1 MINDREADING IN THE ANIMAL KINGDOM? José Luis Bermúdez In R. Lurz (Ed.), Animal Minds (Cambridge University Press, 2009), pages 145-164 Introduction Can non-human animals think and reason about what other creatures are thinking, reasoning, or experiencing? Experimentalists, ethologists, and theorists have answered this deceptively simple question in many different ways. Some researchers have made very strong claims about the socalled mindreading abilities in animals (Byrne and Whiten 1988, 1990, 1991; Premack and Woodruff 1978, Hare, Call, and Tomasello 2001, Tomasello and Call 2006; Hare et al, 2002; Dally, Emery, and Clayton 2006; Tschudin 2001). Others have been critical of such claims (Heyes 1998; Povinelli and Vonk 2006, Penn and Povinelli 2007). Even a cursory look at the extensive literature on mindreading in animals reveals considerable variation both in what mindreading abilities are taken to be, and in what is taken as evidence for them. The first aim of this paper is to tackle some important framework questions about how exactly the mindreading hypothesis is to be stated. In sections 1 and 2, three importantly different versions of the mindreading hypothesis are distinguished. The first (which I call minimal mindreading) occurs when a creature’s behavior covaries with the psychological states of other participants in social exchanges. The second (which I call substantive mindreading) involves attributions of mental states. In section 2, substantive mindreading is further divided into propositional attitude mindreading and perceptual mindreading. In section 3, I present reasons for thinking that the role of propositional attitude psychology in human social life is very much over-stated and show that this very much weakens the analogical case for identifying propositional attitude mindreading in nonlinguistic creatures. And in section 4, I present a revised version of an argument I have given elsewhere (Bermúdez 2003) to show that the most sophisticated form of substantive mindreading (the type of mindreading that exploits the concepts of propositional attitude psychology) is only available to language-using creatures. I Minimal and substantive mindreading My starting-point is that many types of animal are genuine thinkers. I have discussed this at length elsewhere (Bermúdez 2003) and will not rehearse the arguments again here. The evidence from comparative psychology and cognitive ethology overwhelmingly supports taking some forms of animal behavior to be genuinely psychological, generated by primitive forms of belief and desire via processes that have significant commonalities with the forms of reasoning engaged in by language-using creatures. Animals are capable of sophisticated social behaviors. Many of these social behaviors do not have any sort of psychological dimension. Schooling and flocking behaviors are obvious examples. And there are behaviors with a psychological dimension that do not involve any social coordination, as in cases of emotional contagion. But there do appear, at least at first sight, to be forms of social coordination in the animal kingdom that have a psychological dimension and that involve a sensitivity to the psychological states of other participants in the interaction. Here is an example of how non-linguistic creatures can exploit social cues. A well-known set of experiments by Brian Hare and collaborators have revealed that domestic dogs are strikingly successful on object choice tasks with social cues (Hare et al. 2002). One reason these results are striking is that most primates, which are generally thought to have quite sophisticated socialcognitive skills, seem unable to perform above chance on object choice tasks. In a standard object choice task, an experimenter hides a food reward in one of two opaque containers. The 2 subject, which did not see the food being hidden, has to choose between the two containers. Before the animal is presented with the choice the experimenters “signals” which container the food is in by using one of a range of communicative cues (such as pointing to, marking, or looking at the correct container). Hare et al. found that domestic dogs master object choice tasks very quickly, often without any learning. The success of domestic dogs in picking up and exploiting social cues to solve object choice tasks is a paradigm example of social coordination with a psychological dimension – as opposed to coordinated group behaviors, such as schooling or flocking, where (to simplify somewhat) an individual’s behavior depends simply upon changes in the behavior of other participants in the coordinated group behavior. That there is social coordination is obvious. What makes it social coordination involving sensitivity to psychology is that the dogs behave in ways that depend upon changes in the psychological states of the other participant in the interaction. The dogs in the object choice tasks are able to respond to different visual cues. These cues all have something in common. They all have a common cause in the psychological profile of the experimenter – namely, the experimenter’s intention to signal to the animal the location of the reward. This allows us to extrapolate to predictions about how the dogs will behave in future tests – namely, that they will respond to visual cues that have the same cause and origin. In essence, we assume that the dogs are responding to cues in the abstract, rather than to the physical gestures by which those cues are made – a “multi-track” sensitivity, as opposed to a set of contingencies between particular responses and particular stimuli. This account of social coordination involving sensitivity to psychology is quasi-operational. While it goes beyond observed behavior in making reference to the psychological states of the experimenter, it does not go beyond the observed behavior of the experimental subject. The experimental subject is characterized in purely behavioral ways. Saying that a non-linguistic creature displays psychological sensitivity in this quasi-operational sense does not attribute to it any psychological states. Displaying sensitivity to psychology only requires behaving in ways that depend upon the psychological states of another participant in the interaction. It is neither implied nor required that an animal displaying sensitivity to psychology should represent (or even be capable of representing) the psychological states of the other participant(s) in the exchange. I will use the expression “minimal mindreading” to describe what is going in instances of social coordination involving sensitivity to psychology of this type. Here is the official definition: A creature engages in minimal mindreading when its behavior is systematically dependent upon changes in the psychological states of other participants in the interaction. Characterizations of minimal mindreading are descriptive rather than explanatory. To say that an animal is engaged in minimal mindreading is simply to assert that certain contingencies hold between its behavior and the psychological states of the creatures with which it is interacting. It is not in any sense to say why those contingencies hold. The expression “substantive mindreading”, in contrast, is intended to be explanatory. Attributions of substantive mindreading are made in order to explain how and why an animal’s behavior depends systematically upon the psychological states of other participants in the interaction. What explains the dependence, it is typically claimed by those who identify substantive mindreading in the animal kingdom, is the fact that the animal engaged in a social interaction is mentally representing the psychological states of other participants in the interaction. Here is the matching official definition. 3 A creature engages in substantive mindreading when its behavior is systematically dependent on its representations of the psychological states of other participants in the interaction.1 Although the notion of systematic dependence features in the definition both of minimal and of substantive mindreading, it is doing different work in each. In the definition of minimal mindreading, the systematic dependence is not intended to be causal. Minimal mindreading is covariation pure and simple. In contrast, claims of substantive mindreading are intended to be causal. The animal’s behavior is caused by (and hence can be explained by appeal to) how it represents the mental states of others. 2 Two types of substantive mindreading Substantive mindreading occurs when a creature behaves in ways that depend systematically upon how it represents the psychological states of other participants in the interaction. But there are different types of psychological state and, correspondingly, different types of substantive mindreading. In this section I articulate what I take to be the most important distinction in this area, both from a theoretical and from an experimental perspective. We can start from the obvious fact that psychological states do not typically generate behavior on their own. Behavior is the product of complexes of psychological states – of what might be thought of as psychological profiles. We can only think about the behavioral implications of individual psychological states through the prism of the subject’s psychological profile. This is part of what makes studying the psychology of nonlinguistic creatures so challenging. A welldesigned experiment tries to find behavioral criteria for the presence or absence of a particular form of psychological state. No interesting psychological states have unambiguous and unequivocal implications for behavior, however. So assumptions have to be made about the subject’s more general psychological profile and, of necessity, these assumptions are not themselves under investigation and scrutiny. This means that there is always an element of bootstrapping going on when we explore the psychological lives of non-linguistic creatures. Fortunately, psychological states lie on a continuum in terms of the directness of their implications for behavior. At one end of the continuum are psychological states with more or less immediate implications for behavior. At the other end lie the psychological states that feed into action only very indirectly. We can think about where an individual psychological state lies on the continuum in terms of the complexity and particularity of the background psychological profile required for it to issue in action. In some cases the background psychological profile is very simple, given by a relatively fixed set of drives and goals that may well be constant across individuals within a given community, or even across a given species. In other cases the background psychological profile is highly complex and highly individual. This basic fact about the relation between psychology and action has important implications for thinking about substantive mindreading. A creature engages in substantive mindreading to the extent that its behavior depends systematically upon how it represents the psychological states of others. A mindreading creature behaves in ways that reflect its predictions about how other creatures are going to behave – predictions derived from representations of their psychological states. It is clear that certain conditions have to be met for these predictions to be successful. It is not enough simply that the mindreading creature represent the psychological states of other participants. Or even that it represent those psychological states accurately. The success of behavioral predictions stands or falls with their being in conformity with the background psychological profile of the creature whose behavior is being predicted. I say that a prediction is 1 Behavior can be understood in a thin sense here. Preferential looking counts as a behavior, for example. 4 in conformity with a background psychological profile when the predicted behavior is the behavior that would result from the combination of the (accurately represented) psychological state and the background psychological profile. What does it take to secure conformity with the background psychological profile? One way to secure conformity would be through explicitly (and accurately) representing the background psychological profile in order to apply some set of principles that connect psychology with behavior. These principles may be proto-theoretical, as proposed by adherents of the “theory of mind” approach to mindreading. In this case the principles themselves are explicitly represented. Or they may be principles governing the subject’s own decision-making processes, as suggested by supporters of the simulationist approach.2 On the simulationist view the principles are not explicitly represented. The particular principles, and how they are applied, are less important than the “raw materials” on which they operate. In this case the raw materials are representations of psychological profiles. But conformity can also be achieved without explicit representation. In cases where the relevant elements of the background psychological profile are generic and widely held it is possible simply to trade on them. So, for example, if I put a $100 bill in plain view on the sidewalk and I correctly identify someone as catching sight of it, I am fairly safe in predicting that they will bend over to pick it up. The background psychology required to generate this behavior is nothing more than a desire for free money, which can we can safely assume to be constant across the human population – so constant in fact that there is no need explicitly to represent it, and certainly no need to delve any deeper into the particularities of the individual’s psychology. In this case a reasoner who moves directly from the observation that a person has seen the $100 bill will move directly (and with justification) to a prediction that the person will bend over to pick it up. In almost every case this prediction will be accurate. Again, we have two end-points on a continuum. The more complex and variable the relevant elements of the background psychological profile, the more necessary explicit representation becomes, and the more extensive it has to be. I am taking complexity and variability here to be distinct phenomena. Complexity itself does not mandate explicit representation. Predictors can trade on constant elements of the background psychological profiles when those elements are generic and widely held, no matter how complex they are. The real problem is created by variability. If there are many different ways in which an agent’s psychology might be configured relative to the behavior being predicted, then there are all sorts of ways in which a prediction might go wrong, even if the prediction is based on a completely accurate psychological attribution. Making predictions that involve explicitly representing a background psychological profile can be a substantial intellectual achievement. It typically involves, for example, representing the range of different motivational states that an agent has, together with the information they currently possess about the environment and a range of more general beliefs. But it is not enough, of course, simply to represent the states. The representer must also represent how they fit together and how they might jointly determine a particular action. To put it another way, the representer must reason about how the agent might reason their way to a particular action. This reasoning is typically conscious – and even when it proceeds below the threshold of consciousness it is consciously accessible. The activity of explanation and prediction is a personal-level activity (in the sense of Dennett 1969 – see Ch. 1 of Bermúdez 2005 for further discussion). 2 For the debates between simulationist and theory-theory approaches to mindreading see the papers in Davies and Stone 1995 and Carruthers and Smith 1996. 5 These general observations about the different ways in which identifying another’s psychological state can generate predictions of behavior have important implications for how we think about substantive mindreading in nonlinguistic creatures. Both experimentalists (Tomasello and Call 1997 Pt 2, 2006) and philosophers (Bermúdez 2003) have noted that substantive mindreading is not a unitary phenomenon. There are different types of mindreading, varying according to the type of psychological state that they involve. Predictions based on representations of perceptual states, for example, reflect a different type of mindreading from predictions based on representations of beliefs and desires. Standardly these types of mindreading are distinguished simply as a function of differences between the relevant represented psychological states, on the tacit assumption that if psychological states are different in type then representing them requires distinct abilities. The current observations give us a principled way of categorizing mindreading abilities that lines up in important respects with the standard distinctions. Consider, for example, the psychological states that philosophers standardly term propositional attitudes. These includes beliefs, desires, hopes, fears, and so on. What they have in common, from a philosophical point of view, is that they can all be analyzed as attitudes that thinkers and reasoners have towards propositions. We will look in more detail later on at what propositions might be and why this is important for thinking about the mindreading abilities of nonlinguistic creatures. For the moment the important point is that the propositional attitudes are collectively located at one end of our continuum. They typically do not have direct implications for action. There is no single way that a particular belief or desire will feed into action. Individual beliefs and desires feed into action only indirectly, as a result of an agent’s specific psychological profile. This holistic character of the propositional attitudes places a very specific burden on mindreading that involves attribution of propositional attitudes, because it brings into play the distinctiveness of the agent’s background psychological profile. Any creature that exploits propositional attitude attributions in order to predict the behavior of another agent will need also to represent explicitly that agent’s background psychological profile. The success of the prediction will ultimately depend upon the accuracy both of the propositional attitude attribution and of the representation of the background psychological profile. Let us call this complex form of mindreading propositional attitude mindreading. Perceptual mindreading is typically far less complex. In many contexts what a creature perceives has obvious and immediate implications for action – seeing a predator, or a food source, for example. The only elements of the background psychological profile that need to be brought into play are generic and universal. In these cases, therefore, there is correspondingly little or no need for the creature explicitly to represent the agent’s background psychological profile. Of course, not all cases of perceptual mindreading are straightforward. Sometimes predicting how an agent will respond to something in plain view requires delving deep into the agent’s psychology. But experiments exploring the mindreading abilities of nonlinguistic creatures tend to lie at the straightforward end of the spectrum. They typically exploit the fact that seeing a food item has immediate implications for action, as in the much-discussed food competition paradigm developed in Hare and Tomasello 2001. For these reasons propositional attitude mindreading is a more complex and sophisticated intellectual activity than perceptual mindreading. It involves explicitly representing elements of an agent’s background psychological profile, and then reasoning about how the agent will act in the light of what is known of their psychology. The complexity of the explicit representation and the scope of the reasoning will vary depending on the particular propositional attitude. But there will always have to be some explicit representation and some reasoning on the part of the propositional attitude mindreader. Perceptual mindreading is not like this. It is perfectly possible for a creature to be a perceptual mind reader without any capacity for explicitly representing an agent’s background psychological profile, and without any capacity to reason about how psychology issues in action. This is because perceptual mindreaders can often exploit and trade on 6 direct connections between perception and action. The holistic character of the propositional attitudes means that there are no comparable direct connections between propositional attitudes and action. 3 The double analogy Something like the following pattern of reasoning is implicit in many discussions of animal minds. (1) (2) (3) Certain species of non-human animals solve many problems of social interaction and coordination that are analogous to problems solved by humans. Humans solve these problems through mindreading strategies. Hence non-human animals also have to be mindreaders. There is a double analogy here. The first analogy is between the types of social situations confronted by human and non-human animals. The second analogy is between the strategies that human and non-human employ to navigate those situations. There are many important questions that might be raised about whether and how arguments from analogy should be used in these contexts. I will prescind from these questions here. What I want to focus on is the basis on which the second analogy is made. Even if one thinks that it is acceptable to reason analogically from the mindreading strategies of human animals to those of nonhuman animals, it is important to start from an accurate picture of humans solve problems of social interaction and social coordination. Many philosophers assume without argument that some version of what is standardly called folk psychology or commonsense psychology is the principal tool that we employ to navigate the social world. This is standardly understood to involve attributing propositional attitudes. In previous work I have expressed skepticism about this assumption. There are many forms of social understanding and social coordination that proceed without attributions of propositional attitudes. One example that I have discussed (Bermúdez 2003, 2005) are highly stereotypical interactions that can be modeled using frames and routines. This is particularly relevant for thinking about mindreading in nonlinguistic creatures. Many of the social interactions that we engage in are highly stereotypical. We negotiate them successfully because we are able to predict what other participants will do. But those predictions need not, and in fact rarely do, involve forms of propositional attitude mindreading. When one goes into a shop or a restaurant, for example, it is obvious that the situation can only be effectively negotiated because one has certain beliefs about why people are doing what they are doing and about how they will continue to behave. I cannot effectively order dinner without interpreting the behavior of the person who approaches me with a pad in his hand, or buy some meat for dinner without interpreting the person standing behind the counter. But these beliefs about what people are doing do not involve second-order beliefs about their psychological states. Ordering meals in restaurants and buying meat in butcher's shops are such routine situations that one need only identify the person approaching the table as a waiter, or the person standing behind the counter as a butcher. Simply identifying social roles provides enough leverage on the situation to allow one to predict the behavior of other participants and to understand why they are behaving as they are. Social understanding and social coordination exploiting shared knowledge of social routines and stereotypes is a form of reasoning. However, this reasoning is similarity-based and analogybased. Social understanding becomes a matter of matching perceived social situations to prototypical social situations and working by analogy from partial similarities. We do not store general principles about how social situations work, but rather have a general template for 7 particular types of situation with parameters that can be adjusted to allow for differences in detail across the members of a particular social category. Research in computer science and artificial intelligence provides one of modeling this type of social reasoning. Motivated in part by the intractability of rule- and logic-based solutions to these everyday social interactions, computer scientists have proposed what are known as frame-based systems (Nebel 1999). Here is Minsky's original articulation of the notion of a frame: Here is the essence of the theory: when one encounters a new situation (or makes a substantial change in one's view of the present problem) one selects from memory a structure called a frame. This is a remembered framework to be adapted to fit reality by changing details as necessary. A frame is a data structure for representing a stereotyped situation, like being in a certain kind of living room, or going to a child's birthday party. Attached to each frame are several kinds of information. Some of this information is about how to use the frame. Some is about what one can expect to happen next. Some is about what to do if those expectations are not confirmed. We can think of a frame as a network of nodes and relations. The top levels of a frame are fixed, and represent things that are always true about the supposed situation. The lower levels have many terminals – slots that must be filled by specific instances or data. Each terminal can specify conditions its assignments must meet. (The assignments themselves are usually smaller sub-frames.) Simple conditions are specified by markers that might require a terminal assignment to be a person, an object of sufficient value, or a pointer to a sub-frame of a certain type. More complex conditions can specify relations among the things assigned to several terminals. (Minsky 1974, pp. 111-112) The frame-based approach gives a concrete example of the form that a routine-based approach to social understanding and social coordination might take. The key point is that, as stressed earlier, the parameters for the frame (what Minsky calls the terminals) need not be propositional attitude attributions. Social interactions that are sufficiently stereotypical to be modeled in terms of frames can proceed without propositional attitude mindreading. Where they do involve mindreading this can simply be perceptual mindreading. Frames and routines provide a framework for interpreting some of the observational and ethological evidence often cited for propositional attitude mindreading in primates. An important part of the case (as reviewed, for example, in Pt II of Tomasello and Call 1997) comes from observation of behaviors in the wild that seem to involve tactical deception (Byrne and Whiten 1990) and/or communication. Leaving aside the methodological issues raised by the analysis of what has seemed to some observers to be rather anecdotal observations, one issue to explore is whether the observed behaviors could not be viewed as stereotypical and patterned interactions where the parameter-setting does not involve one creature forming beliefs about the desires and beliefs of another creature. Consider communicative behaviors, such as the much-discussed alarm calls of vervet monkeys (Cheney and Seyfarth 1990). One of the key features of vervet monkey alarm calls is that they use different types of alarm call in response to the presence of different types of predator – and that monkeys hearing the alarm call respond in different ways to each type of call. This seems to fit very closely the frame model just outlined, with relatively fixed responses (the calls and the behaviors to which they give rise) triggered by different perceptual experiences (playing the role of the terminals in Minsky’s frames). This interpretation of the vervet alarm calls is consistent with many of the claims that have been made about the degree of cognitive sophistication that 8 they reflect. So, for example, it is consistent with the vervet alarm calls carrying information about events in the environment (as opposed to being expressions of the monkey’s state of arousal). And yet it does not involve bringing into play the machinery of propositional attitude mindreading. The routine does not involve one monkey intending to bring it about that the other monkeys believe that a predator is nearby – or that the other monkeys recognize the first monkey’s intention to bring it about that they form this belief. Tactical deception has been less systematically and longitudinally studied than vervet monkey alarm calls. It is certainly possible that what have been interpreted as episodes of primates intentionally manipulating (or attempting to manipulate) the propositional attitudes of conspecifics will turn out to be complex and sophisticated routines of the type analyzed by Minsky. It is plausible that the parameter-setting will involve a degree of mind-reading, but the default assumption (particularly in the light of the considerations that will emerge in the next section) should be that this will be perceptual mindreading. Certainly, many of the reported instances of tactical deception do seem to be interpretable in terms of intentions to manipulate a conspecific’s visual perspective (rather than their propositional attitudes). Consider the following well-known description of an instance of tactical deception in a troop of baboons in Ethopia. An adult female spent 20 min in gradually shifting in a seated position over a distance of about 2m to a place behind a rock about 50 cm high where she began to groom the subadult male follower of the group – an interaction not tolerated by the adult male. As I was observing from a cliff slightly above [the animals] I could judge that the adult male leader could, from his resting position, see the tail, back and crown of the female’s head, but not her front, arms and face: the subadult male sat in a bent position while being groomed, and was also invisible to the leader. The leader could thus see that she was present, but probably not that she groomed. (Report by Hans Kummer quoted in Byrne 1995 p. 106) We can understand what is going on here in terms of a Minsky-style routine. The female baboon is engaging in a complicated pattern of quasi-stereotypical behavior in which the terminals are filled by instances of perceptual mindreading. One terminal is filled by her calculation of the alpha male’s line of sight. Another by her perception of the rock between the alpha male and the subadult male. There is no need to appeal to the female baboon’s intention to manipulate the beliefs of the alpha male. 4 The limits of nonlinguistic mindreading In the previous section I tried to weaken the temptation to think that the complex forms of social interaction and social coordination that we see in the animal kingdom demand explanations in terms of propositional attitude mindreading. In this section I take a more direct tack. I present a version of an argument I first proposed in Bermúdez 2003. The argument aims to show that propositional attitude mindreading is not available to creatures that lack a public language. In order to make the structure of the argument more perspicuous I make each step explicit, adding comments where applicable. (A) Unlike perceptual mindreading, propositional attitude mindreading involves representing another creature’s attitude to a proposition. A representation of another creature as, say, believing that the food is hidden behind the tree is tripartite in nature. It involves representing (i) (ii) a particular individual as bearing a particular propositional attitude to 9 (iii) a particular proposition. Perceptual mindreading is also tripartite in nature. It involves representing (i) (ii) (iii) a particular individual as perceiving a particular object or state of affairs. Despite this similarity in structure, however, the representations required for propositional attitude are far more complex than those required for perceptual mindreading. Consider a perceptual mindreader M representing another agent α as perceiving a state of affairs S. The perceptual mindreader is already perceiving S. In order to represent α as perceiving S, M needs simply to add to its representation of S a representation of a relation between α and S. As we saw in the previous section, in many cases of perceptual mindreading this additional representation can be very straightforward. It can be simply a matter of representing S as lying in α’s line of sight. The representational skills required are basic geometric skills, on a par with those involved in working out possible trajectories of objects (including the agent’s own body) through the environment. As we saw in section 2, moreover, the fact that S is in α’s line of sight can often have very immediate implications for action. This means that moving from a representation of α as perceiving S to a prediction of how α will behave is often completely straightforward. Now consider a propositional attitude mindreader M* representing another agent β as believing a proposition P. Here it is not typically the case that P corresponds to a state of affairs in the distal environment that M* is already perceiving. In fact (as we saw in the last section), in many of the case where propositional attitude mindreading is identified in the animal kingdom P is actually false. The aim in tactical deception as standardly interpreted, for example, is to generate false beliefs in another agent – and hence where the deceiver must intend to bring it about that an agent believe that p where p is false. So, the question arises: What is it to represent a proposition (particularly one that one does not oneself believe)? One answer here is that propositions just are states of affairs, and so representing a proposition is no more and no less complicated than representing states of affairs. On this interpretation propositional attitude mindreading does not come out as fundamentally different in kind from perceptual mindreading. It is true that propositional attitude mindreading can involve representing states of affairs that do not exist (as in tactical deception cases), but it is widely accepted that many types of non-human animals can represent non-existent states of affairs. After all, we can only explain the behavior of non-linguistic creatures in psychological terms if we attribute to them desires, and having a desire often involves representing a non-existent state of affairs. The obvious problem with this view is that states of affairs lack some of the fundamental characteristics of propositions. In particular, propositions are true or false, while states of affairs are not the sort of things that can be either true or false. On many standard ways of thinking about propositions and states of affairs, states of affairs are the things that make propositions true or false – precisely because propositions are representations of states of affairs that can be true or false.3 3 Things are not quite as simple as this, since propositions can be logically complex and it is not clear that there are logically complex states of affairs. Even if one thinks that the proposition that the table is red is made true by state of affairs of the table being red, it is far from obvious that the proposition the table is not red is made true by the state of affairs of the table not being red. Many philosophers would deny that 10 The “truth-aptness” of propositions is absolutely fundamental to the whole enterprise of propositional attitude mindreading. Propositional attitude mindreading allows us to explain and predict the behavior of other subjects in terms of the representational states that generated it. It is such a powerful tool because it works both when other subjects represent the world correctly, and when they misrepresent it. Moreover (as will become very important in the following) the truth-aptness of propositions is what explains the inferential connections between propositional attitudes that are not belief-like. Desires, for example, are not true or false in the way that beliefs are. But desires have contents (propositions) that stand in logical relations to the contents of desires and other beliefs. These logical relations are constantly exploited in practical reasoning. In sum, my claim is that propositional attitude mindreading involves representing another agent’s representation of a state of affairs. I will use the term “proposition” to abbreviate “representation of a state of affairs”. Anyone who thinks about propositions in a different way is invited to use the unabbreviated expression. (B) Propositional attitude mindreaders must represent propositions in a way that allows them to work out how the relevant propositional attitudes will feed into action. This should not be controversial in the light of discussion earlier in the paper. Propositional attitude mindreading is a way of explaining and predicting behavior. Obviously it is not enough simply to represent the propositional attitudes of another agent. Those propositional attitudes must be represented in a format (what is sometimes called a vehicle) that can be exploited in reasoning that leads to explanation and prediction. (C) Working out how a particular set of propositional attitudes will feed into action depends upon the inferential relations between those propositional attitudes and the agent’s background psychological profile. This is an immediate consequence of the earlier discussion in section 2. Propositional attitudes do not feed directly into action in the way that many perceptual states do. Exploiting propositional attitude attributions to explain and predict behavior requires working out the different possible inferential connections between those propositional attitudes and the agent’s background psychological profile, as well as the information that they have through perception about the distal environment. A propositional attitude mindreader must be able to represent propositions in a way that makes clear the logical and inferential relations between th attributed propositional attitudes, the assumed background psychological profile, and the anticipated action. (D) The representations exploited in propositional attitude mindreading are consciously accessible constituents of a creature’s psychological life. When a propositional attitude mindreader forms beliefs about the mental states of another agent, those beliefs are integrated with rest of the mindreader’s propositional attitudes – with their beliefs about the distal environment, with their short-term and long-term goals, for example. This integration is required for the results of propositional attitude mindreading to feature in a creature’s practical decision-making. And, since practical decision-making takes place at the conscious level, beliefs about the mental states of other agents must also be consciously accessible. (E) The representational format for propositional attitude mindreading must be either language-like or image-like. negative states of affairs exist. Nonetheless, the fact remains that when the table is not red is true, its truth consists in the holding of some state of affairs (the table being black, for example). 11 This basic distinction here can be put in a number of different ways – as the distinction between digital representations and analog representations, for example. It is a very basic distinction for cognitive science, as reflected for example in what has come to be known as the imagery debate.4 The central idea is straightforward, although there are many different ways of working out the details. A language-like representational format allows complex representations to be built up in a rule-governed way from basic representational units. The representational units and complex representations built up from them function as symbols, with no intrinsic connection to what they represent. An image-like representational format, in contrast, functions as a picture. It is not built up from basic representational units and it represents through similarity relations between the structure of the representation and the structure of what is being represented. As far as propositional attitude mindreading is concerned, the most obvious candidates for an image-like representational format comprise mental models theory in the psychology of reasoning (originally proposed in Craik 1943 but most comprehensively developed in JohnsonLaird 1983) and the conception of mental maps put forward by Braddon-Mitchell and Jackson (Braddon-Mitchell and Jackson 1996). The idea of structural isomorphism is key to both approaches. Mental maps and models are isomorphic to what they represent, in the sense that the relations holding between elements in the map/model can be mapped onto relations holding between elements in what is represented. (F) The representational format for propositional attitude mindreading must exemplify the structure of the represented propositions. This follows from (B) and (C) above. Propositional attitude mindreading allows mindreaders to navigate the social world by making sense of and predicting the behavior of other agents. Attributions of propositional attitudes do not lead immediately to predictions and explanations – precisely because propositional attitudes themselves do not themselves feed directly into action. So, a propositional attitude mindreader needs to be able to work out the implications of the attributed propositional attitudes for the agent’s behavior, in the light of what is known (or conjectured) about the agent’s background psychological profile. This is a matter of reasoning about the logical and inferential relations between propositional attitudes. Accurate predictions depend upon the predictor being able in some sense to track the reasoning that the agent might themselves engage in. This means that propositions (and hence, of course, propositional attitudes) must be represented in a way that allows the mindreader to reason about the logical and inferential connections between propositions. In particular, they must be represented in a way that makes clear the structure of the relevant propositions. This is because many of the most basic logical and inferential relations between propositions hold in virtue of their structure. Consider, for example, the most fundamental form of logical thinking – that codified in the propositional calculus and involving the basic logical connectives, such as disjunction, conjunction, and the material conditional (if. . . then. . .). It is hard to see how a mindreader could reason about the propositional attitudes and background psychological profile of another agent without attributing to them conditional beliefs. These conditional beliefs might, for example, dictate possible behaviors contingent upon particular environmental factors (e.g. if the prey goes into the woods then I will follow it). Or they might record regularities and contingencies in the environment (e.g. if it rains then there will be more insects on the leaves). In order to reason about how an individual with these conditional beliefs might behave they must be represented in a way that reflects their structure – in a way that reflects the fact that they are attitudes to a complex proposition that relates two other propositions. Unless the conditional belief is 4 For classic discussions see Anderson 1978 and Pylyshyn 1981. Kosslyn et al. 2006 is a more recent contribution. 12 represented in this way it will be impossible to reason, for example, that since the agent has seen the prey going into the woods it will follow it. (G) Representing propositions in a pictorial or image-like format does not reveal their canonical structure. It is certainly true that images have a type of structure. So do mental models and maps. We can identify distinct parts in analog representations and indeed identify them across different representations. Without this it would be impossible for analog representations to represent in virtue of their isomorphic structural resemblance to the state of affairs that they represent. But this is not the right sort of structure for them to represent propositions in the way that (F) requires. Braddon-Mitchell and Jackson themselves bring out both how imagistic representations can be structured, and why that is not enough to represent propositions in a manner that will allow them to feature in inferences. There is no natural way of dividing a map at its truth-assessable representational joints. Each part of a map contributes to the representational content of the whole map, in the sense that had that part of the map been different, the representational content of the whole would have been different. Change the bit of the map of the United States between New York and Boston, and you change systematically what the map says. This is part of what makes it true that the map is structured. However, there is no preferred way of dividing the map into basic representational units. There are many jigsaw puzzles you might make out of the map, but no single one would have a claim to have pieces that were all and only the most basic units. (BraddonMitchell and Jackson, p.171) As this very helpful formulation brings out, analog representations do not have a canonical structure. Their structure can be analyzed in many different ways (corresponding to many different jigsaw puzzles that one can construct from it), but none of these can properly be described as giving the structure of the map. This is why maps (and other analog representations) are not well-suited to represent propositions. Propositions have what might be a termed a canonical structure and in order to understand how the inferential connections in which a proposition might stand a thinker needs to understand that canonical structure. The canonical structure of a proposition corresponds to what BraddonMitchell and Jackson describe as the “preferred way of dividing” the proposition into basic representational units. The canonical structure of a conditional proposition, for example, is ‘If A then B’, where ‘A’ and B’ are the basic representational units (in this case the basic representational units are themselves propositions). But it is clear that this canonical structure cannot be captured in any sort of analog representation. We have no idea what a conditional map might look like, for example. (H) The canonical structure of a proposition is only revealed when propositions are represented in a linguistic format. According to (E), propositions must be represented in either an image-like or a language-like representational format. From (F) we have that an image-like representation of a proposition cannot reveal its canonical structure. So, (H) will follow provided that language-like representations of propositions can reveal their canonical structure. But it is obvious that the canonical structure of a proposition is revealed when it is represented linguistically. Viewed in the abstract, language is a mechanism for creating complex representations through the combination of basic representational units according to independently identifiable combinatorial rules. Language contains markers (such as the logical connectives) corresponding to the basic inferential 13 connections between propositions. Indeed, for many philosophers it is almost a tautology that sentences express propositions. (I) The linguistic representations required for propositional mindreading must involve natural language sentences. (H) tells us that propositions must be represented in a linguistic format for propositional mindreading. But it does not say anything about that linguistic format, beyond that it must be capable of revealing the structure of a proposition. There are two different candidate formats. On the one hand, the vehicles for propositional attitude mindreading might be the sentences of a public language. Or, on the other, they might be the sentences of what is sometimes termed the language of thought. The language of thought is proposed as a representational format for certain types of cognitive information-processing (Fodor 1975). It is a key element in what is sometimes called the computational or representational model of the mind. There is no need, however, to explore the details of the arguments for and against the language of thought hypothesis and the computational model of the mind (for further discussion see Bermúdez 2005 and Bermúdez forthcoming). The important point is that the language of thought hypothesis is an explanatory hypothesis in cognitive science. It is a hypothesis about the machinery of subpersonal information processing. Information-processing that exploits sentences in the language of thought takes place below the threshold of consciousness. This means that sentences in the language of thought cannot be consciously accessible constituents of a creature’s psychological life. But we saw in (D) that the representations exploited in propositional attitude mindreading must be consciously accessible constituents of the mindreader’s psychological life. They must be capable of featuring in a creature’s conscious practical decision-making. 5 Conclusion This paper has introduced two fundamental distinctions to be kept clearly in view when thinking about mindreading in the animal kingdom. The first is between minimal mindreading and substantive mindreading. Minimal mindreading occurs in a social interaction when a creature’s behavior depends systematically upon changes in the psychological states of other participants in the interaction. In contrast, a creature engages in substantive mindreading when its behavior is systematically dependent upon how it represents the psychological states of other participants. Substantive mindreading is not a unitary phenomenon. There is a principled distinction between perceptual mindreading and propositional attitude mindreading, stemming from the holistic character of the propositional attitudes. Using attributions of propositional attitudes to predict and explain behavior involves complex forms of reasoning that exploit information about the agent’s background psychological profile. The principal aim of this paper is to argue that propositional attitude mindreading does not and cannot exist in the absence of language. The temptation to identify propositional attitude mindreading in the animal kingdom often rests on the tacit assumption that most if not all of the complexities of human social understanding and social coordination depend upon propositional attitude mindreading. We looked at an important type of interaction that can be modeled without making this assumption, and considered how that model might be applied to behaviors in the animal kingdom that are sometimes taken as evidence for propositional attitude mindreading. Finally, I developed an argument to the effect that propositional attitude mindreading cannot exist in the absence of language. If this argument is sound, it shows that perceptual mindreading is the only form of substantive mindreading that can exist in the (non-human) animal kingdom. References Allen, C., & Bekoff, M. (1997). Species of Mind: The Philosophy and Biology of Cognitive 14 Ethology. Cambridge MA: MIT Press. Anderson, J. R. (1978). Arguments concerning representations for mental imagery. Psychological Review, 85, 249-277. Axelrod, R. (1984). The Evolution of Coperation. New York: Basic Books. Bermudez, J. L. (2003). Thinking without Words. New York: Oxford University Press. Bermudez, J. L. (2005). Philosophy of Psychology: A Contemporary Introduction. London: Routledge, Bermudez, J. L. (Forthcoming). Cognitive Science: An Introduction to the Science of the Mind. Cambridge: Cambridge University Press. Braddon-Mitchell, D., & Jackson, F. (1996). Philosophy of Mind and Cognition. Oxford: Blackwell. Byrne, R. W., Whiten, A. (Ed.). (1988). Machiavellian Intelligence. Oxford: Oxford University Press. Byrne, R. W., & Whiten, A. (1990). Tactical deception in primates: The 1990 database. Primate Report, 27(1-101). Byrne, R. W., & Whiten, A. (1991). Computation and mindreading in primate tactical deception. In A. Whiten (Ed.), Natural Theories of Mind: Evolution, Development, and Simulation of Everyday Mindreading. Oxford: Blackwell. Carruthers, P., & Smith, P. K. (Eds.). (1996). Theories of Theory of Mind. Cambridge: Cambridge University Press. Cheney, D. L., & Seyfarth, R. M. (1995). How Monkeys See The World. Chicago: University of Chicago Press. Craik, K. (1943). The Nature of Explanation. Cambridge: Cambridge University Press. Dally, J. M., Emery, N. J., & Clayton, N. S. (2006). Food-caching western scrub jays keep track of who was watching when. Science, 312, 1662-1665. Davidson, D. (1975). Thought and talk. In S. Guttenplan (Ed.), Mind and Language. Oxford: Oxford University Press. Davies, M., & Stone, T. (Eds.). (1995). Folk Psychology: The Theory of Mind Debate. Oxford: Basil Blackwell. Dennett, D. (1969). Content and Consciousness. London: Routledge Kegan Paul. Fodor, J. (1975). The Language of Thought. Cambridge MA: Harvard University Press. Hare, B., Brown, M., Williamson, C., & Tomasello, M. (2002). The domestication of social cognition in dogs. Science, 298, 1634-1636. Hare, B., & Tomasello, M. (2001). Do chimpanzees know what conspecifics know? Animal Behavior, 61, 139-151. Heil, J. (1982). The Nature of True Minds. Cambridge: Cambridge University Press. Heyes, C. M. (1998). Theory of mind in nonhuman primates. Behavioral and Brain Sciences, 21, 101-134. Hurley, S., & Nudds, M. (2006). Rational Animals? Oxford: Oxford University Press. Johnson-Laird, P. (1983). Mental Models. Cambridge: Cambridge University Press. Kosslyn, S. M., Thompson, W. L., & Ganis, G. (2006). The Case for Mental Imagery. Oxford: Oxford University Press. Maynard Smith, J. (1982). Evolution and the Theory of Games. Cambridge: Cambridge University Press. Nowak, M. A., Sasaki, A., Taylor, C., & Fudenberg, D. (2004). Emergence of cooperation and evolutionary stability in finite populations. Nature, 428, 646-650. 15 Penn, D. C., & Povinelli, D. (2007). On the lack of evidence that non-human animals possess anything remotely resembling a 'theory of mind'. Philosophical Transactions of the Royal Society B, 362, 731-744. Povinelli, D., & Vonk, J. (2006). We don't need a microscope to explore the chimpanzee's mind. In S. Hurley & M. Nudds (Eds.), Rational Animals? Oxford: Oxford University Press. Premack, D., Woodruff, G. (1978). Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(515-526). Pylyshyn, Z. (1981). The imagery debate: Analogue media versus tacit knowledge. Psychological Review, 88, 16-45. Tomasello, M., & Call, J. (1997). Primate Cognition. Oxford Oxford University Press. Tomasello, M., & Call, J. (2006). Do chimpanzees know what others are seeing - or only what they are looking at? In S. Hurley & M. Nudds (Eds.), Rational Animals? Oxford: Oxford University Press. Tschudin, A. (2001). 'Mind-reading' mammals: Attribution of belief tasks with dolphins. Animal Welfare, 10, S119-127. Wimmer, H., & Perner, J. (1983). Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition, 13, 103-128.