CNRS, Paris

Just as bats are unique in their ability to use echolocation, so are humans unique in their ability to use metarepresentations. Other primates may have some rather rudimentary metarepresentational capacities. We humans are massive users of metarepresentations, and of quite complex ones at that. We have no difficulty, for instance, in processing a three-tiered metarepresentation such as that expressed in (1):

The fact that humans are expert users of metarepresentations, is, I would argue, as important in understanding human behavior, as the fact that bats are expert users of echolocation is in understanding bat behavior.

How has the human metarepresentational capacity evolved? In order to contribute to the ongoing debate on this question, I will focus on three more specific issues:

1. How do humans metarepresent representations?
2. What came first: language, or metarepresentations?
3. Do humans have more than one metarepresentational ability?

Metarepresentations are representations of representations, but not all representations of representations are metarepresentations in the relevant sense. The human metarepresentational capacity we are interested in here is, first and foremost, a capacity to represent the content of representations.

Consider:

Statements (2a) and (2b) are about representations but they do not in any way represent their contents. They are not metarepresentations in any useful sense. One can imagine a biological or artificial device that could detect the presence of representations, but not at all their content properties. It could detect, say, mental representations by being sensitive to the appropriate manifestations of brain activity, or it could detect public representations such as utterances by being sensitive to their phonetic properties. Such a device, lacking access to the content properties of the representations it would represent, would not have a metarepresentational capacity in the sense intended.

Consider now:

Statements (3a-d) do attribute a more or less specific tenor to representations they are directly or indirectly about. Thus (3a) is true only if Mary tends to form unwarranted beliefs the gist of which is that she is seriously ill. (3b) is true only if Henry is making claims to the effect that certain things are not the way they ought to be. (3c) is true only if John accepts the central tenets of the creationist doctrine. (3d) is true only if the claim referred to attributes to some agent some undesirable property. Representations (3a-d) do state or imply something about the content of some representations, however they do not articulate the contents of these representations. They are metarepresentational only in a rudimentary way.

There may be animals capable of detecting the presence and tenor of some mental representations in others but who do so by means of unstructured representations. Thus an animal might detect the fact that a conspecific wants to mate, and represent it by means of a single unarticulated symbol wants-to-mate as in (4a). Similarly an animal might detect the fact that a conspecific has been suitably impressed with some display of strength, and represent it by means of a single unarticulated symbol knows-that-I-am-stronger-than-him as in (4b).

Such animals would possess a very rudimentary metarepresentational capacity, lacking compositionality and recursion. They could only metarepresent a short and fixed list of representations.

Imagine a species with quite limited cognitive capacities. The contents of its mental representations belong to a limited repertoire: "there is a predator (at a certain location)"; "there is a prey"; "there is a mating partner," and so forth. Suppose it were advantageous for members of this species to be capable of representing some of the mental states of their conspecifics. Then, a rudimentary metarepresentational ability of the kind I have just evoked might evolve.

Consider now a species with rich cognitive capacities, capable of forming and of using in its behavior mental representations of indefinitely varied contents. Suppose again that it were advantageous for members of this species to be capable of representing indefinitely may of the mental states of their conspecifics. A rudimentary metarepresentational ability of the kind just considered would not do this time. Such ability would be sufficient to represent only a narrow and fixed subset of the indefinitely varied mental representations of the members of the species.

Another way to put the same point is this. Consider first the case of a species with a system of internal representations consisting, roughly, in a small list of representations-types that can be occasionally tokened (and indexed to the situation). Then the internal representations of members of this species could be metarepresentated by means of a rudimentary metarepresentational system with a different mental symbol for each metarepresentable representation type. Consider now a species with system of internal representations that is - or is equivalent to - a mental language with compositionality and recursion. Then the internal representations of members of this species could be metarepresentated only by means of metarepresentational system consisting in - or equivalent to - a meta-language no less rich than the language the expressions of which it serves to metarepresent. If, moreover, several levels of metarepresentation are possible (as in (1) above), the metarepresentational system must equally rich or richer at each level. The only cost-effective way to achieve this is to have the expressions of the object-language do double service as expression of the meta-language (or, if n levels of metarepresentation are possible, do n+1-tuple service). A full-fledged metarepresentational capability such as the one found in human languages and in human thinking is based on the possibility of interpreting any expression-token as representing another token of the same expression, or the expression-type, or more generally some expression type or token it resembles in relevant respects. Thus, in (5a-e) the expressions italicized represent not the state of affairs that they describe, but representations (mental, public, or abstract) the contents of which they serve to render.

Examples (1) and (5a-e) are linguistic ones, but I take it that humans are capable of entertaining their mental equivalent. In other words, expressions in the internal system of conceptual representations (however close or distant from natural language this system might be) can serve to represent expression types or tokens which they resemble in relevant respects (and identity is best seen here as just a limiting case of resemblance - see Sperber & Wilson 1986/1995, chapter 4).

Imagine an organism endowed with a rich internal system of conceptual representations, but without the ability to use these "opaquely" or metarepresentationally, i.e. as iconic representations of other representations (types or tokens). Could such an organism learn to do so? Think of what is involved. The organism has no more knowledge of its internal representations qua representations than it has of its patterns of neural activities. Its representations are wholly transparent to it. Such an organism might be capable of representing, say, the fact that it is raining, but never of representing the fact that it is representing the fact that it is raining.

It is hard to imagine what combination of external stimulations and internal dispositions might ever cause individual organisms of this type to become capable of using their repertoire of mental representations in order to represent not what these representations transparently represent, but, in virtue of resemblance relationships, to represent opaquely other representations. This seems as implausible as learning to guide one's spontaneous movement by means of echolocation in the absence of a genetically determined domain-specific disposition. Such abilities speak of biological-evolutionary, rather than cognitive-developmental transitions. I am of course aware of the relative weakness of hard-to-imagine-otherwise arguments. Still, if you are arguing that a full-fledged metarepresentational ability is something learnt (in the sense of learning theory), and don't illuminate how this learning might take place, then you are signing a huge promissory note the equivalent of which has never been honored (whereas if you assume that a metarepresentational ability is a biological development, the promissory note you are signing is more modest: there are well-understood biological developments of much greater complexity).

Language - rather than metarepresentational ability - is usually taken to be the most distinctive feature of the human species. The two are clearly linked: natural languages serve as their own meta-language and thus incorporate a full-fledged metarepresentational capacity. Linguistic utterances are public representations, and are typical objects of mental metarepresentation. Speakers in intending an utterance and hearers in interpreting an utterance mentally represent it as a bearer of specified content, i.e. they metarepresent it.

Language and metarepresentations are made possible by biologically evolved mental mechanisms, which, it has been argued in both cases, are domain-specific. Noam Chomsky arguing for a domain-specific language faculty (e.g. Chomsky 1975, 1980) introduced the very idea of domain-specificity in the cognitive sciences. The idea of metarepresentations became familiar in cognitive science through work on naïve psychology or "Theory of Mind". In suggesting in their 1978 article "Does the chimpanzee have a theory of mind?", that the ability to attribute mental states to others was also found among non-linguistic animals, Premack and Woodruff were implying that this ability, at least in its simpler forms, is independent of language (see also Premack 1988). Developmental psychologists have argued that theory of mind is based on a domain-specific mental module (Leslie 1987, 1994; Baron-Cohen 1995).

If one accepts, as I do, the existence of two dedicated mental mechanisms, one for language, the other for metarepresentations, it seems reasonable to assume that, in humans, they have co-evolved. While the fully developed version of each of these two mechanisms may presuppose the development of the other, it still makes sense to ask which of these two, the linguistic or the metarepresentational, might have developed first to a degree sufficient to bootstrap the co-evolutionary process. At first blush, the language-first hypothesis seems quite attractive (and has attracted, for instance Daniel Dennett 1991).

The hypothesis of the language faculty evolving first may seem, moreover, to offer a way of explaining how a metarepresentational ability might emerge in individual cognitive development, even in the absence of a biologically evolved specialized disposition. Linguistic communication fills the environment with a new kind of object: utterances, i.e. public representations. Utterances can be perceived, attended to, thought about, just as any other perceptible object in the environment. At the same time, they are representations, they have meaning, content. It may seem imaginable, then, that children find linguistic representations in their environment, grasp the representational character of these utterances because they are linguistically equipped to assign them content, and, as a result, develop an ability to represent representations qua representations. This acquired metarepresentational ability would apply first to linguistic utterances, and then would extend to other types of representations, in particular mental representations.

This hypothesis loses much of its appeal under scrutiny. Spelling it out results in a worrisome dilemma. Either ancestral linguistic communication, though presumably simpler in many respects, was based on the same kind of communicative mechanism as modern linguistic communication. If so, it presupposed metarepresentational ability, and therefore could not precede it. Or else, ancestral linguistic communication was strictly a coding-decoding affair, like other forms of non-human animal communication, and then there is no reason to assume that our ancestors had the resources to become aware of the representational character of their signals, anymore than bees or vervet monkeys do.

When we, modern humans, communicate verbally, we decode what the words mean in order to find out what the speaker meant. Discovering sentence meaning is just a means to an end. Our true interest is in speaker's meaning. A speaker's meaning is a mental representation entertained by the speaker which she intends the hearer to recognize and to which she intends him to take some specific attitude (e.g. accept as true). Verbal understanding consists in forming a metarepresentation of a representation of the speaker (in fact a higher order metarepresentation, since the speaker's representation is itself a metarepresentational intention). Moreover, there is a systematic gap between sentence meaning and speaker's meaning. Sentences are typically ambiguous and must be disambiguated ; they contain referring expression the intended referent of which must be identified ; they underdetermine speaker's meaning in yet many other ways. Linguistic utterances fall short, typically by a wide margin, of encoding their speaker's meanings. On the other hand, utterances are generally good pieces of evidence of these meanings. Inference to the best explanation of the speaker's linguistic behavior generally consists in attributing to her the intention to convey what actually was her meaning in producing her utterance. Linguistic comprehension is an inferential task using decoded material as evidence. The inferences involved are about speaker's meaning, that is, are aimed at metarepresentational conclusions.

If the ability to communicate linguistically had preceded the ability to use metarepresentations, then this pre-metarepresentational, ancestral verbal ability would have been radically different from the kind of verbal ability we modern humans use - which is metarepresentational through and through. The ancestral language would have been a coding-decoding affair, as are the many forms of non-human animal communication we know about. This, in itself, is an unattractive speculation, since it implies a radical change in the mechanism of human linguistic communication at some point in its evolution.

Even more importantly, there is no reason to assume that a decoding animal experiences or conceptualizes the stimulus it decodes as a representation: for non-human animals, coded communication and attribution of mental states to others are two unrelated capacities (the second apparently much rarer than the first). What seems to happen, rather, is that the decoding of the stimulus automatically puts the animal in a specific cognitive, emotional, or motivational state appropriate to the situation. For instance, the decoding of an alarm signal typically triggers a state of fear and activates flight plans. It would be quite unparsimonious to hypothesize that the decoding animal is able to recognize the signal as a signal endowed with meaning. It is more sensible (and the burden of proof would be on whoever maintained otherwise) to assume that, to mere coding-decoding communicators, signals are transparent. Such signals play a role comparable to that of proximal stimuli in perception : they occur at an uncognized stage of a cognitive process.

If our ancestors were such coders-decoders, and had no evolved disposition to metarepresent, then there is no sensible story of how the presence of utterances in their environment would have led them to discover their representational character, to metarepresent their content, and to use for this their own mental representations in a novel, opaque, manner.

Out goes the language-first hypothesis.

What about the metarepresentations-first hypothesis ? A metarepresentational, and more specifically a metapsychological ability, may be advantageous and may have evolved on its own. This has been convincingly argued in much recent literature on 'Machiavellian intelligence' (Byrne and Whiten 1988; Whiten and Byrne 1997). The ability to interpret the behavior of intelligent conspecifics not just as bodily movement but as action guided by beliefs and desires gives one a much-enhanced predictive power. Predicting the behavior of others helps to protect oneself from them, to compete successfully with them, to exploit them, or to co-operate more profitably with them. A metarepresentational ability is a plausible adaptation quite independently of communication.

Moreover, a well-developed metarepresentational ability makes certain forms of communication possible quite independently from any code or language. Organisms with metarepresentational abilities live in a world where there are not only physical facts but also mental facts. An individual may form beliefs or desires by emulating those it attributes to another individual. An individual may want to modify the beliefs and desires of others. It may want others to become aware of its beliefs and desires and to emulate these. Let me illustrate. Imagine two one of our hominid ancestors, call one Mary and the other Peter.

Once the level of metarepresentational sophistication of our fifth scenario is reached, indeed, a dramatic change occurs. Whereas, before, in order to fulfil her first-level informative intention, Mary had to engage in behavior - picking the berries - that was best explained by attributing to her the belief that the berries were edible and the desire to collect the berries, she can now resort to symbolic behavior the best explanation of which is simply that she is trying to fulfil an informative intention, her desire to inform (or misinform) Peter that the berries are edible. Instead of actually picking the berries, she might, for instance, mime the action of eating the berries.

Typically, symbolic behavior such as miming has no plausible explanation other than that it is intended to affect the beliefs or desires of an audience. This generally is the one effect of such a behavior that could clearly have been intended by the agent. Thus Mary miming eating the berries does not feed her, does not help her in any easily imaginable way, except through the effect it has on Peter. Her miming behavior triggers in Peter's mind, through the perception of a resemblance between the miming behavior and the behavior mimed, the idea of eating the berries. Moreover, if it is at all relevant to Peter to know whether or not the berries are edible, Mary's behavior suggests that they are. Provided that Peter sees this miming behavior as intended by Mary to inform him of an informative intention of hers, he is justified in assuming that the very idea triggered in his mind was one Mary wanted him to entertain, elaborate and accept.

The same result achieved by miming, could, provided Peter and Mary shared an appropriate code, be achieved by means of a coded signal, or by means of some combination of iconic and coded behavior. Suppose Mary and Peter shared a signal meaning something like "good". Then Mary might point to the berries and produce this signal, again, triggering in Peter's mind the idea that the berries were good to eat, doing so in an manifestly intentional manner, and therefore justifying Peter in assuming that she intended him to believe that the berries were edible.

Note that, if Mary used the coded symbol "good" in this way, she would, nevertheless, be very far from encoding her meaning that these berries are edible. She would merely be giving evidence of her intention to cause Peter to come to accept this meaning as true. The use of coded signals as part of the process of communication is no proof that the communication in question is of wholly or even essentially of the coding-decoding kind.

Metarepresentational sophistication allows a form of inferential communication independent of the possession of a common code. This type of inferential communication, however, can take advantage of a code. It can do so even if the signals generated by the code are ambiguous, incomplete, and context-dependent (all of which linguistic utterances actually are). By triggering mental representations in the audience, coded signals provide just the sort of evidence of the communicator's informative intention that inferential communication requires, even if they come quite short of encoding the communicator's meaning.

To conclude this section, there is a plausible scenario where a metarepresentational ability develops in the ancestral species for reasons having to do with competition, exploitation and co-operation, and not with communication per se. This metarepresentational ability makes a form of inferential communication possible, initially as a side effect, and, probably, rather painstakingly at first. The beneficial character of this side effect turns it into a function of metarepresentations, and creates a favorable environment for the evolution of a new adaptation, a linguistic ability. Once this linguistic ability develops, a co-evolutionary mutual enhancement of both abilities is easy enough to imagine.

Current discussions have focused on the metapsychological use of a metarepresentational ability to represent mental states, the beliefs and desires of others and of oneself. Humans however metarepresent not just mental representations, but also public representations and representations considered in the abstract, independently of their mental or public instantiation. The three types of metarepresented representations are simultaneously illustrated in (1) and individually highlighted in (6a-c):

In (6a), the italicized phrase metarepresents a mental representation of Peter's, in (6b), it metarepresents an utterance, i.e. a public representation of Mary's, and in (6c), it metarepresents a representation considered in the abstract (as a hypothesis), independently of whoever might entertain or express it.

Representations considered in the abstract are reduced to their logical, semantic, and epistemic properties: they may be true or false, self-contradictory or necessarily true, plausible of implausible, standing in relationships of entailment, of contradiction, of warrant, of being a good argument one for the other, of meaning similarity, etc. They may be normatively evaluated from a logico-semantic point of view (and also from an aesthetic point of view).

Mental and public representations have the same content properties as their abstract counterparts - or arguably, abtract representations are nothing but the content properties of concrete representations, abstracted away. Mental and public representations also have specific properties linked to their mode of instantiation. A mental representation occurs in one individual; it is causally linked to other mental and non-mental states and processes of which the individual is wholly or partially the locus. A mental representation can be sad or happy, disturbing or helpful, it can be normatively evaluated in psychological terms as poorly or well reasoned, as imaginative, as delusional, and so on. A public representation typically occurs in the common environment of two or more people; it is an artifact aimed at communication. It exhibits such properties as having a certain linguistic form, as being used to convey a certain content in a certain context, as being more or less attention-grabbing, more or less comprehensible to its intended audience, and so on. It can be normatively evaluated from a communicative point of view as sincere or insincere, intelligible or unintelligible, relevant or irrelevant.

The properties of these three types of representations, mental, public, and abstract, do not constitute three disjoints sets and are not always easy to distinguish. As mentioned above, mental and public representations have the logico-semantic properties of their abstract counterparts. A belief or an utterance is said, for instance, to be true or false when its propositional content is. Public representations serve to convey mental representations and have, at least by extension, some of the properties of the mental representations they convey. An utterance is said, for instance, to be imaginative or delusional when the thought it expresses is. Still, many properties are best understood as belonging essentially to one of the three types of representation, and as belonging to another type of representation, if at all, only by extension, in virtue of some relational property that holds between them (such as the relationship of expression that holds between an utterance and a thought).

There is no question that we, modern humans, can attend in different ways to these three types of representations. We can attribute mental states, interpret public representations, and reflect on the formal properties of abstract representations. Are these performances all based on a single metarepresentational ability or do they, in fact, involve different competencies? In the latter case, is it plausible that these competencies might each be a distinct evolved adaptation? Could there be several metarepresentational "modules"? In the evolutionary psychology literature on the one hand, in the developmental psychology literature on the other, the only metarepresentational adaptation envisaged is a metapsychological "theory of mind" the main function of which is to predict behavior of others. Even a peculiar behavior such as collective pretense play, which involves essentially the creation and manipulation of public representations, is often treated as a regular use of "theory of mind."

What does a metarepresentational ability whose function is basically metapsychological do? What kind of inferences does it draw? Which properties of representations does it attend to? It draws inferences from situation and behavior to mental states as in (7a-c), from mental states to other mental states as in (8a-b), and from mental states to behavior as in (9a-b):

A metapsychological ability assumes that others have some basic knowledge and basic drives, and attributes to them specific beliefs and desires derived through perception and inference. Does such an ability, with the kind of power we are entitled to grant it on evolutionary and developmental grounds, suffice to explain the kind of metarepresentational processes we engage in when we interpret public representations, in particular utterances, or when we attend to the formal properties of abstract representations?

Comprehension (or its pragmatic layer) is an inferential process, using as input the output of linguistic decoding, and aiming at discovering the speaker's meaning. Comprehension consists, therefore, in inferring a mental state (an intention of a specific kind) from behavior (an utterance). It might seem that this is precisely the kind of result a metapsychological ability should be able to achieve. In the story of Mary, Peter, and the berries above, I tried to illustrate how, in principle, a multi-leveled metapsychological ability might make inferential communication possible. The communication achieved between Mary and Peter in the fifth scenario, was a metarepresentational prowess. In modern humans however, comparable or greater metarepresentational achievements have become routine: we communicate at great speed much more complex contents than Mary and Peter ever could. This is a first reason to hypothesize that a more specialized adaptation aimed at comprehension has evolved. A second reason has to do with the pattern of inference from observed behavior to attributed intention.

In the simplest case, behavior (say, throwing a stone) is observed to have several effects (making a noise, casting a moving shadow, killing a bird). One of these effects (killing a bird, as it might be) can be seen as desirable to the agent. It is then inferred that the agent performed the behavior with the intention of achieving this desirable effect. Inferring intentions is somewhat harder when the behavior fails to achieve its intended effect. Say the bird is merely frightened away, rather than killed. Competent attributers of mental states will nevertheless recognize that throwing the stone could have killed the bird and that the agent could expect this effect (with greater or lesser confidence). They will then infer that this unachieved effect was the one intended.

In the case of attributions of speaker's meaning, these standard patterns of inference (from behavior to intention through identification of a desirable actual or failed effect) are not readily available. The essential effect intended by a speaker, i.e. comprehension of her meaning, cannot be achieved without the very recognition of her intention to achieve this effect. The intended effect cannot, therefore, be independently observed to occur and then be recognized as desirable and presumably intentional. Nor does it make sense to imagine that the comprehender might recognize as the effect intended an effect that might plausibly have occurred but that in fact failed to, since the recognition of the intended effect would secure its occurrence.

There are, in the literature, two basic ways of solving the puzzle raised by inferential comprehension. The first way is Grice's (1989). It aims at showing a way around the difficulty while remaining within the limits of standard belief-desire psychology. It consists in assuming that the decoding of the linguistic signal provides a default assumption regarding the speaker's meaning. By default, the speaker is assumed to mean what the sentence she utters means. This default assumption can be inferentially complemented or otherwise corrected when there is a mismatch between it and general assumptions about standing goals the speaker is presumed to aim at in her verbal behavior, goals codified by Grice in term of his Co-operative Principle and Maxims. Such inferential correction involves a form of metarepresentational reasoning about the speaker's intentions where not only the conclusion (an attribution of meaning to the speaker) is metarepresentational, but where also some of the premises are. For instance, in the case of a metaphor such as (10), the hearer is supposed to reason somewhat as in (11):

It is as if, by flouting the maxim of truthfulness, the speaker deliberately failed to achieve a certain effect, thus suggesting that the truly intended effect is in the vicinity of the overtly failed one.

As this example illustrates, Gricean pragmatics can be seen as an account where the inferential part of comprehension consists in applying commonsense psychology to verbal behavior. If Grice were right, a general metapsychological ability, together with a presumably socially acquired knowledge of the Co-operative Principle and the Maxims, would be sufficient to account for inferential comprehension.

Gricean pragmatics might seem attractively parsimonious, since it does not require any specialized comprehension ability. However, the economy in terms of the number of mental devices one may be led to postulate is more than offset by the cumbersome character of the inferences Gricean pragmatics necessitates every time speaker's meaning diverges from sentence meaning (and we have argued that it always so diverges). Do we really have, in the case of implicature, of indirect speech acts, of metaphor, or of irony, to reflect on what the speaker knows we know she knows, on what respecting the maxims requires of her, on what she might mean and not mean? Does it take more effort and hence longer to process such utterances? (The answer is no, see Gibbs 1994, and this volume). Do we want to attribute to young children these complex inference patterns, or to deny them the ability to comprehend metaphor and other forms of so-called indirect speech? As a rational reconstruction of how inferential comprehension might be possible, Grice's account, though not without problems, is certainly appealing. As a psychologically realistic account of the mental processes actually involved in comprehension, it is much less so.

Though we owe a great deal to Grice's inspiration, Deirdre Wilson and I have criticized his pragmatics and, among other aspects, this account of metaphor. We have developed relevance theory as another way to solve the puzzle raised by inferential comprehension (Sperber and Wilson 1986/1995). In relevance theory, we assume that comprehension follows a very specific inferential pattern suited for the discovery of the informative intentions of communicators. We define the relevance of a cognitive input to an individual as a positive function of the cognitive effects achieved by processing this input and as a negative function of the amount of effort involved in this processing. We argue that every utterance (and more generally every communicative act) conveys a presumption of its own relevance. We show that this justifies the following inferential procedure: follow a route of least effort in constructing an interpretation - taking the very fact that an element of interpretation comes first to mind as an argument in its favor - until the effects achieved are sufficient to warrant the presumption of relevance conveyed by the utterance, and then stop. To illustrate briefly: Given the context, utterance (10) might be capable of activating in the mind of the hearer the ideas in (11a-g) in that order:

Suppose that the utterance, when interpreted as conveying (11a-d), satisfies the expectations of relevance it has itself raised. Then, we predict that the interpretation will stop at this point, and that (11g), i.e. the literal interpretation won't even be considered. In another context, the order in which the elements of interpretation might come to the mind of the hearer would be different, and the stopping point might be such that the overall interpretation would include (11g) and be literal. Suppose, for instance, that "John is a soldier" was said in response to the question "What does John do for a living?" Then (11a-g) would probably be accessible in the reverse order, from (g) up. The hearer, moreover would likely stop at (f) or (e). The resulting interpretation (11g-e) would be literal, and not even overlap with the metaphorical interpretation (11a-d) of the same sentence uttered in a different context. In all cases, however, the comprehension procedure is the same: follow the path of least effort until adequate relevance is achieved. This may yield a literal, a loose, or a metaphorical interpretation without the comprehender having to take notice of the type of interpretation achieved.

The conclusion of such an interpretation process is an attribution of a meaning to the speaker, hence a metarepresentation, but the premises in the inference process need not be metarepresentational. Therefore this procedure can be followed by a relatively unsophisticated metarepresenter, for instance by a young child capable, as young children are, of comprehending metaphor. On the other hand, how would the child, or for that matter the adult, discover this procedure and recognize that it is reliable means to infer a speaker's meaning? A plausible answer is that this procedure is not individually discovered, but is biologically evolved. It is an evolved module.

The relevance-based comprehension procedure could not be soundly applied to the discovery of non-communicative intentions. Non-communicative behavior carries no presumption of relevance to possible observers. The order in which elements of interpretation come to the mind of the observer has no particular epistemic value. There is no level where, expectations of relevance being satisfied, the observer is thereby justified in believing that his interpretation of the agent's intentions is complete.

What could be the relationship between a relevance-based comprehension module and a more general metapsychological module? The former might be a sub-module of the latter, in the manner in which linguistic acoustic abilities are a sub-module of general acoustic abilities. Speech sounds are sounds, and their perception recruits the hearing system. At the same time, speech sounds are, from birth onwards, attended and processed in a proprietary way. Similarly the recognition of communicative intentions might be a biologically differentiated and stabilized sub-system of human naïve psychology.

Is the human metapsychological ability enough to explain the human ability to attend to formal, and in particular to logical properties of representations? A metapsychological ability attributes inferences to others. Inferences must, on the whole, respect logico-semantic relationships such as entailment or warrant - i.e. relationships holding among representations in the abstract - or else they are not truly inferences. Successful attribution of inferences to others must also respect these relationships. Still, if we suppose that the organisms to which mental states are attributed have quite limited inferential abilities, then the attributing organisms - the metapsychologists - need attend only to a few logical relationships. Moreover, this attention need not be reflective. That is, the metapsychologists need not be logicians, thinking about logical relationships; it is enough that they be able to make themselves, offline so to speak, the inferences they attribute. (This is sometimes presented as the simulation view, but there is a weak and sufficient version of it that is equally compatible with the simulation view à la Goldman, and with the theory of mind view à la Leslie.)

The type of metapsychological competence that is likely to be an evolved adaptation is unlikely to explain the modern human ability to attend to abstract representations, and in particular to do formal logic. However, it is dubious that such an ability is widely shared. It might be just a skill acquired with some difficulty by a minority of people in scholarly institutions.

In fact, psychologists favorable to an evolutionary point of view have expressed doubts as to the existence of a genetically determined domain-general "logical" ability. If by a "logical ability" what is meant is a unitary, domain-general ability that would govern all human inferences, then indeed, positing such an ability would go against well-known evolutionary arguments for the domain-specificity of mental mechanisms (see Sperber 1996, chapter 5). On the other hand, if we think of the kind of logical ability that is involved in formulating or evaluating arguments, far from being domain-general, this is a highly domain-specific metarepresentational ability: its domain is some specific properties of abstract representations. An appearance of domain-generality might come from the fact that the representations that such a device would handle could be representations of anything. The device, however, would just attend to some formal properties of these representations, independent of what they happen to represent, and would be as specialized as, say a word-processor grammar-checker, which is able to process statements about anything, but is obviously not domain-general for all that.

Still, what is the plausibility, from an evolutionary point of view, that a logical module specialized in checking the validity of arguments would have evolved? It might seem low, but let me speculate and offer an argument as to why such a logical module with a fairly specific function is not so implausible.

Though the conclusion is quite different, the structure of my argument is parallel to Leda Cosmides's argument on cheater detection (Cosmides 1989). She draws on the neo-Darwinian argument according to which "reciprocal altruism" is unstable unless there is some form of control for cheating, since otherwise cheaters would be more successful until there were not enough genuine altruists left to cheat. In the case of complex forms of reciprocal exchange, as found among humans, the prevention of cheating is an elaborate cognitive task that is likely, Cosmides argues, to have caused the evolution of an ad hoc adaptation, a cheater-detection mechanism. As usual, it is an open question as to whether the adaptation that handles the problem is precisely calibrated for the task, or is in fact a narrower, or on the contrary a larger ability that was still the best biologically available solution for the task. I don't want here to discuss Cosmides's basic argument, which I find illuminating, but to draw inspiration from it (for a critical discussion of her use of the selection task to provide empirical evidence for this argument, see Sperber, Cara & Girotto 1995).

Communication is a form of co-operation that seems particularly advantageous for animals that depend as much as humans do on their cognitive resources. Instead of being restricted in one's knowledge to the products of one's own experiences and thinking, communication makes experience and thinking available by proxy. Alas, as with other forms of co-operation, communication makes one also vulnerable to misinformation, deception, and misguidance. Of course, one could protect oneself from the harms of deception by being systematically mistrustful, but then one would lose, by the same token, the benefits of communication. Communication is advantageous only if it is paired with mechanisms that ensure the proper calibration of trust. It is even more advantageous, if, while protected from the deception of others without being overprotected, you can penetrate their protection and deceive them.

The human reliance on communication is so great, the risks of deception and manipulation so ubiquitous, that it is reasonable to speculate that all cost-effective available modes of defense are likely to have evolved. There are several such modes. One is to be discriminative regarding whom to trust and to accept authority quite selectively. Clearly, humans do this. Another is to be sensitive to subtle signs of deceptive intent, to read the relevant attitudinal and emotional signs. These safeguards can be breeched, as they are by professional swindlers who typically "look absolutely trustworthy."

I want to focus on yet another possible protective mechanism against misinformation: Check for the consistency of the information communicated, both for its internal consistency and its consistency with what you already believe. As anybody who has ever tried to lie knows, the liar's problem is to maintain consistency, both internally and contextually. An organism that drew its knowledge only from its senses and a few reliable inferential routines would probably waste energy in checking the consistency of the information it acquired. Inconsistencies in perception-based information occur, but they are rare, and trying to eliminate them may not be worth the cost. For richly communicating animals like us, eliminating inconsistencies may not only be worth the cost, but indeed, literally life-saving.

So, I hypothesize the emergence of an ability to check for consistency and to filter incoming information on this basis. But this is only the first step in a persuation-counterpersuation arm race. More advantageous than merely protecting yourself from misinformation, is to combine this with the ability to freely inform and misinform others, to persuade them of what it is in your interest to persuade them of, whether true or false. Persuaders addressing consistency-checkers cannot do better than displaying the very consistency - or at least the appearance of it - that their audience is likely to check for. Communicators now express not just the propositions they want their audience to accept, but arguments to accept these propositions, and the very argumentative structure that leads to the intended conclusion. The language is enriched with logical terms ("and", "or", "if", etc.) and para-logical terms ("therefore", "but", "since", "although", "even", etc.) that display the real or alleged consistency and logical force of the communication.

The next step in the persuation-counterpersuation arms race is the development of the ability to scrutinize these argumentative displays, and to find fault with them. In other terms, I am surmising, on evolutionary grounds, the development of a very special kind of "logical", or logico-rhetorical ability. It is special in that it attends to logical relationships not per se, nor for the sake of the benefits that good reasoning can give the individual thinker, but as a means to filter communicated information, and, on the other hand, as a means to penetrate the filters of others. Such a speculation has experimentally testable implications. It predicts, for instance, that logically equivalent tasks will yield significantly better performance when they are presented in a context where subjects are scrutinizing arguments plausibly aimed at persuading them, than when they are evaluating these arguments in the abstract (as happens in most experimental tasks).

My aim in the present chapter has been to envisage the possibility that humans might be endowed, not with one, but with several evolved metarepresentational abilities. I have considered argument as to why, beside the standard metapsychological ability, they might have a comprehension module aimed at the online interpretation of utterances, and a logico-argumentative module, aimed at persuading others and not being too easily persuaded oneself. If there are indeed three such metarepresentational modules, the obvious next step would be to envisage their evolution not separately, but, much more systematically than I have done here, as a case of co-evolution, involving also the "language instinct", and - dare I say it? - consciousness.

I am grateful to Steven Davis, Gloria Origgi, and Deirdre Wilson for their useful comments on earlier versions of this chapter.