The Folk Psychology of Free Will: Fits and Starts

According to agent-causal accounts of free will, agents have the capacity to cause actions, and for a given action, an agent could have done otherwise. This paper uses existing results and presents experimental evidence to argue that young children deploy a notion of agent-causation. If young children do have such a notion, however, it remains quite unclear how they acquire it. Several possible acquisition stories are canvassed, including the possibility that the notion of agent-causation develops from a prior notion of obligation. Finally, the paper sets out how this work might illuminate the philosophical problem of free will. The language of all mankind, and their ordinary conduct in life, demonstrate, that they have a conviction of some active power in themselves to produce certain motions in their own and in other bodies, and to regulate and direct their own thoughts. This conviction we have so early in life, that we have no remembrance when, or in what way we acquired it (Reid, 1969 [1788], p. 269). The problem of free will has its roots in commonsense. No scientific expertise is required to bring the problem to life for the uninitiated. For the notion of free will is a commonsense notion, a part of our folk psychology. Over the last quarter century, philosophical work on the notion of free will has flourished (see Kane, 2002 for state-of-the-art overviews). Much of this work is devoted to giving an analysis of the notion of free will that fits with the intuitions of philosophical grown-ups. While this analytic literature on free will has become wildly sophisticated, naturalistic philosophers and developmental psychologists have produced a highly developed literature on the child’s emerging capacity for predicting and explaining people’s behavior, ‘mindreading’ (see, e.g. Goldman, 1989; Gopnik and Wellman, 1994; Gordon, 1986; Harris, 1992; Nichols and Stich, 2003; Perner, 1991; Wellman, 1990). Unfortunately, although the notion of free will is indisputably part of folk psychology, the notion of free will has been almost entirely neglected in this naturalistic literature on folk psychology in children. This is especially disappointing for those of us who regard the developmental work as a I would like to thank the teachers, parents, and children of the N.E. Miles Early Childhood Development Center at the College of Charleston. Thanks to Hilary Martin, Sara Martell, and Emily Askey for assistance in data collection. Many thanks to Trisha Folds-Bennett for help getting subjects and pilot data for experiment 2. I would also like to thank John Doris, Larry Krasnoff, Heidi Maibom, Eddy Nahmias, Philip Robbins, Eric Schwitzgebel, two anonymous referees and an editor at Mind & Language for helpful comments on an earlier draft. Address for correspondence: Department of Philosophy, College of Charleston, Charleston, SC 29424, USA. Email: [email protected] Mind & Language, Vol. 19 No. 5 November 2004, pp. 473–502. #BlackwellPublishingLtd. 2004, 9600GarsingtonRoad,Oxford,OX42DQ,UKand350MainStreet,Malden,MA02148,USA. promising source for philosophical illumination. In this paper, I attempt to forge some links between these literatures. Since the issue has not been systematically joined previously, the paper will be rather exploratory and programmatic. I will argue that young children have a notion of agent-causation, according to which (i) actions are caused by agents and (ii) for a given action, an agent could have done otherwise. In the first section, I’ll set out the notion of agent-causation as it has been developed in philosophy. Then I’ll briefly discuss the absence of this notion in standard accounts of the child’s understanding of decision making. In section 3, I’ll discuss some evidence and present some experimental results that support the claim that children exploit a notion of agent-causation. In section 4, I’ll set out the difficult question of acquisition. If children do have a notion of agent-causation, how do they acquire such a thing? I’ll describe and discuss several possible acquisition stories. Finally, in the fifth section, I’ll consider possible implications for the philosophical debate on free will. 1. The Philosophical Characterization of Agent-Causation In the philosophical literature on free will, agent-causal theories have a long, if somewhat irregular, history. One of the earliest accounts that is unequivocally agent-causal comes from Thomas Reid (1969 [1788]). According to William Rowe, the foremost contemporary exegete of Reid’s view, on Reid’s account you count as a true agent-cause of a change in the world when ‘you had the power to bring about that change, you exerted that power by acting, and finally, you had the power not to bring about that change’ (1989, p. 159). Such agent-causal theories are typically framed against the thesis of determinism, the claim that every event is an inevitable consequence of the prior conditions and the natural laws. Historically, agent-causal theorists have had to fight on two very different fronts. On the one hand, agent-causal theorists are in combat with ‘compatibilists’ who claim that free will and determinism are perfectly consistent once one gets clear about the proper interpretation of ‘free will’. The proper understanding of ‘free will’, according to agent-causal theorists, is one that is thoroughly at odds with determinism. On the second front, agent-causal theorists resist ‘hard determinists’, who maintain that while free will is indeed at odds with determinism, that means that free will doesn’t exist, since determinism is true. Agent-causal theorists maintain rather that it is a fact that we have free will and that this entails that determinism is false, at least when it comes to agents. For most of the Twentieth century, agent-causation had little cachet in analytic philosophy. In the mid-Twentieth century, the view was bravely resurrected by 1 As a result, agent-causal theories count as ‘libertarian’ approaches to free will. However, some libertarian accounts reject agent-causation (e.g. Ginet, 1990; Kane, 1996). 474 S. Nichols # Blackwell Publishing Ltd. 2004 Roderick Chisholm (1966) and a few others, in the face of rather a good deal of abuse. But recently agent-causation has undergone a renaissance. There are a variety of agent-causal theories, and there has been a resurgence of work on these kinds of accounts (see, e.g. Clarke, 1993, 1996, 2003; Ekstrom, 2000; O’Connor, 1995, 2000; Rowe, 1989, 1991). Despite the important differences between different accounts of agent-causation, all agent-causal theorists agree on two points: 1. An agent is a causal factor in the production of an action. 2. For a given action of an agent, the agent could have not caused it. Roughly, the agent could have done otherwise. The first condition has it that the power to make the change, to produce the action, lies in the agent. The important feature here is that agent-causal theories quantify over agents, and they locate in the agent causal powers. However, this first condition alone will not rescue agents from the hegemony of determinism. The problem is that just because the agent caused something to happen doesn’t mean that the outcome wasn’t inevitable. It doesn’t mean that the agent could have refrained from the action. Rowe makes the point by adverting to an example of causal powers outside the realm of agency: Suppose a piece of zinc is dropped into some acid, and the acid dissolves the zinc. In this example, we might say that the acid has the power to bring about a certain change in the zinc. . . . But can we reasonably say that the acid had the power not to bring about this change? Clearly we cannot. The acid has no power to refrain from dissolving the zinc . . . The acid, therefore, is not an agent-cause of the zinc’s dissolving (Rowe, 1989, p. 159). The acid has the power to dissolve the zinc, but it does not have the power not to dissolve the zinc. One might similarly claim then that an agent was a causal factor in producing an action without maintaining that the agent could have done otherwise than produce the action. Hence, to capture the relevant anti-determinist notion of agent causation, we need to add the condition that the agent could have done otherwise, even while all the other factors were exactly the same. The interposed hyphen produces the technical term for this notion, agentcausation. Agent-causal theories attract skeptics of all sorts. Perhaps the deepest reservation is that agent-causation is unintelligible. For an agent-caused action is produced in a way that is not deterministic, but neither can the action be produced in a random fashion. And that, the worry goes, leaves us fresh out of possibilities. I will prescind 2 In this paper, I intend for the ‘could have done otherwise’ condition to be understood as indeterminist, unless otherwise noted. The Folk Psychology of Free Will 475 # Blackwell Publishing Ltd. 2004 from such questions about whether agent-causation has any remote plausibility. I should probably confess that I am deeply skeptical that we are agent-causes. But here I won’t assume or argue that agent-causal theories are wrong. I want to remain entirely neutral on the actual metaphysics of agency. Until section 5, my focus will be on two questions: i. Do children believe in agent-causation? ii. How might the belief in agent-causation be acquired? These questions would not be settled by determining whether or not agentcausation obtains. Children (and adults) believe in all sorts of crazy things, and they fail to believe in all sorts of true things. So we could hardly settle whether children believe in agent-causation by settling the actual metaphysics. Similarly, the belief in agent-causation would demand an acquisition story regardless of whether agent-causation is real or not. For even if it’s real, we need a story about how children come to recognize this. Even if we are agent-causes, that does not explain why young children believe that we’re agent-causes. 2. Mindreading and Agent-Causation As noted in the introduction, there has been a great deal of work on the child’s capacity for mindreading. The research shows that by the age of four, children are remarkably adept at predicting and explaining behavior (e.g. Gopnik and Meltzoff, 1997). This presumably requires a facility at explaining and predicting decisions. Researchers on mindreading have provided increasingly detailed models of children’s folk psychological reasoning. So one might expect that this literature is the best place to look to determine whether children believe in agentcausation. As it happens, on standard accounts of mindreading, agent-causal notions are never invoked. Indeed, models of folk psychology in the mindreading literature make no reference to free will as an element of the lay theory of decision-making. Gopnik and Wellman, for instance, characterize the young child as deploying a version of the practical syllogism: ‘If an agent desires x, and sees that x exists, he will do things to get x.’ Even that form of the practical syllogism is a powerful inferential folk psychological law. It allows children to infer, for example, that if John wants a cookie and sees one in the cookie jar, he will go there for it. If he doesn’t want it, or doesn’t see it, he won’t (Gopnik and Wellman, 1994, p. 265; see also Wellman and Bartsch, 1988; Wellman, 1990). In our recent work, Steve Stich and I dispute this picture of early mindreading, but we do not invoke agent-causal notions either. Rather, to explain how children 476 S. Nichols # Blackwell Publishing Ltd. 2004 predict an agent’s decisions, we advert to a ‘Mindreading Coordinator’ mechanism. This mechanism works as follows on our theory: When a prediction about the future behavior of a target is required, the Coordinator’s first job is to assemble information about the target’s goals or desires. . . . After this is done, the Coordinator collects whatever information it can find in the Belief Box about the target’s desires and sends a call to the Planner [the mechanism for constructing plans]. The Planner’s charge is to come up with the best plan for satisfying those desires for an agent in the target’s situation. When the Planner reports back with a plan of action, the Coordinator then generates a belief that the target will try to act in accordance with that plan. And it is that belief that is used to predict the target’s behavior (Nichols and Stich, 2003, p. 81). So, on our proposal, the Mindreading Coordinator begins with information about the agent’s goals and desires and generates a prediction that the agent will try to act in accordance with the best plan for satisfying his desires. Thus, the Mindreading Coordinator generates a prediction, but it does not generate an attribution of agent-causation. Similarly, Alan Leslie, who occupies yet another view on the theoretical landscape, does not advert to agent-causal notions as part of the lay view of agency. Leslie maintains that evolution has built us to track three features of agents: Mechanical properties (e.g. ‘having an internal and renewable source of energy’), Actional properties (‘Agents act in pursuit of goals and re-act to the environment as a result of perceiving’), and Cognitive properties (‘The behavior of Agents is determined by cognitive properties, e.g. holding a certain attitude to the truth of a proposition’). On Leslie’s view, these three features of agents are tracked by three different modules that deliver our ‘core notions of agency’ (Leslie, 1995). Leslie never suggests either that agent-causation is a genuine feature of agents or that agent-causation is a part of our core notions of agency. Hence, mindreading theorists with very different allegiances share widely in the omission of agent-causal notions from accounts of the lay prediction of behavior. How are we to interpret the evident omission of agent-causation from accounts of mindreading? One possibility is that agent-causation is a wrongly neglected aspect of the mindreading system. Perhaps Gopnik, Wellman, Leslie, Stich, and I have all overlooked a central aspect of the mindreading system, of the mechanisms underlying the child’s prediction of behavior. Another possibility, however, is that the notion of agent-causation is not implicated in the mindreading system—in the domain of folk prediction and explanation, the notion of agent-causation plays no useful role, so the mindreading accounts are right to neglect it. An even more skeptical possibility looms. Perhaps the notion of agent-causation has no part in the child’s (or the adult’s) view of the mind. Perhaps agent-causation is an invention of philosophers that fails entirely to connect with the child’s view of decision making. At this point, we arrive at a resolutely empirical issue—do children have a notion of agent-causation? The Folk Psychology of Free Will 477 # Blackwell Publishing Ltd. 2004 3. Do Children Believe in Agent-Causation? As we saw in section 1, there are two crucial claims to an agent-causal theory. First, actions are caused by agents. Second, actions are not inevitable—for a given action the agent could have caused a different action. There is reason to think that children embrace both claims, but making the case for this requires some care. I’ll start by arguing that children regard agents as having causal powers to produce actions. Then I’ll present some experimental work indicating that children accept that agents could have done otherwise. 3.1. Agents Are Causes Recent work on the understanding of agency in infants and toddlers provides the basis for thinking that young children regard agents as having causal powers. The primary evidence here comes from developmental psychologists, of course. But to make the case for causal powers, we will need to recruit some recent work on causal attribution as well. The goal in this section is to argue that very young children accept the following: The Causal Principle: An agent is a causal factor in the production of an action; briefly, agents have causal powers to produce actions. First, I’ll rely on developmental evidence to argue that very young children accept the following: The Correlation Principle: If there is an action, there is an agent. Then I’ll argue that it’s plausible that children who accept the Correlation Principle will also accept the Causal Principle. 3.1.1. Developmental Evidence. The available evidence from developmental psychology indicates that infants embrace the Correlation Principle. The first important claim to establish is that infants quantify over agents. As with everything in this paper, the existing data are more suggestive than definitive, but recent work on agency detection provides some reason to think that infants, at least in the second year, do quantify over agents. To begin to see this, we can look to the work of Susan Johnson and her colleagues. In an elegant experiment, 12-month-old infants were shown a fuzzy brown object under a variety of different conditions 3 Most accounts of children’s mindreading probably do maintain that children think that actions are caused by agents. It is the second claim of agent-causal theories that isn’t reflected in contemporary accounts of mindreading. However, since the notion of agent-causation encompasses both claims, it will be important to consider directly whether children embrace both claims. 478 S. Nichols # Blackwell Publishing Ltd. 2004 (Johnson, Slaughter and Carey, 1998). In one condition, the fuzzy brown object had eyes, in another condition, the fuzzy brown object was eyeless but interacted contingently with the infant (by beeping and flashing lights when the infant babbled or moved), and in another condition, the fuzzy brown object had neither eyes nor did it interact contingently. In all conditions, children’s looking behavior was measured when the fuzzy brown object ‘gazed’ at one of two objects by making a smooth, 45 degree turn towards the object and remaining in this orientation for several seconds. What Johnson and colleagues found was that infants would follow the ‘gaze’ of the fuzzy brown object when it had eyes or when it interacted contingently, but infants were less likely to follow the ‘gaze’ of the fuzzy brown object if it had neither of these features. Johnson proposes that what happened in the experiment is that the infants followed the gaze when the fuzzy brown object was coded as an agent, an entity that has mental states (Johnson, 2000, p. 22). For gaze-following is often taken to reflect the ‘implicit attribution of a mind to the gazer’ (Johnson et al., 1998, p. 233), and the experiments of Johnson and colleagues controlled to exclude many other deflationary interpretations of gaze-following in their tasks (1998, p. 237). Further, they note that experiments on adults using the same fuzzy brown object yielded complementary results. Adults described the fuzzy brown object’s turning behavior in mentalistic terms in the same conditions that generated gaze-following in infants (1998, p. 237). So either having a face or interacting contingently apparently triggers a representation of an object as an agent. Further support for the idea that infants code certain objects as agents comes from evidence of diverse behavioral responses to agents. In follow-up work, Johnson, Booth and O’Hearn (2001) explored the reactions of 15-month-old infants to a stuffed orangutan doll that had eyes and exhibited contingent interaction. They found that the infants made communicative gestures to the orangutan, but not to a grossly similar object that lacked eyes and contingent interaction (Johnson et al., 2001, p. 652). Hence, there is some evidence that infants do represent certain objects as agents, insofar as they systematically treat agents distinctively in gaze-following and communication. Furthermore, the work of Johnson and colleagues suggests that the triggers for the representation of agency include having a face and exhibiting contingent interaction. Apparently, then, infants do quantify over agents. Recall the Correlation Principle, which states that if there’s an action, there’s an agent. One way to explore whether children accept this principle is by investigating whether children think that if there is no agent, there is no action. Using an ingenious imitation methodology, Meltzoff found that infants attributed goals to people but not to a mechanical device. In Meltzoff’s experiment, 18-month-old 4 Agent-causal theorists often have a stronger notion of agent in mind. Reid, for instance, regarded agents as immaterial substances. The evidence on agency detection doesn’t reveal whether or not children share this view of the ontology of agency. And the discussion to follow focuses on a weaker notion of agency according to which agents are entities that have mental states. The Folk Psychology of Free Will 479 # Blackwell Publishing Ltd. 2004 infants observed an adult ‘failing’ to carry out a goal. For instance, in one of the tasks infants observed an adult trying but failing to pull apart a small dumbbell. Infants pulled apart the dumbbell when they watched a person fail, but not when they watched a mechanical contraption fail. According to Meltzoff, the results from this experiment ‘demonstrate that physical movements performed by a machine are not ascribed the same meaning as when performed by a person. Therefore, even a weak reading of the data suggests that infants are thinking in terms of goals that are connected to people and not to things’ (1995, p. 848). For our purposes, the results suggest the following: the infants do not interpret the machine’s behavior as an action because they do not categorize the machine as an agent. Meltzoff’s study suggests that infants think that if there is no agent, there’s no action (see also Woodward, 1998), and this supports the claim that infants embrace the Correlation Principle. Further evidence for this claim might be drawn out of the studies by Johnson and her colleagues. As noted above, Johnson and colleagues (1998) found that the infants would follow the ‘gaze’ of the eyeless fuzzy brown object, but only if it interacted contingently with the infants. If the (eyeless) fuzzy brown object did not exhibit such contingent interaction, the infants would not follow its ‘gaze’. One implication of this work is that non-human objects are not categorically excluded from being agents. More importantly for our purposes, this provides further evidence that children accept the Correlation Principle that if there is an action there is an agent. The infants evidently inferred that the fuzzy brown object was an agent from observing the contingent interaction of the fuzzy brown object. A natural interpretation of this is that the infant regarded the contingent interaction as action, from which they inferred that there was an agent. That is, the results suggest that babies are inferring that the contingent behavior counts as action and hence, the fuzzy brown object is an agent. Together, the results from Meltzoff and Johnson and colleagues provide promising initial evidence that young children do indeed accept the Correlation Principle that if there is an action, there is an agent. That does not immediately generate the result that children accept the Causal Principle that agents are causes. But now we might appeal to work on causal attribution to build a bridge. 3.1.2. Causal Attribution. The work on causal attribution that I’ll invoke is, not surprisingly, quite deliberately non-Humean. Humean accounts of causal attribution eschew causal powers altogether, so it makes no sense on a Humean account of causal attribution to argue that children attribute causal powers to agents or anything else. There is, by now, an important body of empirical work 5 Indeed, much of the motivation behind the work of Johnson and colleagues is to undermine the claim that infants will only regard humans as agents (Johnson et al., 2001). In the orangutan-doll studies mentioned above, one task is specifically designed to parallel Meltzoff’s imitation study, but with an object that is manifestly non-human. Using Meltzoff’s method, Johnson and colleagues found that infants would imitate apparently goal-directed actions of the orangutan-doll. 480 S. Nichols # Blackwell Publishing Ltd. 2004 supporting the claim that lay views of causation are decidedly non-Humean (e.g. Ahn and Bailenson, 1996; Bullock et al., 1982; Cheng, 1997; Michotte, 1963 [1946]). Since I’m explicitly trying to see whether children attribute causal powers to agents, I will draw from this non-Humean tradition. The notion of causal power used in this literature is summarized as follows by Patricia Cheng: ‘causal power . . . is the intuitive notion that one thing causes another by virtue of the power or energy that it exerts over the other’ (Cheng, 1997, p. 368). Elsewhere, she writes, ‘entities and events may have causal powers with respect to other entities or events’ (Cheng, 1999, p. 227). Hence, the question for us is whether we can use the fact that children accept the Correlation Principle to infer that children regard agents as entities that have causal powers with respect to actions. One possibility, of course, is that the Correlation Principle is the product of a prior belief in a version of the Causal Principle. That is, it might be that in the experiments reviewed above, the babies already believe that only agents have causal powers to produce actions, and it is this belief that underlies their responses in the experiments. Indeed, one obvious possibility is that the Causal Principle is innate, that babies are pre-wired to believe that agents are causes. Even if babies don’t have a prior belief in the Causal Principle, there’s reason to think that children would likely infer the Causal Principle from the Correlation Principle. On the prevailing account of causal induction, regular covariation is used as evidence for causal induction. There are exceptions to this rule, of course. Sometimes there are spurious correlations. There is currently a flurry of work that tries to build a more tightly constrained account of how correlational information is used to infer causal powers. For instance, Cheng offers an influential account for how subjects estimate causal powers by consulting contingency information (Cheng, 1997, 1999; see also Glymour, 2001). But on all of the prominent accounts (including Cheng’s), it is a presupposition of causal attribution that, ceteris paribus, if an effect only occurs in the presence of a candidate cause, then that candidate cause does indeed exert causal powers over the effect. To be sure, the promise of these accounts hasn’t been fully worked out, and the current formulations will no doubt be subject to counterexamples. But if anything much like the current accounts are right, the Causal Principle would likely be inferred from the Correlation Principle. The Correlation Principle says that effects of the type action never occur in the absence of candidate causes of the type agent. In that case, given the above presupposition of causal attribution, the Causal Principle would follow. As a result, on this approach, since young children accept the Correlation Principle, if they apply causal reasoning to the domain of action (as seems likely), they will likely come to grant causal powers to agents. Hence, the developmental evidence suggests that very young children accept the Causal Principle. For the Correlation Principle either reflects a prior belief in the Causal Principle, or, alternatively, the Correlation Principle would likely lead children to infer the Causal Principle. The upshot of this is that the developmental evidence does indeed give us a basis for concluding that young children accept the The Folk Psychology of Free Will 481 # Blackwell Publishing Ltd. 2004 first condition of agent-causal theories. The developmental research indicates that, from a very young age, children regard agents as distinctive and causal. They quantify over agents, and they locate in the agent a causal power to produce an action. What we have found thus far is that children apparently attribute causal powers to agents. This alone does not get us to agent-causation. For children might regard agents as causes and also expect such causal activity to fit comfortably into a deterministic framework. Recall the example from Rowe. We might say that the acid has a power to bring about a change in the zinc and that the acid exerted its power. But did the acid have the power not to bring about this change? No. That’s what excludes it from counting as a true agent-cause. So, what we do not yet have evidence for is the other condition on agent-causal theories: an agent is an agent-cause of a given action only if the agent could have not caused it, that is, only if the agent could have done otherwise. Neither the evidence on agency detection nor that on causal attribution gives us any reason to think that children accept this condition. 3.2. Agents Could Have Done Otherwise Although there is excellent work on agency detection, there is, as far as I can tell, no evidence whatsoever on whether children think that an agent could have done otherwise. So I collected data in some very simple experiments. The clearest way to pose the ‘could have done otherwise’ question is, of course, as a counterfactual. As a result, a rudimentary facility with counterfactuals is necessary for understanding basic questions about whether an agent could have done otherwise. Fortunately, there is good evidence that children do understand counterfactuals from a young age. Three-year olds are good at answering simple counterfactual questions, and four-year-olds can answer even somewhat complicated counterfactual questions (Harris et al., 1996; German and Nichols, 2003). All of the children in the experiments to follow were at least 3-years-old, so it’s likely that the children are competent with counterfactuals. 6 More broadly, there is little evidence on children’s understanding of voluntary action. However, there has been intriguing work by Josef Perner on the child’s understanding of reflexes and voluntary action. Perner maintains that children don’t understand voluntary actions until they have a representational theory of mind (1991), and Perner and Birgit Lang have new results on children’s understanding of reflexes and voluntary action. In a group of children aged 3 to 5 years, knee-reflex responses were elicited and the child was told ‘Look your leg moved! Did you mean to do this?’ Surprisingly, a majority of the children mistakenly said that they did mean to move their leg. Moreover, there was a clear correlation between failing a false belief task and mistakenly judging the knee-reflex to be intentional (Lang and Perner, 2002). The interpretation of this as evidence in favor of Perner’s (1991) proposal is complicated by the fact that in some scenarios, children do seem to succeed at recognizing that certain of their actions were not intentional (see Perner, 1991, p. 219). Nonetheless, this is an important area for further research. 482 S. Nichols # Blackwell Publishing Ltd. 2004