant that a reasonable—if not optimal—treatment of lies, jokes and fictions might be developed along these lines. Tropes, and more generally loose uses of language (e.g. approximations, sense extensions), present a much more pressing challenge. After all, many, if not most, of our serious declarative utterances are not strictly and literally true, either because they are figurative, or simply because we express ourselves loosely.

An utterance can be said to have a literal meaning which is capable of being either true or false when the result of combining its linguistic sense with its reference is a proposition. We ourselves do not claim that all utterances have a literal meaning, and we will be arguing that even when a literal meaning is available, it is not automatically the preferred interpretation of an utterance. In fact, literalness plays no role in our account of language comprehension, and nor does the notion of what is said. By contrast, to those who argue that there is an expectation of truthfulness in what is said, literal meanings matter. For Grice, what is said (as distinct from what is implicated) is the literal meaning of an utterance. For Lewis, what is said is specifiable on the basis of the utterance’s literal meaning. Without such an appeal to literal meaning in the determination of what is said, the claim that there is a maxim or convention of truthfulness in what is said would be, if not vacuous, at least utterly vague.

2. The case of tropes

Lewis (1975) considers the case of tropes:

Objection: Suppose the members of a population are untruthful in their language £ more often than not, not because they lie, but because they go in heavily for irony, metaphor, hyperbole, and such. It is hard to deny that the language £ is used by such a population.

Reply: I claim that these people are truthful in their language £, though they are not literally truthful in £. To be literally truthful in £ is to be truthful in another language related to £, a language we can call literal-£. The relation between £ and literal-£ is as follows: a good way to describe £ is to start by specifying literal-£ and then to describe £ as obtained by certain systematic departures from literal-£. This two-stage specification of £ by way of literal-£ may turn out to be much simpler than any direct specification of £. (Lewis 1975, p. 183)

Lewis's reply rests on a widely-shared view which dates back to classical rhetoric. On this view,

(2) (a) Figurative and literal utterances differ not in the kind of meanings they have (thus, if literal meanings are truth-conditional, so are figurative meanings), but in the way these meanings are generated.

(b) The meanings of figurative utterances are generated by systematic departures from their literal meanings.

For example, consider (3) and (4), where (3) is a metaphor and (4) is intended as hyperbole[1]:

(3) The leaves danced in the breeze.

(4) You’re a genius.

Lewis might want to say that in literal-English, sentences (3) and (4) have just their literal meanings. In actual English, the language in which a convention of truthfulness and trust holds among English speakers, (3) and (4) are ambiguous. They have their literal meanings plus other, figurative meanings: thus, (3) has the metaphorical meaning in (5), and (4) the hyperbolical meaning in (6):

(5) The leaves moved in the breeze as if they were dancing.

(6) You’re very clever.

However, it is not as if any language £ (in the sense required by Lewis, where the sentences of £ can be assigned truth-conditional meanings) had ever actually been specified on the basis of a corresponding ‘literal-£’. So what is the justification for accepting something like (2a) and (2b)? How are figurative meanings derived from literal meanings, and under what conditions do the derivations take place? Lewis does not explain, and there are no generally accepted answers to these questions. We have argued (Sperber and Wilson 1986a,b; 1990; 1998a) that figurative interpretations are radically context-dependent, and that the context is not fixed independently of the utterance but constructed as an integral part of the comprehension process. If so, then the very idea of generating the sentences of a language £ on the basis of a corresponding ‘literal-£’ is misguided.

Grice is often seen as providing an explanation of how figurative interpretations are conveyed. Consider a situation where the speaker of (3) or (4) manifestly could not have intended to commit herself to the truth of the propositions literally expressed: it is common knowledge that she knows that leaves never dance, or that she does not regard the hearer as a genius. She is therefore overtly violating the maxim of truthfulness: in Grice's terms, she is flouting it. Flouting a maxim indicates a speaker's intention: the speaker intends the hearer to retrieve an implicature which brings the full interpretation of the utterance (i.e. what is said plus what is implicated) as close as possible to satisfying the Co-operative Principle and maxims. In the case of tropes, the required implicature is related to what is said in one of several possible ways, each characteristic of a different trope. With metaphor, the implicature is a simile based on what is said; with irony, it is the opposite of what is said; with hyperbole, it is a weaker proposition, and with understatement, a stronger one. Thus, Grice might analyse (3) as implicating (5) above, and (4) as implicating (6).

Note that this treatment of tropes does not differ radically from Lewis’s, or from the classical rhetorical account. Grice's approach, like Lewis's, is based on assumption (2a) and, more importantly, on assumption (2b) (that the meanings of figurative utterances are generated by systematic departures from their literal meanings). The only difference is that Lewis sees these departures as systematic enough to be analysed in code-like terms: the figurative meaning of a sentence is a genuine linguistic meaning specified in the grammar of £ by a derivation which takes the literal meaning of the sentence as input. The sentences of £ (unlike those of literal-£) are systematically ambiguous between literal and figurative senses. For Grice, by contrast, sentences have only literal meanings. Figurative meanings are not sentence meanings but utterance meanings, derived in a conversational context. However, the derivations proposed in Grice’s pragmatic approach to tropes are the same as those hinted at by Lewis in his linguistic approach, and neither differs seriously from the classical rhetorical account.

Grice's treatment of tropes leaves several questions unanswered, and we will argue that it is inconsistent with the rationale of his own enterprise. In particular, there is room for doubt about what he meant by the maxim of truthfulness, and the role it was intended to play in his framework. This doubt is created by two possible interpretations of his notion of saying. On the first interpretation, saying is merely expressing a proposition, without any necessary commitment to its truth. Understood in this way, the maxim of truthfulness means ‘Do not express propositions you believe to be false.’ The function of this maxim, and more generally of the Quality maxims, would be to account for the fact—to the extent that it is a fact—that a speaker actually commits herself to the truth of what she says. Tropes would then be explained by the claim that flouting the maxim triggers the recovery of an implicature in the standard Gricean way. However, there is a problem. In general, the recovery of implicatures is meant to restore the assumption that the maxims have been observed, or that their violation was justified in the circumstances (as when a speaker is justified by her ignorance in providing less information than required) (Grice 1989a, p. 370). In the case of tropes, the maxim of truthfulness is irretrievably violated, and the implicature provides no circumstantial justification whatsoever.

On the second, and stronger, interpretation, saying is not merely expressing a proposition but asserting it: that is, committing oneself to its truth. Understood in this way, the maxim of truthfulness means ‘Do not assert propositions you believe to be false.’ On this interpretation, saying already involves speaker commitment, and the function of the maxim of truthfulness, and more generally of the Quality maxims, would be to ensure that speakers do not make spurious commitments. This seems to fit with Grice's above remark that the function of the Quality maxim is to guarantee that contributions are genuine rather than spurious. However, understood in this way, it is hard to see why a maxim of truthfulness is needed at all. It seems to follow from the very notion of an assertion as a commitment to truth (perhaps together with a proper understanding of commitment) that your assertions should be truthful. In fact, the only pragmatic function of the maxim of truthfulness, on this interpretation, is to be violated in metaphor and irony, thus triggering the search for an implicature. Without it, Grice would have no account of figurative utterances at all.

Which notion of saying did Grice have in mind in proposing the maxim of truthfulness? There is evidence of some hesitation. On the one hand, he treats the tropes as ‘Examples in which the first maxim of Quality is flouted’ (Grice 1967, p. 34). On the other, he comments that in irony the speaker ‘has said or has made as if to say’ [our italics] something she does not believe, and that in metaphor what is communicated must be obviously related to what the speaker ‘has made as if to say’ (ibid. p. 34). If the speaker of metaphor or irony merely ‘makes as if to say’ something, then the stronger notion of saying must be in force; on the other hand, if the speaker of a trope merely ‘makes as if’ to say something, then surely the maxim of truthfulness is not violated. But if the maxim of truthfulness is not violated, how does Grice's analysis of metaphor and irony go through at all?

Elsewhere in his philosophy of language, where the notion of saying plays a central role, it was the stronger rather than the weaker notion that interested Grice. He says, for example, ‘I want to say that (1) “U (utterer) said that p” entails (2) “U did something x by which U meant that p”.’ (Grice 1967, p. 87). For Grice, what is meant is roughly co-extensive with what is intentionally communicated: that is, with the information put forward as true. On this interpretation, saying involves speaker commitment: that is, it means asserting. Among his commentators, Stephen Neale (1992) treats these broader considerations as decisive: ‘If U utters the sentence “Bill is an honest man” ironically, on Grice's account U will not have said that Bill is an honest man: U will have made as if to say that Bill is an honest man.’ (Neale 1992, section 2).

How can we reconcile these two claims: that metaphor and irony are deliberate violations of the maxim of truthfulness, and hence must say something, and that in metaphor and irony the speaker merely makes as if to say something? A possible answer would be to distinguish two phases in the utterance interpretation process. In the first, the utterance of a declarative sentence would provide prima facie evidence for the assumption that an assertion is being made. In the second, this assumption would be evaluated and accepted or rejected. In the case of metaphor and irony, this second phase would involve an argument of the following sort: if it is common ground that the utterer U doesn't believe p, then U cannot assert p; it is common ground that U doesn't believe p; hence, U hasn't asserted p. In this way, we get a consistent interpretation of the notion of saying, and we can see why Grice hesitates between ‘saying’ and ‘making as if to say.’

However, if this interpretation is correct, then a trope involves no real violation of the maxim of truthfulness at any stage: since the speaker was not saying p, she was not saying what she believed to be false. A flouting, so understood, is a mere appearance of violation. So why should it be necessary to retrieve an implicature in order to preserve the assumption that the maxims have been respected? The Gricean way to go (although Grice himself did not take this route) would be to argue that it is not the maxim of truthfulness but some other maxim that is being violated. Quite plausibly, the maxim of Relation (‘Be relevant’) is being violated, for how can you be relevant when you speak and say nothing? Surely the first maxim of Quantity (‘Make your contribution as informative as is required’) is being violated, for if nothing is said, no information is provided. The implicature should thus be seen as a way of providing a full interpretation of the utterance in which these maxims are respected.

The problem with this analysis of tropes (and with the alternative analysis on which floutings of the maxim of truthfulness are genuine violations) is that it leads to an interpretation of figurative utterances which irretrievably violates the Manner maxims. In classical rhetoric, a metaphor such as (3) or a hyperbole such as (4) is merely an indirect and decorative way of communicating the propositions in (5) or (6). This ornamental value might be seen as explaining the use of tropes, insofar as classical rhetoricians were interested in explanation at all. Quite sensibly, Grice does not appeal to ornamental value. His supermaxim of Manner is not ‘Be fancy’ but ‘Be perspicuous.’ He assumes, this time in accordance with classical rhetoric, that figurative meanings, like literal meanings, are fully propositional, and always paraphrasable by means of a literal utterance. Which raises the following question: isn't a direct and literal expression of what you mean always more perspicuous (and in particular less obscure and less ambiguous, cf. the first and second Manner maxims) than an indirect figurative expression? (Remember: you cannot appeal to the subtle extra effects of tropes, since they are not considered, let alone explained, within the Gricean framework.)

It would be presumptuous to attribute Grice's apparent hesitation between two senses of saying to a lack of conceptual rigour on his part. We see it rather as arising from the difficulty of deploying a notion of saying which is both close enough to common usage to justify the choice of this word, and yet precise enough to make a contribution to a theory of language use. We will argue in §7 that there is no such notion. We ourselves do not use ‘saying’ as a theoretical term, except in rendering the views of others.

3. The case of loose use

Tropes are the most striking examples of serious utterances where the speaker is manifestly not telling the strict and literal truth. Even more common are loose uses of language (e.g. rough approximations, sense extensions), where an expression is applied to items that fall outside its linguistically-determined denotation, strictly understood. Consider the examples in (7)–(10):

(7) The lecture starts at five o’clock .

(8) Holland is flat.

(9) Sue: I must run to the bank before it closes.

(10) Jane: I have a terrible cold. I need a Kleenex.

If the italicised expressions in (7)–(10) are understood in the most restrictive way (and ignoring issues of ambiguity or polysemy for a moment), these utterances are not strictly and literally true: lectures rarely start at exactly the appointed time, Holland is not a plane surface, Sue must hurry to the bank but not necessarily run there, and other brands of disposable tissue would do just as well for Jane. Such loose uses of language are very common. Some are tied to a particular situation, produced once and then forgotten. Others may be regular and frequent enough to give rise to an extra sense, which may stabilise in an individual or a population: lexical broadening (along with lexical narrowing and metaphorical transfer) has been seen as one of the main pragmatic factors driving semantic change (Lyons 1977, Chs.13.4, 14.5). What concerns us here is not so much the outcome of these historical macro-processes as the nature of the pragmatic micro-processes that underlie them, and we will largely abstract away from the question of whether, or when, a word such as ‘flat’, or ‘run’, or ‘Kleenex’ may be said to have acquired an extra stable sense (see Sperber and Wilson 1998a for some discussion).

How should loose uses such as those in (7)–(10) be analysed? Are they lapses, the result of sloppy speech or thought, accepted by hearers whose expectations have been reduced to realistic levels by repeated encounters with normal human failings? Is it reasonable to assume that there really is a convention or maxim of truthfulness, although speakers quite commonly fall short of strictly obeying it? As hearers, would we always—and as speakers, should we always—prefer the strictly true statements in (11)–(14) to the loose uses in (7)–(10)?

(11) The lecture starts at or shortly after five.

(12) Holland has no mountain and very few hills.

(13) I must go to the bank as fast as if I were running.

(14) I need a Kleenex or other disposable tissue.

Clearly not. In most circumstances, the hearer would not be misled by strictly untrue approximations such as (7)–(10), and their strictly true counterparts in (11)–(14) would not provide him with any more valuable information. Indeed, since these strictly true counterparts are typically longer, the shorter approximations may be preferable.

Loose uses of language present few problems for speakers and hearers, who are rarely even aware of their occurrence; but they do raise a serious issue for any philosophy of language based on a maxim or convention of truthfulness. We have suggested above that appeals to ambiguity (or polysemy) merely defer the problem, since such ambiguities ultimately derive from repeated instances of loose use (for further discussion, see §6 below). In this section, we will consider other solutions proposed in the literature, paying particular attention to Lewis’s treatment of pragmatic vagueness. We will argue that no single solution, nor any combination of proposed solutions, is adequate to handle the full variety of loose uses of language, which go well beyond the types of pragmatic vagueness dealt with on Lewis’s account.

In Grice’s framework, loose uses such as (7)–(10) apparently violate either the maxim of truthfulness or the second maxim of Quality (‘Have adequate evidence for what you say’). However, they do not really fit into any of the categories of violation listed in §1 above. They are not covert violations, designed to deceive the hearer into believing the proposition strictly and literally expressed. They are not like jokes or fictions, which suspend the maxims entirely. One might try to analyse them as floutings: overt violations (real or apparent), designed to trigger the search for a related implicature (here a hedged version of what was literally said or quasi-said); but the problem is that loose uses are not generally perceived as violating the Quality maxims at all. In classical rhetoric, they were not treated as tropes involving the substitution of a figurative for a literal meaning. They do not have the striking quality that Grice associated with floutings, which he saw as resulting in figurative or quasi-figurative interpretations. Loose talk involves no overt violation, real or apparent; or at least it does not involve a degree of overtness in real or apparent violation which might trigger the search for an implicature. While we are all capable of realising on reflection that utterances such as (7)–(10) are not strictly and literally true, these departures from truthfulness pass unattended and undetected in the normal flow of discourse. Grice's framework thus leaves them unexplained.

Perhaps we should reconsider the apparent platitudes we started with. Maybe we should have said that as speakers, we expect what we say to be accepted as approximately true, and as hearers, we expect what is said to us to be approximately true. But this is far too vague to do the required work of explaining how speakers and hearers manage to communicate successfully. Approximations differ both in kind and in degree, and their acceptability varies with content and context. There is no single scale on which the degrees of approximation in disparate statements such as (7)–(10) can be usefully compared. The same statement can be an acceptable approximation in one situation and not in another. Thus, suppose the speaker of (7) expects the lecture to start sometime between five o’clock and 5.10: then (7) would be an acceptable approximation to a student who has just asked whether the lecture starts at five or six o’clock , but not to a radio technician preparing to broadcast the lecture live. Moreover, as we will argue below, there are cases where the notion of ‘degrees of approximation’ does not really apply.

A convention of truthfulness and trust in a language (if there were one) might play a valuable role in explaining how linguistic expressions acquire their conventional meanings, and how speakers and hearers use these meanings to communicate successfully. If all that speakers and hearers are entitled to are vague expectations of approximate truth, it is hard to see how the resulting convention of approximate truthfulness could be robust enough to establish common meanings. As we have shown, the same approximation may be differently intended and understood in different circumstances. Unless it is supplemented by some account of how speakers and hearers may converge on these more specific understandings—an account which might then be doing most of the explanatory work—a convention of approximate truthfulness and trust is inadequate to explain how the co-ordination necessary for successful communication is achieved. Still, this is the direction that David Lewis proposes to explore. He writes:

When is a sentence true enough? […] this itself is a vague matter. More important for our present purposes, it is something that depends on context. What is true enough on one occasion is not true enough on another. The standards of precision in force are different from one conversation to another, and may change in the course of a single conversation. Austin 's ‘ France is hexagonal’ is a good example of a sentence that is true enough for many contexts but not true enough for many others. (1979, pp. 244–45)

We agree with Lewis (and Unger 1975, Ch.2) that ‘hexagonal’ and ‘flat’ are semantically absolute terms, and that their vagueness should be seen as pragmatic rather than semantic (so in our terms they are genuine cases of loose use). However, Lewis’s analysis of pragmatically vague terms such as ‘flat’ is very similar to his analysis of semantically vague terms such as ‘cool’.[2] For Lewis, a semantically vague term has a range of possible sharp delineations, marking different cut-off points between, say, ‘cool’ and ‘warm’. ‘This is cool’ may be true at some but not all delineations, and depending on our purposes, we may be willing or unwilling to assert it: hence its vagueness (1970, pp. 228–29). On Lewis’s account, semantically absolute but pragmatically vague terms are handled on similar lines, except that semantic delineations are replaced by contextually-determined standards of precision (so if ‘flat’ were semantically rather than pragmatically vague, the analysis would not be very different). On this approach, ‘ Holland is flat’ would be true according to some fairly low standard of precision, but false given higher standards.

Semantic vagueness clearly exists (‘bluish’ and ‘flattish’ are good examples); its analysis raises problems of its own, about which we have nothing to say here (see Williamson 1994; Keefe and Smith 1996). What we do want to argue against is the idea that loose use can be successfully treated as a pragmatic analogue of semantic vagueness. As we have suggested above (and will argue in more detail below), there are many varieties of loose use, not all of which can be satisfactorily handled by appeal to contextually-determined standards of precision. For the cases that cannot be handled on Lewis’s lines, an alternative analysis must be found. We will propose such an analysis, and argue that it generalises straightforwardly to all varieties of loose use (and indeed to all utterances, literal, loose, or figurative), making the appeal to standards of precision as a component of conversational competence unnecessary.

In fact, there are problems even in some cases where the appeal to contextually-determined standards of precision looks initially plausible. Consider a situation where (7) (‘The lecture starts at five o’clock ’) would be accepted as true enough if the lecture started somewhere between five o’clock and 5.10. On Lewis’s account, it might be claimed that here the contextually-determined standard of precision allows for a give or take of, say, fifteen minutes around the stated time. It should then follow that a hearer in the same situation, with the same standard of precision in force, would be equally willing to accept (7) as true enough if the lecture started somewhere between 4.50 and five o’clock . But there is an obvious asymmetry between the two cases. Intuitively, the reason is clear enough: the audience won't mind or feel misled if they get to the lecture a few minutes early, but they will if they get there a few minutes late, so the loosening is acceptable only in one direction. In a different situation—when the speaker is talking about the end of the lecture rather than the beginning, for example—there may be an asymmetry in the other direction. Again, the reason is intuitively clear: the audience won't mind or feel misled if they can get away a bit earlier than expected, but they will if they have to stay longer. It is hard to explain these obvious intuitions by appeal to the regular notion of contextually-determined standards of precision as described above. One might, of course, try building the asymmetries into the standards of precision themselves, but then two different standards would have to be invoked to explain how (15) is quite naturally understood to mean something like (16):

(15) The lecture starts at five o’clock and ends at seven o’clock .

(16) The lecture starts at five o’clock or shortly after and ends at seven o’clock or shortly before.

This is clearly ad hoc. It would be better to find an alternative account of these asymmetries—but such an account might make the appeal to contextually-determined standards of precision redundant.

A more serious problem for Lewis is that in some cases of loose use, the appeal to contextually-determined standards of precision does not seem to work at all. Lewis’s account works best when there is a continuum (or ordered series) of cases between the strict truth and the broadest possible approximation. ‘Flat’ is a good example, since departures from strict flatness may vary in degree. ‘ Five o’clock ’ also works well in this respect, since departures from exactness may vary in degree. But with ‘run’ in (9) (‘I must run to the bank’) and ‘Kleenex’ in (10) (‘I need a Kleenex’), no such continuum exists. There is a sharp discontinuity between running (where both feet leave the ground at each step) and walking (where there is always at least one foot on the ground). Typically (though not necessarily), running is faster than walking, so that ‘run’ may be loosely used, as in (9), to indicate the activity of going on foot (whether walking or running) at a speed more typical of running. But walking at different speeds is not equivalent to running relative to different standards of precision. Similarly, ‘Kleenex’ is a brand name: other brands of disposable tissue are not Kleenex. The word ‘Kleenex’ may be loosely used, as in (10), to indicate a range of tissues similar to Kleenex. But there is no continuum on which being similar enough to Kleenex amounts to actually being Kleenex relative to stricter or looser standards of precision. ‘Run’, ‘Kleenex’ and many other words have sharp conceptual boundaries and no ordered series of successively broader extensions which might be picked out by raising or lowering some standard of precision. Yet these terms are often loosely used. This supports our claim that looseness is a broader notion than pragmatic vagueness.[3]

Again, for someone with no particular theoretical axe to grind, it is easy enough to see intuitively what is going on. Suppose you have a lecture one afternoon, but don’t know exactly when it is due to start. Someone tells you, ‘The lecture starts at five o’clock .’ From the literal content of the utterance, together with other premises drawn from background knowledge, you can derive a number of conclusions that matter to you: that you will not be free to do other things between five and seven o’clock, that you should leave the library no later than 4.45, that it will be too late to go shopping after the lecture, and so on. To say that these conclusions matter to you is to say that you can use them to derive still further non-trivial contextual implications, of a practical or a theoretical nature. These initial conclusions are the main branches of a derivational tree with many further branches and sub-branches. You would have been able to derive all these direct and indirect conclusions from the strictly true utterance ‘The lecture starts at or shortly after five o’clock ,’ but at the extra cost required to process a longer sentence and a more complex meaning. There are other conclusions¾false ones this time¾that you would have been able to derive from the approximation, ‘The lecture starts at five o’clock ’, but not from its strictly true counterpart: that the lecture will have begun by 5.01, for instance. But you are unlikely to derive them. They don't matter, because they are derivational bare branches which yield no further non-trivial implications.

Suppose that Peter and Mary, who are both rather unfit, are discussing where to go on their next cycling holiday. Mary suggests Holland , adding, ‘ Holland is flat.’ From the strictly false proposition that Holland is flat—just as easily as from the strictly true hedged proposition that Holland is approximately flat—Peter can derive the true conclusion that cycling in Holland would involve no mountain roads and would not be too demanding. Unlike the true hedged proposition, the false approximation also has false implications (that there are no hills at all in Holland , for instance). But it is unlikely that Peter would even contemplate deriving any of these.

Suppose that Sue, chatting with friends in the street, looks at her watch and says, ‘I must run to the bank before it closes.’ Her friends will take her to mean that she must break off their chat and hurry to the bank. For them, that much information is worth deriving. Whether she will actually get to the bank by running, walking fast or a mixture of both is of no interest to them, and they will simply not attend to this aspect of the literal meaning of her utterance.

Suppose that Jane and Jack are in the cinema waiting for the film to start. By saying, ‘I have a terrible cold. I need a Kleenex,’ Jane provides a premise from which Jack can infer that she wants to borrow a tissue to use in dealing with her cold. Her utterance also provides a premise from which he could derive the possibly false conclusion that she does not want a tissue of any other brand than Kleenex; but he is unlikely to draw such a conclusion, since his expectations of relevance in this context are satisfied by the weaker interpretation on which she wants a tissue (and in a context where his expectations of relevance would encourage an interpretation on which Jane was specifically requesting a Kleenex—e.g. if she was angrily throwing away tissues of another brand—the utterance would not be understood as a case of loose talk).

As these examples show, hearers have no objection to strictly false approximations as long as the conclusions they bother to derive from them are true. In fact, they might prefer the shorter approximations to their longer-winded but strictly true counterparts for reasons of economy of effort. There is some evidence that speakers take account of the perceived preferences of their audience in deciding how strictly or loosely to speak. In a series of experiments on truthfulness and relevance in telling the time, Van der Henst, Carles and Sperber (forthcoming) showed that when people in public places are asked the time by a stranger, they tend to respond with a time that is either accurate to the minute or rounded to the nearest multiple of five, depending on how useful in the circumstances they think a more accurate answer would be. This includes people with digital watches, from which a time that is accurate to the minute is most easily read off. These people have to make an extra effort to produce a rounded answer which, to the best of their knowledge, is not strictly true, but is easier for their audience to process.

Anticipating the arguments of the next section, let us say that an utterance is relevant when the hearer, given his cognitive dispositions and the context, is likely to derive some genuine knowledge from it (we will shortly elaborate on this). Someone interested in defending a maxim or convention of truthfulness might then suggest that expectations of relevance do play a role in comprehension, but in a strictly limited way. It might be claimed, for example, that while utterances in general create expectations of truthfulness, approximations alone create expectations of relevance, which have a role to play in the case of loose talk, but only there. This account (apart from being unparsimonious) raises the following problem. As noted above, while we are all capable of realising on reflection that an utterance was an approximation rather than a strictly literal truth, the fact that an approximation has been used is simply not noticed in the normal flow of discourse, and is surely not recognisable in advance of the comprehension process. But in that case, how could loose talk and literal talk be approached and processed with different expectations?

Here is the answer. It is not just approximations but all utterances¾literal, loose or figurative¾that are approached with expectations of relevance rather than truthfulness. Sometimes, the only way of satisfying these expectations is to understand the utterance as literally true. But just as an utterance can be understood as an approximation without being recognised and categorised as such, so it can be literally understood without being recognised and categorised as such. We will argue that the same is true of tropes. Literal, loose, and figurative interpretations are arrived at in the same way, by constructing an interpretation which satisfies the hearer’s expectations of relevance (for earlier arguments along these lines, see Sperber and Wilson 1986a,b; 1990; 1998a).

No special machinery is needed to explain the interpretation of loose talk. In particular, contextually-determined standards of precision play no role in the interpretation process. They do not help with cases such as ‘run’, or ‘Kleenex’, which are neither semantically nor pragmatically vague; and to appeal to them in analysing cases such as ‘flat’ or ‘at five o’clock’, which might be seen as involving a pragmatic form of vagueness, would be superfluous at best.

4. Relevance: theory

Grice's maxim of truthfulness was part of what might be called an inferential model of human communication. This contrasts with a more classical code model, which treats utterances as signals encoding the messages that speakers intend to convey. On the classical view, comprehension is achieved by decoding signals to obtain the associated messages. On the inferential view, utterances are not signals but pieces of evidence about the speaker's meaning, and comprehension is achieved by inferring this meaning from the evidence provided. An utterance is, of course, a linguistically coded piece of evidence, so that the comprehension process will involve an element of decoding. But the linguistically-encoded sentence meaning need not be identical to the speaker's meaning¾and we would argue that it never is¾since it is likely to be ambiguous and incomplete in ways the speaker's meaning is not. On this approach, the linguistic meaning recovered by decoding is just one of the inputs to an inferential process which yields an interpretation of the speaker's meaning.

Grice, Lewis and others who have contributed to the development of an inferential approach to communication have tended to minimise the gap between sentence meaning and speaker's meaning; they treat the inference from sentence meaning to speaker’s meaning as merely a matter of assigning referents to referring expressions, and perhaps of calculating implicatures. While the slack between sentence meaning and speaker's explicit meaning cannot be entirely eliminated, a framework with a maxim or convention of truthfulness has the effect of reducing it to a minimum. But why should this be something to be a priori expected or desired? Comprehension is a complex cognitive process. From a cognitive point of view, how much of the work is done by inference and how much by decoding depends on how efficient the inferential processes are. We have argued (Wilson and Sperber 1981; 1993; forthcoming; Sperber and Wilson 1986a; 1998a) that relevance-oriented inferential processes are efficient enough to allow for a much greater slack between sentence meaning and speaker's meaning than is generally assumed. Here, we summarise the theory briefly for purposes of the present discussion.

We characterise relevance as a property of inputs to cognitive processes which makes them worth processing. (‘Relevance’ is used in a technical sense which is not meant to capture any of the ordinary senses of the word.) These inputs may be external stimuli (e.g. a smell, the sound of an utterance), or internal representations which may undergo further processing (e.g. the recognition of a smell, a memory, the linguistic decoding of an utterance). At each point in our cognitive lives, there are many more potential inputs available than we can actually process: for example, we perceive many more distal stimuli than we can attend to, and have many more memories than we can reactivate at a single time. Efficiency in cognition is largely a matter of allocating our processing resources so as to maximise cognitive benefits. This involves processing inputs that offer the best expected cost/benefit ratio at the time.

Here we will consider only one type of cognitive benefit: improvements in knowledge (theoretical or practical). This is plausibly the most important type of cognitive benefit. There may be others: improvements in memory or imagination, for example (although it might be argued that these are benefits only because they contribute indirectly to improvements in knowledge; better memory and imagination lead to better non-demonstrative inference, and therefore to better knowledge). In any case, for our present purposes, there is another important reason for identifying cognitive benefits with improvements in knowledge.

In a situation where it is clear to both participants that the hearer's goal in listening to the speaker's utterances is not the improvement of knowledge—say, he just wants to be amused—there is no reason why the speaker should be expected to tell the truth. Thus, one way of challenging the maxim or convention of truthfulness would be to start by questioning whether humans are much interested in truth (e.g. Stich 1990). Here, we want to present a more pointed challenge to Grice’s and Lewis’s ideas, based on the nature of human communication rather than the goals of cognition. We will therefore grant that one of the goals of most human communication (though certainly not the only one) is the transmission of genuine information and the improvement of the hearer's knowledge. We will consider only cases where hearers are interested in truth. Our claim is that even in these cases, hearers do not expect utterances to be literally true.

The processing of an input in the context of existing assumptions may improve the individual’s knowledge not only by adding a new piece of information, but by revising his existing assumptions, or yielding conclusions not derivable from the new piece of knowledge alone or from existing assumptions alone. We define an input as relevant when and only when it has such positive cognitive effects.[4] Relevance is also a matter of degree, and we want to characterise it not only as a classificatory notion but also as a comparative one. There are inputs with at least some low degree of relevance all around us, but we cannot attend to them all. What makes an input worth attending to is not just that it is relevant, but that it is more relevant than any alternative potential input to the same processing resources at that time. Although relevance cannot be measured in absolute terms, the relevance of various inputs can be compared.

For our purpose, which is to characterise a property crucial to cognitive economy, the relevance of inputs must be comparable not only in terms of benefits (i.e. positive cognitive effects), but also in terms of costs (i.e. processing effort). We therefore propose the following comparative notion of relevance:

(17) Relevance of an input to an individual at a time

(a) Everything else being equal, the greater the positive cognitive effects achieved in an individual by processing an input at a given time, the greater the relevance of the input to that individual at that time.

(b) Everything else being equal, the smaller the processing effort expended by the individual in achieving those effects, the greater the relevance of the input to that individual at that time.

Here is a brief and artificial illustration. Peter wakes up feeling unwell and goes to the doctor. On the basis of her examination, she might make any of the following true statements:

(18) You are ill.

(19) You have flu.

(20) You have flu or 29 is the square root of 843.

The literal content of all three utterances would be relevant to Peter. However, (19) would be more relevant than either (18) or (20). It would be more relevant than (18) for reasons of cognitive effect, since it yields all the consequences derivable from (18) and more besides. This is an application of clause (a) of the characterisation of relevance in (17). It would be more relevant than (20) for reasons of processing effort: although (19) and (20) yield exactly the same consequences, these consequences are easier to derive from (19) than from (20), which requires an additional effort of parsing and inference (in order to realise that the second disjunct is false and the first is therefore true). This is an application of clause (b) of the characterisation of relevance in (17).

Given this characterisation of relevance, it is, ceteris paribus, in the individual’s interest to process the most relevant inputs available. We claim that this is what people tend to do (with many failures, of course). They tend to do it not because they realise that it is in their interest (and they certainly do not realise it in those terms), but because they are cognitively-endowed evolved organisms. In biological evolution, there has been constant pressure on the human cognitive system to organise itself so as to select inputs on the basis of their expected relevance (see Sperber and Wilson 2002). Hence:

(21) The First, or Cognitive, Principle of Relevance

The human cognitive system tends towards processing the most relevant inputs available.

The tendency described in the Cognitive Principle of Relevance is strong enough, and manifest enough, to make our mental processes at least partially predictable to others. We are in general fairly good at predicting which of the external stimuli currently affecting some other individual’s nervous system she is likely to be attending to, and which of the indefinitely many conclusions that she might draw from it she will in fact draw. What we do, essentially, is assume that she will pay attention to the potentially most relevant stimulus, and process it so as to maximise its relevance: that is, in a context of easily accessible background assumptions, where the information it provides will carry relatively rich cognitive effects.

This mutual predictability is exploited in communication. As communicators, we provide stimuli which are likely to strike our intended audience as relevant enough to be worth processing, and to be interpreted in the intended way. A communicator produces a stimulus¾say an utterance¾which attracts her audience's attention, and she does so in an overtly intentional way. In other words, she makes it manifest that she wants her audience's attention. Since it is also manifest that the audience will tend to pay appropriate attention only to an utterance that seems relevant enough, it is manifest that the communicator wants her audience to assume that the utterance is indeed relevant enough. There is thus a minimal level of relevance that the audience is encouraged to expect: the utterance should be relevant enough to be worth the effort needed for comprehension.

Is the audience entitled to expect more relevance than this? In certain conditions, yes. The communicator wants to be understood. An utterance is most likely to be understood when it simplifies the hearer's task by demanding as little effort from him as possible, and encourages him to pay it due attention by offering him as much effect as possible. The smaller the effort, and the greater the effect, the greater the relevance. It is therefore manifestly in the communicator's interest for the hearer to presume that the utterance is not just relevant enough to be worth his attention, but more relevant than this. How much more? Here, the communicator is manifestly limited by her own abilities (to provide appropriate information, and to present it in the most efficient way). Nor can she be expected to go against her own preferences (e.g. against the goal she wants to achieve in communicating, or the rules of etiquette she wishes to follow). Still, it may be compatible with the communicator’s abilities and preferences to go beyond the minimally necessary level of relevance. We define a notion of optimal relevance (of an utterance, to an audience) which takes these ideas into account, and propose a second principle of relevance based on it:

(22) Optimal relevance of an utterance

An utterance is optimally relevant to the hearer iff:

(a) It is relevant enough to be worth the hearer's processing effort;

(b) It is the most relevant one compatible with the speaker's abilities and preferences.

(23) The Second, or Communicative, Principle of Relevance

Every utterance conveys a presumption of its own optimal relevance.

In interpreting an utterance, the hearer invariably has to go beyond the linguistically-encoded sentence meaning. There will be ambiguities and referential indeterminacies to resolve, and other underdeterminate aspects of explicit content that we will look at shortly. There may be implicatures to identify, illocutionary indeterminacies to resolve, metaphors and ironies to interpret. All this requires an appropriate set of contextual assumptions. The Communicative Principle of Relevance and the definition of optimal relevance suggest a practical procedure for constructing a hypothesis about the speaker’s meaning. The hearer should consider interpretive hypotheses (disambiguations, reference assignments, implicatures, etc.) in order of accessibility—that is, follow a path of least effort—and stop when he arrives at an interpretation which satisfies the expectations of relevance raised by the utterance itself.[5]

What makes it reasonable for the hearer to follow a path of least effort is that the speaker is expected (within the limits of her abilities and preferences) to make her utterance as easy as possible for the hearer to understand. Since relevance varies inversely with effort, the very fact that an interpretive hypothesis is easily accessible gives it an initial degree of plausibility (an epistemic advantage specific to communicated information).

What makes it reasonable for the hearer to stop at the first interpretation which satisfies his expectations of relevance is that either this interpretation is close enough to what the speaker meant, or she has failed to communicate her meaning. A speaker who produced an utterance with two or more significantly different interpretations, each yielding the expected level of cognitive effect, would put the hearer to the gratuitous and unexpected extra effort of choosing among them, and the resulting interpretation (if any) would not satisfy clause (b) of the presumption of optimal relevance.[6] Thus, when a hearer following the path of least effort finds an interpretation which satisfies his expectations of relevance, in the absence of contrary evidence, this is the best possible interpretive hypothesis. Since comprehension is a non-demonstrative inference process, this hypothesis may well be false. Typically, this happens when the speaker expresses herself in a way that is inconsistent with the expectations she herself has raised, so that the normal inferential routines of comprehension fail. Failures in communication are common enough: what is remarkable and calls for explanation is that communication works at all.

This relevance-theoretic account not only describes a psychological process but also explains what makes this process genuinely inferential: that is, likely to yield true conclusions (in this case, intended interpretations) from true premises (in this case, from the fact that the speaker has produced a given utterance, together with contextual information). On the descriptive level, it has testable implications, some of which have been tested and confirmed by a growing body of research in experimental pragmatics.[7] On the explanatory level, it claims that what makes this relevance-guided process genuinely inferential is the fact that it typically yields a single interpretation for a given quadruple of speaker, hearer, utterance and situation. Given that such interpretations are predictable by the speaker, the best utterance for a speaker to produce is the one that is likely to be interpreted in the intended way, and the best interpretation for a hearer to choose is the one arrived at by use of the relevance-guided procedure, which is therefore likely to have been predicted and intended by the speaker. Communication is a form of co-ordination, and runs into co-ordination problems which are partly standard, and partly specific to communication. Relevance-guided comprehension takes advantage of the communication-specific aspects of these problems, and provides a solution which is, of course, imperfect, but is nonetheless effective (for further discussion, see Sperber and Wilson 2002).

5. Relevance: illustration

An utterance has two immediate effects: it indicates that the speaker has something to communicate, and it determines an order of accessibility in which interpretive hypotheses will occur to the hearer. Here is an illustration.

Lisa drops by her neighbours, the Joneses, who have just sat down to supper:

(24) Alan Jones: Do you want to join us for supper?

Lisa: No, thanks. I've eaten.

A standard semantic analysis of the second part of Lisa’s utterance would assign it the following truth condition:

(25) At some point in a time span whose endpoint is the time of utterance, Lisa has eaten something.

Clearly, though, Lisa means something more specific than this. She means that she has eaten that very evening, and not just anything, but a supper or something equivalent: a few peanuts wouldn't do.[8]

Here is our explanation of how Alan understands Lisa’s meaning. Her utterance activates in his mind, via automatic linguistic decoding, a conceptual structure which articulates in the grammatically specified way the concepts of Lisa, of eating, and of a time span whose endpoint is the time of utterance.[9] He does not have to reason, because it is all routine, but he might reason along the following lines: she has caused me a certain amount of processing effort (the effort required to attend to her utterance and decode it). Given the Communicative Principle of Relevance, this effort was presumably not caused in vain. So the conceptual structure activated by her utterance must be a good starting point for inferring her meaning, which should be relevant to me.

Lisa’s utterance, ‘I have eaten’, immediately follows her refusal of Alan’s invitation to supper. It would be relevant to Alan (or so she may have thought) to know the reasons for her refusal, which have implications for their relationship: Did she object to the offer? Would she accept it another time? It all depends on the reasons for her refusal. The use of the perfect ‘have eaten’ indicates a time span ending at the time of utterance and starting at some indefinite point in the past. Alan narrows the time span by assuming that it started recently enough for the information that Lisa has eaten during that period to yield adequate consequences: here, the relevant time span is that very evening (for discussion, see Wilson and Sperber 1998). He does the same in deciding what she ate. In the circumstances, the idea of eating is most easily fleshed out as eating supper, and this, together with the narrowing of the time span, yields the expected level of cognitive effect. Alan then assumes that Lisa intended to express the proposition that she has eaten supper that evening, and to present this as her reason for refusing his invitation. Although this attribution of meaning is typically a conscious event, Alan is never aware of the process by which he arrived at it, or of a literal meaning equivalent to (25).

The process by which Alan interprets Lisa’s meaning may be represented as in Table 1, with Alan’s interpretive hypotheses on the left, and his basis for arriving at them on the right. We have presented the hypotheses in English, but for Alan they would be in whatever is the medium of conceptual thought, and they need not correspond very closely to our paraphrases.