R1. Introduction
It is a delight to receive such thoughtful commentaries, and a gift to be able to reply. They offer many important improvements to the target article. My goal is to make these as clear as possible. To this end, several sections of this reply summarize the most valuable revisions to the theory of rationalization as representational exchange offered by the commentaries (see R2.1, R4.3, R5.4, and R7.4). The target article focused the representational exchange function of rationalization, but the commentaries show that is important to also acknowledge its self-serving functions and to keep these distinct. The target article focused on post hoc rationalization, but these commentaries extend it to cover antecedent and more dynamic cognitive processes. The target article focused mostly on the personal benefits of representational exchange, but these commentaries extend it to cover the many important interpersonal and collective benefits. Most significantly, several of these commentaries translate the academese of the target article (“rationalization extracts implicit information”) into the native language of human experience. We constantly try to find, or make, meaning in our lives. We do this individually, and we do it collectively. We do it through self-reflection: By interrogating the purpose of our actions and attitudes and by striving for consistency and coherence among our beliefs. And, whether we realize it or not, we do it for self-improvement. The theory of representational exchange offers a natural account of these behaviors. I conclude by considering the challenges, opportunities, and pitfalls of self-improvement by rationalization.
Perhaps, then, I should not have titled the target article “Rationalization Is Rational,” but instead “Self-Reflection Is Useful.” Although I can't shake the feeling that there was something right about the original title, I'll have to think a little harder about what that might be.
R2. Motivated reasoning versus representational exchange
There must be at least two kinds of rationalization. One of these is proposed by the target article, and Ullman calls it “self-benefitting.” Through representational exchange it improves our beliefs, desires, and ultimately our behavior. But several commentaries urge that we not overlook the other kind of rationalization, which Ullman calls “self-serving” (see also Quilty-Dunn), and Ellis & Schwitzgebel call “rationalization in the pejorative sense.” It is basically a variety of motivated reasoning. In order to convince ourselves or others that we did the right thing, we invent explanations for our behavior that cast it in favorable light. Such rationalizations do not distill a liquor from the mash of non-rational influences, but instead slap a fraudulent label on a jug of rotgut. Ellis & Schwitzgebel, Quilty-Dunn, Tierney & Uhlmann, Altay & Mercier, D'Cruz, and Brody & Costa argue that most rationalization is self-serving. Whether or not it is more common, it is certainly more prototypical. Perhaps Ellis & Schwitzgebel say it best: “Rationalization has a bad name for good reason, and that will be missed on Cushman's theory.”
How badly is it missed? The target article certainly doesn't deny that rationalization is sometimes self-serving. To the contrary: “This family of theories likely explains a part of the function of rationalization. If the present account also explains a part, then it is a complementary but largely independent explanation” (target article, sect. 2.3, para. 4). Yet, by calling these two kinds of rationalization by the same name, the target article implied an unnecessary competition between them.
The wiser course, urged by Ellis & Schwitzgebel, Ullman, Quilty-Dunn, and D'Cruz is to allow each its own name. Following Ullman, I will call them self-serving (i.e., rationalization as motivated reasoning) and self-benefitting (i.e., rationalization as representational exchange). Self-benefitting rationalization is not necessarily distinct from self-serving rationalization because of consequences – that is, adaptive versus maladaptive outcomes. Presumably both are usually good for people but sometimes bad. Neither is it necessarily distinct in its mechanisms. It is possible that both forms of rationalization share a common mechanism. Rather, what distinguishes self-benefitting rationalization from self-serving rationalization is its function: representational exchange. Following from this function, self-benefitting rationalization should generally improve the quality of reasoning, while self-serving rationalization will generally diminish it.
Of course, the key question is whether self-benefitting rationalization actually happens. Insofar as the human mind is designed to rationalize, is representational exchange ever the ultimate goal? Several commentaries raise important doubts. For instance, existing research shows that rationalization is often empirically associated with a worse, not better, quality of reasoning and action. Summarizing some this evidence, Tierney & Uhlmann entertain the possibility that “rationalizers could tend to be unhappy, socially unpopular underperformers, rejected and ineffective due to their self-serving arguments and lack of insight into their own actions” – a vision more or less shared by D'Cruz, Brody & Costa, and Tierney & Uhlmann. (Quilty-Dunn shares the view that rationalization degrades reason, but views it as a useful “drive state that pushes us to reorient our beliefs in ways that keep our self-esteem up, even when we are irrational, unstable, or immoral, and thereby keep ourselves motivated by avoiding the awful truth.”)
In support of these claims, Tierney & Uhlmann review evidence the rationalization of gender bias in hiring decisions is associated with greater gender bias in hiring decision. Brody & Costa review the evidence that rationalization of violence and criminality is associated with greater antisocial behavior. Quilty-Dunn reviews evidence that rationalization of meat consumption is associated with the belief that it is “necessary, normal, natural, and nice.”
If you believe that gender bias, violence, criminality, and eating meat are bad things, then surely it is bad to rationalize them. But does that mean that rationalization itself is bad? Obviously, good objects can be used for bad ends: A chef's knife makes a fist fight worse, but it is awfully handy for cooking. When we look for bad effects of rationalization, we can find them; what would happen if we instead looked for good effects? For instance, is rationalizing charitable acts associated with more prosocial behavior?
Brody & Costa review evidence that high levels of rationalization characterize several mental disorders. They conclude that rationalization is generally maladaptive. But it is easy to see the flaw in arguments of this kind. Anxiety disorders are characterized by excessive fear, but this doesn't mean that fear is generally maladaptive. Presumably many mental disorders arise from the dysregulation of thought processes that, in moderation, are essential to proper psychological function. In fact, as Saxe & Nettle note, self-benefitting rationalization (i.e., critical self-reflection about the causes of one's behavior) is essential to cognitive behavioral therapy, one of the most effective, evidence-based treatments available for mental disorders. It is an advantage of the representational exchange framework that we can identify both why this kind of behavior can be helpful and also how its misapplication can cause harm.
De Neys summarizes further evidence consistent with self-benefitting rationalizations. Across several experiments, they find that rationalization is often used in service of finding the right explanations for the right actions, but in cases where the actions themselves were a product of intuition. For instance, consider the famous “bat and a ball” question that appears in the Cognitive Reflection Task (Frederick Reference Frederick2005). When people report the correct answer to this word problem, they often cannot say exactly why it is correct. Rather, they have to deliberate for a moment about why it is correct, eventually explaining their answer in terms of sound reasons, but ones that were only rendered explicit by post hoc rationalization. This illustrates rationalization of a self-benefitting kind.
R2.1. Revision: Distinguishing two functions of rationalization
The pejorative model of rationalization currently dominates. It proposes that while rationalization may have some benefits, improving reasoning is not among them. Not surprisingly, experiments designed to show examples of rationalization in the pejorative sense tend to find that it is associated with bad outcomes. The target article was not intended to deny the existence of self-serving rationalization. It exists. We've documented it. We understand it pretty well. Rather, the contribution of the main article is to point out that the same mechanism – or a very similar one – can also improve reasoning.
Consider an analogy to stereotyping. Prototypically, stereotypes are bad things that make us worse off. This occurs in part because stereotypes can become a vehicle for motivated reasoning, allowing us to distort our factual understanding of the world in order to serve our social interests. Yet, everybody also recognizes that there is also a “rational” side to stereotypes. Given that our minds must make the best use of limited data and limited computation, it makes sense to understand and predict the behavior of tokens (e.g., individual people) by drawing on statistical generalizations at the level of the type (e.g., social group). A common form of computation and representation at the mechanistic level – the stereotype – can serve two quite different functions. One is generally helpful; another can be extraordinarily harmful.
A key outstanding question is whether, and to what extent, self-serving and self-benefitting rationalization share a common mechanism. There is a spectrum of possibilities. At one extreme, it may be that fully distinct mechanisms engage in rationalization of self-serving and self-benefitting kinds, and that the only connection between these mechanisms is in the abstract form they take. On this view, they are as distinct as the umbrella and the parasol sitting beside each other in a closet – although similarly constructed, they are both physically and functionally distinct. Or, it may be that there is a single, common mechanism that happens to serve two independent purposes. On this view, rationalization is like a roof: A single object that serves the dual purposes of sheltering from rain and from the sun.
R3. Does rationalization really happen?
Tierney & Uhlmann, Stafford, Dahl & Waltzer, and Pärnamets, Johansson, & Hall (Pärnamets et al.) argue that rationalization is a rare and weak force in our minds. These concerns must be taken seriously because they are supported by some of the very evidence cited in the target article. In a study by Sharot et al. (Reference Sharot, Velasquez and Dolan2010), for instance, the experimentally induced perception of having chosen one vacation destination over another brought about only a 0.06 point shift in subsequent preference, on an 6-point scale.Footnote 1 Although there are some published large effect sizes of rationalization, these commentaries make a persuasive case that the effects of standard laboratory methods (such as the “free choice paradigm”) are usually quite small.
It is notable, however, that standard experimental methods of eliciting rationalization focus specifically on cases where it would be irrational. In other words, social psychologists tend not to induce rationalization by prompting people to acts in an adaptive manner and then, through a bit of self-reflection, drawing useful lessons from it. Instead, they tend to devise circumstances in which the experimenter has cleverly induced somebody to act in a random or even maladaptive manner, and thus any rationalization would necessarily corrupt reasoning.
This is a useful approach from the standpoint of experimental design. In order to draw strong inferences about psychological mechanisms, it is helpful to put a participant's behavior under experimental control and then demonstrate that they do something predicted by the mechanism, but otherwise irrational and, therefore, very difficult to explain in any other way (see Saxe Reference Saxe2005).
Unfortunately, the very same features of an experiment that allow us to draw strong inferences may tend to generate weak effects (Mook Reference Mook1983). If rationalization “approximate(s) a well-understood form of rational inference” (target article, sect. 3.3, para. 3) then, when people are tricked into believing they chose Greece over Thailand, they may rationally infer little from it. Rather, they may conclude that this “choice” was mostly a matter of chance (Gershman Reference Gershman2019; Hawthorne-Madell & Goodman Reference Hawthorne-Madell and Goodman2019). And, therefore, they may exhibit little change in their beliefs or attitudes.
To see why, it helps to focus on the details of this particular experiment: Although exquisitely designed for strong causal inference, it is wholly unlike ordinary life. Participants are instructed to click on one of two nonsense words that, they are told, contain subliminal information eliciting an unconscious decision between vacation destinations. In fact, the experimenter randomly assigns the participant one vacation destination or the other. Upon “learning” that their “choice” was Greece (for instance), the participant might begin to introspect a bit. “How do I feel about Greece? Do I really have strong feelings that it is a better vacation destination?” Due to the very logic of this experiment, on average the answer to this question must be “no,” since the participant's apparent choice is randomly assigned. In light of the peculiar circumstances of the choice, and lacking any actual underlying attitude supporting it, people may reasonably conclude that there are few “true beliefs” or “useful desires” to be gleaned. Evidence for rationalization in this context allows for a strong causal inference, but we ought to expect the effect size to be very small.
Contrast this with a person who is actually choosing between vacations in Greece and Thailand – say, in a travel agency, with colorful brochures and detailed information. They feel a strong, intuitive pull toward Greece, but can't yet put their finger on why. They try to explain their feelings to themselves and to the travel agent. In this circumstance, they might reasonably conclude that there is something true, and something important, to be learned during self-reflection.
Dahl & Waltzer raise a second objection to the thesis that rationalization makes much contribution to human thought and behavior. They write, “If rationalization were widespread, we would have to abandon a premise of much discourse: that our beliefs and actions are generally based on reasons.” Along similar lines, Tierney & Uhlmann summarize findings that “Time 1 explicit attitudes predict Time 2 behaviors far better than past behaviors predict future self-reported attitudes, calling into question the prevalence of post hoc rationalizations for past actions… [and] relegating the ‘rationalizations are rational thesis’ to address only a small portion of the attitude-behavior relationship.”
These objections depend on the premise that rationalization is opposed to reason, and so the more influential one is, the less the other must be. Yet, the very premise of the target article is that rationalization can actually improve reasoning. Indeed, the theory depends upon the premise that we are often rational creatures whose Time 1 beliefs and desires appropriately shape our Time 2 behaviors; after all, there is no point exchanging representations in order to improve reasoning if it is impotent or inert! (See sect. R5, where the target article is defended against the critique that we never actually reason).
By analogy, consider the proposition: “Humans frequently sleep in order to benefit their waking lives.” It hardly counts as an objection to say, “But if we sleep so frequently, this leaves no time for us to be awake!” Likewise, “Research suggests we accomplish nearly all tasks while awake, relegating to sleep to an inconsequential role!” Just as a few hours of sleep render our many waking hours more productive, so too might a few moments of rationalization support many moments of reasoning.
Berthelette & Kalbach agree that nearly all adaptive behavior traces back to reasoning, but they arrive there by a different route. They argue that when habits work well – that is, when they are appropriately attuned to circumstances and recommend the correct behavioral policy – they are “best explained as being caused by a standing belief and desire” Thus, there can be no rationalization “because the relevant beliefs and desires were there all along.” The implied psychological model seems to be that habits arise exclusively through the stamping-in of rational actions through repetition (something like practicing a piano piece until it becomes “muscle memory”). This kind of habit formation is well documented (Dezfouli & Balleine Reference Dezfouli and Balleine2012; Reference Dezfouli and Balleine2013; Dezfouli et al. Reference Dezfouli, Lingawi and Balleine2014; Miller et al. Reference Miller, Shenhav and Ludvig2019). But it is also well-documented that we cache the context-dependent value of actions based on a history of reward (Schultz et al. Reference Schultz, Dayan and Montague1997; Glimcher Reference Glimcher2011; Morris & Cushman Reference Morris and Cushman2019). In such cases the habitual action does not depend, either presently or historically, on a generative causal model – that is, on beliefs (Dolan & Dayan Reference Dolan and Dayan2013). In other words, habits are not always just “precompiled rational acts.” Rather, they are sometimes formed by a computationally and representationally distinct process.
In summary, the objection that “rationalization doesn't happen” begins with the premise that most rationalization is self-serving. Standard social psychology paradigms are well designed to estimate the effect of this “bad” kind of rationalization, and to the extent that it occurs, it stands in opposition to sound reasoning. In contrast, however, the target article proposes a second kind of rationalization. Its influence may be systematically underestimated by standard paradigms, and it is not opposed to reason. Future experiments should explore whether, how often, and how powerfully it shapes our thinking.
R4. The mechanics of rationalization: Prioritized representational exchange
Several commentaries focused on the mechanisms underlying rationalization and representational exchange (Harmon-Jones & Harmon-Jones, Sievers, Simon & Holyoak). The target article focused on a simple case: The post hoc rationalization of action. Here, a person has already performed an action and adjusted their beliefs and desires in a manner that would have rendered the action rational.
The commentaries offer two basic extensions of this model. First, rationalization is often antecedent: “we rationalize our actions not only after we perform them, but also before we perform them and sometimes as a condition of performing them” (Ellis & Schwitzgebel). Similarly, we often rationalize not an action, but an attitude or emotion (Railton). Second, rationalization is not merely a process of updating beliefs and desires conditional on action, but a more general process of achieve coherence among a variety of mental state representations, with multidirectional influences.
R4.1. Antecedent rationalization
Sievers and Simon & Holyoak assemble impressive evidence for antecedent rationalization, in which a person adjusts their beliefs and desires prior to action (see also D'Cruz, Ellis & Schwitzgebel, and Tierney & Uhlmann). As Sievers points out, “Nothing entailed by antecedent rationalization precludes the correctness of Cushman's model of rationalization as a representational exchange mechanism. It does, however, demand an expansion of that model.” This is an important point: Antecedent rationalization happens. It fosters representational exchange when the anticipated action is influenced by a non-rational adaptive control mechanism (e.g., habit, norm, or instinct).
In a paradigmatic case of antecedent rationalization, we bring beliefs and desires into alignment with a non-rational system, such as habit. For instance, suppose that cached value representations (habit) recommend choosing cake instead of tea for dessert, but that a planning system (reason) recommends tea over cake. We would adjust our reasons for tea (“I've already eaten too many calories in this meal…”) so that they become aligned with our habitually favored choice (“…but it's Lisa's birthday, and she'd be disappointed if I didn't have some cake”). In this case our cached values remain constant while our reasons change.
Extending this idea, Railton proposes that we often rationalize our emotions, and not just our actions. Thus, for instance, suppose that a coworker inspires intense feelings – perhaps admiration, jealousy, anger, love or contempt. Even if these feelings do not move us to any particular action, they still may shape our beliefs by rationalization. In this case, we are learning what to think or want not by observing how we act, but instead by observing how we feel. In this sense it is akin to antecedent rationalization, which is also prompted by a thought, rather than an act.
But once we acknowledge that certain non-rational forms of thought (e.g., emotion, or habit) can antecedently adjust the representations that contribute to reasoning (beliefs and desires), surely we must also acknowledge that the influence could also run the opposite way, as well. To return to the example above, couldn’t the reasons remain the same, while the habit-based values are adjusted so that tea, rather than cake, becomes the habitual response (e.g., Gershman et al. Reference Gershman, Markman and Otto2014)? These are both instances of representational exchange, but they push information in opposite directions. In one case, information flows from habit to reason; in the other case, from reason to habit.
Presumably, the controller that you “trust” more – the one that is most likely to generate the correct policy – should both determine your action and serve as the source for representational exchange. Deciding which controller to trust is often called the metacontrol problem. Classically, metacontrol is construed as a process of deciding which system governs action. (In this manner, it is pivotal to post hoc rationalization, which depends upon the control system underlying action). Importantly, however, it could also be repurposed as a mechanism for determining the direction of representational exchange, for instance, during antecedent rationalization. Metacontrol determines which system should be trusted, and it is an active area of research with several promising theoretical approaches in development (e.g., Daw et al. Reference Daw, Niv and Dayan2005; Griffiths et al. Reference Griffiths, Lieder and Goodman2015; Kool et al. Reference Kool, Gershman and Cushman2017; Shenhav et al. Reference Shenhav, Botvinick and Cohen2013).
In summary, antecedent rationalization surely occurs. But it draws our attention to a crucial problem faced by the theory of representational exchange: How is the direction of exchange established? The existing literature on metacontrol, which is similarly concerned with the problem of deciding which control systems to trust, offers promising leads. It also immediately suggests a more dynamic model in which rational and non-rational processes continually exert influence upon each other in a manner prioritized by metacontrol.
R4.2. Coherence
Simon & Holyoak and Harmon-Jones & Harmon-Jones both contrast the theory of representational exchange with alternative models focused on coherence. To quote Simon & Holyoak, “In models of coherence-based reasoning… complex decision situations are represented by networks in which the relevant variables are interconnected via excitatory and inhibitory links. Constraint satisfaction mechanisms settle the network into a stable state of coherence, in which mutually supportive connections (i.e., those that “go together”) activate one another and collectively inhibit their rivals” (emphasis added).
Simon & Holyoak’s commentary has a lot to say about the mechanisms by which coherence is achieved, but little to say about why coherence itself is a useful property of the mind. Several other commentaries furnish possible adaptive explanations. Harmon-Jones & Harmon-Jones appeal to the idea that coherence reduces internal conflict: “The organism will have a difficult time enacting a course of behavior when trying to follow the action tendencies evoked by conflicting cognitions.” And we have already summarized several related proposals that coherence has the useful social function of rendering our behavior more predictable to others (Levy; Saxe & Nettle; Stanford, Thomas, & Sarnecka [Stanford et al.]). Notably, however, these explanations focus on intrinsic benefits of coherence: that is, the lack of discrepancy or conflict between representations. Rather than positing that we prioritize “better” representations and adjust “worse” ones, the default assumption seems to be that conflicting representations will exert equal and opposite influences upon each other.
It is possible, of course, that people are designed to achieve coherence among their mental states merely for the sake of coherence itself, without attempting to increase the average quality of the resulting representations. But as long as our minds are achieving coherence, shouldn't they hold relatively fixed those mental state representations that are most likely to be adaptive, while adjusting those that are most likely to be maladaptive? By analogy, consider a marketing company with one headquarters in New York and another in L.A. It would make the company more efficient and nimble to consolidate into a single headquarters. One possibility is to split the difference and relocate to Topeka – at least it would increase coherence! Another possibility, however, would be to carefully consider the performance of the New York and L.A. offices, and then to move the weaker headquarters to the location of the stronger one. This solution achieves coherence with a much more promising outlook for the bottom line. But it also makes greater demands on the decision maker. Like antecedent rationalization, it requires a method of prioritizing some representational exchanges over others.
R4.3. Revision: Coherence with priority
These commentaries offer two important amendments to the original theory of representational exchange: One mechanistic, and one functional. The mechanistic amendment is simply to point out that rationalization is not always ex post, but may often be ex ante, or a dynamic process in which coherence is achieved among many sorts of representations at any particular time. The functional amendment addresses the question at the heart of the target article: Why bother with rationalization? The target article offered two reasons; the commentaries suggest a third.
First, the target article suggested that we rationalize in order to improve reasoning, imparting truer beliefs and more useful (i.e., fit) desires. But this answer must be incomplete, because natural selection does not directly favor superior reasoning for its own sake; it only favors superior behavior. Why would we expect rationalization to improve the overall behavior of the organism (see Berthelette & Kalbach)? To the extent that the organism is already acting well based on some non-rational system, what use is there for improvement to its beliefs and desires?
The target article addressed this question in the broader context of representational exchange, offering a second kind of explanation. Different ways of representing information have different costs and benefits. Among the distinctive benefits of reasoning are its flexibility (i.e., the ability to rapidly adjust in light of new information) and its unique form of generalization (i.e., the ability to generate a policy in novel circumstances according to underlying causal principles). Thus, extracting information from non-rational systems may enhance an organism's capacity for flexibility and generalization, improving its behavior overall. Meanwhile, other directions of representational exchange (e.g., from reasoning to habit, as during habitization) can leverage the distinct advantages of other forms of behavioral control (e.g., the speed and cognitive efficiency of habitual action).
The commentaries, however, draw our attention to another manner in which representational exchange can improve our behavior. If it can successfully prioritize from moment to moment which behavioral systems get to be the source versus the receiver of representational exchanges, this would ensure that representational exchange would lead not only to coherence among systems, but also improvement among them. This stands out as a key area of development for the theory of representational exchange. Existing theories of metacontrol provide a natural starting point.
R5. Social functions of rationalization
The theory of representational exchange proposes an “ultimate” or “computational”-level account of rationalization (Marr Reference Marr1982; Tinbergen Reference Tinbergen1963). In other words, it is supposed to answer the question, “Why would natural selection favored rationalization?” Several commentaries also take up this question. Some extend the theory of representational exchange, while others propose alternatives to it. Nearly all of these contributions, however, focus on social functions of rationalization.
R5.1. Cohesion
Van Bavel, Sternisko, Harris, & Robertson (Van Bavel et al.), and Gelpi, Cunningham, & Buchsbaum (Gelpi et al.) propose that rationalization promotes social cohesion. When we rationalize the others’ behavior, or collective cultural practices, we bring our beliefs and desires into closer alignment with that of others. The result is cohesion and a sense of shared identity, which, they propose, has direct personal benefits. As Gelpi et al. write, “Sharing a belief with those in one's community is therefore beneficial not only when (and because) that belief is true, but also when (and because) it provides an individual with the benefits of a group, such as a sense of belonging and easily accessible shared knowledge.”
Certain versions of this thesis are closely related to the persuasion and blame-avoidance functions of rationalization discussed in the target article (Altay & Mercier, see also Haidt Reference Haidt2001; Mercier & Sperber Reference Mercier and Sperber2011; Tedeschi et al. Reference Tedeschi, Schlenker and Bonoma1971; Tetlock Reference Tetlock2002; Von Hippel & Trivers, Reference Von Hippel and Trivers2011). For instance, suppose a political leader is caught performing an apparently corrupt action. In this case she and her supporters might rationalize her behavior by attempting to provide benign explanations for it (Altay & Mercier). But the commentaries also consider extensions of this thesis to many other kinds of cases: We might rationalize collective practices like holidays and sports; shared norms of how we spend time, organize families, and maintain friendships; and, perhaps above all, moral values (Gelpi et al. and Graham).
Van Bavel et al. argue that, when it supports social cohesion, “rationalization often facilitates distortion of rather than approximation to truth.” Altay & Mercier and Gelpi et al. mostly share this view. In other words, theirs is an account of self-serving, pejorative rationalization (albeit one that serves an important social purpose). They do not consider it maladaptive, however, because it pays to get along with one's peers. (In contrast, Graham and Laurin & Jettinghoff raise the possibility that rationalizing collective action can sometimes be adaptive due to its self-benefitting nature. This possibility is considered at length in sect. R7.2.)
R5.2. Coordination
Saxe & Nettle, Stanford et al., and Levy propose an alternative social function of rationalization: Facilitating coordinated joint action. Humans are nature's paramount cooperators; we have an unmatched ability to flexibly coordinate our behavior to accomplish joint goals. We succeed in part because we can predict each other's behavior. We mostly predict each other's behavior by assuming rational planning and choice (Baker et al. Reference Baker, Saxe and Tenenbaum2009; Dennett Reference Dennett1987). Thus, the commentators propose, it will be easier to explain and predict another person's action if it exhibits consistent adherence to norms of rationality across time. Rationalization may be useful because it imposes such consistency. As Levy writes, “By making ourselves predictable, we enable more efficient cooperation, which is essential for the flourishing of social animals like us (Tomasello Reference Tomasello2014). If each of us can predict how the others will behave, we can more efficiently play our part in joint actions, without interfering with one another or introducing redundancies.”
If anything, this understates the case. Successful joint action often requires not only that we model each other's intentions, but also that we adopt joint intentions, in which the behavior of each partner is generated by shared goals and common knowledge (Kleiman-Weiner et al. Reference Kleiman-Weiner, Ho, Austerweil, Littman, Tenenbaum, Papafragou, Grodner, Mirman and Trueswell2016; Tomasello Reference Tomasello2005). Possibly, then, rationalization allows us to translate the behavioral prescriptions of non-rational systems into a format amenable to joint intentional action.
R5.3. Communication
Levy and Pärnamets et al. consider a final social function of rationalization: to foster communication between people. Whether or not reasoning is the language of thought, it is certainly the language of language. We explain ourselves with words, and these words typically express our beliefs, desires, plans, and reasons. Perhaps this is because beliefs, desires, plans, and reasons are especially effective kinds of representations for communication.
It is obvious, for instance, that one person cannot describe their instinct (i.e., innate behavior) to another person and thereby cause the other person to acquire an instinct. Rather, some change of representational format is required. In theory, a person could express their instinct (or habit, or norm) simply as a behavioral prescription. But it will often be much more useful to learn the reasons behind a person's policy than to simply learn the policy itself and be forced to copy it. This is because each of us occupies different circumstances, and so the optimal policy for one person may be suboptimal for another. Suppose, for instance, you love peanuts, but I am allergic to them. If you tell me, “Eat these cookies!” I may be substantially worse off than if you tell me, “I eat these cookies because I love the peanuts in them.” Rationalizing prior to communication presents information in a format such that a person can flexibly adapt socially learned information to their own unique circumstances.
The target article contrasted two possible pathways of social learning (see Fig. 3 of the target article). In the first, a person acts, an observer engages in theory of mind to extract the beliefs and desires implicit in the action, and the observer finally adopts those beliefs and desires herself (i.e., by informational conformity). In the second, a person acts, an observer directly copies the behavior (i.e., by normative conformity) and, upon rationalizing her own behavior, the observer finally extracts its implicit beliefs and desires and adopts them. The key insight of Levy and Pärnamets et al. is to add a third pathway: A person acts, he rationalizes his own behavior (extracting beliefs and desires from it), and he verbally reports these to an observer who then adopts them. This is a promising addition to the framework.
R5.4. Revision: The social dimension of representational exchange
Each of these reviews posits that rationalization has important and diverse social benefits, fostering cohesion, coordination, and communication. Some of these functions, such as group cohesion, share little overlap with the theory of representational exchange. They more likely explain self-serving rationalization. Others of these, such as coordination and communication, are likely explanations of self-benefitting rationalization. They show that effective representational exchange improves not just our personal thoughts and actions, but also our collective ones.
What is most remarkable about this set of commentaries, however, is their implied commitment to a bold thesis: That social benefits are not just an important explanation of rationalization, but also of reason itself. In fact, there is an even more extreme version of this hypothesis, according to which reason (beliefs, desires, plans, etc.) plays no important role in guiding our behavior, but is instead a socially useful way of talking about our behavior. The next section considers several proposals that engage this possibility.
R6. Do we reason at all?
The target article argued that often the function of rationalization is to improve reasoning by improving its raw materials: The beliefs and desires that it assembles into plans. Levy, Roskies, and Veit, Dewhurst, Dolega, Jones, Stanley, Frankish, & Dennett (Veit et al.) question whether humans actually reason – in other words, whether we hold beliefs and desires and use them to make plans. None of these authors dispute that we talk in terms of reasons, beliefs, desires, and plans. And, they agree that such talk has meaning and utility. But, they suggest, maybe it's just talk. (Some endorse this possibility more strongly than others). The target article described rationalization as a “useful fiction,” but is the real fiction reason itself?
According to these critiques, people exhibit “flexible response[s] to environmental and internal information,” but they do so in ways that do not depend on “explicit representations” of belief and desire (Levy). Rather, what exist are “some sort of proto-mental states” (Veit et al.), the “great majority of these states are dispositional and may never be explicitly represented” (Levy). Roskies offers a clear introduction to these ideas, employing the helpful example of a thermostat: “To the degree that we can predict and explain a system's behavior by imputing folk psychological states [e.g., beliefs, desires, plans, and reasons] to that system, so we are warranted in that imputation. For Dennett [Reference Dennett1987], even simple systems such as thermostats are legitimate targets of the intentional stance, even though no standard realist would support the view that thermostats have mental states.” The proposed function of the intentional stance, not dissimilar to Bem's (Reference Bem1972), is not to improve one's own reasoning, but rather to allow us to describe, predict, and explain our own and others’ actions.
There is much to like in these commentaries, but also something amiss. Surely, it is often useful to impute reasons to psychological systems that do not reason. This is the essence of the target article's model of rationalization as representational exchange. And, as discussed in section R4, rationalization has benefits not only for the individual who rationalizes, but also for those who depend upon her for collaboration and communication. But here is the key question: Do these facts explain away reasoning itself? If we accept that our experience of having reasoned in sometimes a fiction, do we have any reason to accept that it is ever a reality?
We do. As defined in the target article, to act by reasoning is to estimate the expected value of candidate actions by consulting a generative model of their consequences, and to then choose actions proportional to their expected value (following, e.g., Daw & Dayan Reference Daw and Dayan2014). There is ample evidence that people learn and represent generative causal models (Buckner et al. Reference Buckner, Andrews-Hanna and Schacter2008; Gershman Reference Gershman2018; Gläscher et al. Reference Gläscher, Daw, Dayan and O'Doherty2010) and can use these to assess the likely consequences of candidate actions (Deserno et al. Reference Deserno, Huys, Boehme, Buchert, Heinze, Grace, Dolan, Heinz and Schlagenhauf2015; Doll et al. Reference Doll, Duncan, Simon, Shohamy and Daw2015; Simon & Daw Reference Simon and Daw2011; Vikbladh et al. Reference Vikbladh, Meager, King, Blackmon, Devinsky, Shohamy, Burgess and Daw2019). By considering these consequences in light of our goals our preferences, we can estimate the instrumental value of candidate actions (Bakkour et al. Reference Bakkour, Palombo, Zylberberg, Kang, Reid, Verfaellie, Shadlen and Shohamy2019; Daw et al. Reference Daw, Gershman, Seymour, Dayan and Dolan2011; Jones et al. Reference Jones, Esber, McDannald, Gruber, Hernandez, Mirenzi and Schoenbaum2012). These value estimates reliably influence choice (Cushman & Morris Reference Cushman and Morris2015; Daw et al. Reference Daw, Gershman, Seymour, Dayan and Dolan2011; Deserno et al. Reference Deserno, Huys, Boehme, Buchert, Heinze, Grace, Dolan, Heinz and Schlagenhauf2015; Doll et al. Reference Doll, Duncan, Simon, Shohamy and Daw2015; Simon & Daw Reference Simon and Daw2011). These are precisely the kinds of representations and computations that can be improved by representational exchange.
In summary, these commentaries emphasize the striking and illuminating fact that we can profitably understand a thermostat by treating it as rational, even if it is not (i.e., by adopting the intentional stance). This immediately suggests a subversive and alluring possibility: Maybe, despite feeling and talking like rational creatures, we are more like very complicated thermostats. Yet, as an empirical matter, we are not. When a thermostat adjusts the temperature, it is not by encoding a generative model of the world, computing the likely consequences of various actions, and then choosing actions proportional to the resulting instrumental value estimates. But, when a human adjusts the thermostat, it often is.
R7. Making meaning
We cannot always know the causes of our behavior, but rather must infer them (Bem Reference Bem1972; Nisbett & Wilson Reference Nisbett and Wilson1977). We do this in part because it brings order and meaning in our lives. Many theories posit that we search for order and meaning because it makes us feel better, which is a proximate explanation (Baumeister Reference Baumeister1991; Hasselkus Reference Hasselkus2011; Reker et al. Reference Reker, Peacock and Wong1987; Webster & Kruglanski Reference Webster and Kruglanski1994). The theory of representational exchange provides a complementary ultimate explanation: The meaning we make out of our actions, attitudes, and emotions (Railton) renders useful information available for reasoning. A final set of commentaries takes this idea of meaning-making as its point of departure. The authors observe how phenomena as diverse as daydreaming (“everyday thought experiments”: Ullman), system justification (Laurin & Jettinghoff), and ideological commitment (Graham) can be understood as attempts to find meaning, breaking new ground for the theory of representational exchange. They make a convincing case that these are powerful and constant forces in our thinking. They also offer a cautionary note, questioning what happens when even self-benefitting rationalization is misapplied (Weinberg & Weinberg, Pärnamets et al., Graham).
R7.1. Rationalization and personal meaning
People daydream about all kinds of fantastical things. They love it. They'll do it on their own, and as Ullman shows, they'll happy to do it if you ask. This is remarkable because many of the questions posed by McCoy et al. (Reference McCoy, Paul, Ullman, Goldman and McLaughlin2019, p. 248) are patently outlandish:
Imagine that aliens come down to Earth, and give you the option to go with them on their travels throughout the universe. The aliens are friendly and honest, and tell you that you would see amazing things on your travels with them if you decide to go with them.
If you decide to go, you will have a week to say goodbye to your family and friends. Once you leave, you will never again return to Earth, nor be able to communicate with people on Earth.
Do you go?
Different people give different answers, and people are bad at guessing what others will say, yet most people report being highly confident in their own personal answers. Not only do people say they enjoy thinking about things like this, they say that they learned something from it – mostly things about themselves:
That I am more risk averse when faced with transformative decisions.
That I am attached to my actual, earthly existence more than I might have thought otherwise.
That I value personal relationships more than I thought.
These findings comport with a large literature showing that when people have downtime, their minds turn toward thoughts about imagined possibilities for themselves, in the past and in the future (Buckner et al. Reference Buckner, Andrews-Hanna and Schacter2008).
As Ullman notes, the theory of representational exchange “can help explain why such everyday thought experiments are informative, and also why people like to engage in them.” When people surprise themselves with their own confident answers to hypothetical questions, it means that they are discovering implicit information in their automatic reactions, including emotions (Railton) – information that often contradicts their explicitly held beliefs and values. They consider this information valuable and true – they take it to reflect important insights about who they are and how they ought to feel. By organizing our impulses around meaning, purpose, and narrative, we unpack implicit information from non-rational systems and make it available to improve subsequent reasoning.
R7.2. Rationalization and social meaning
Although the target article focused mostly on the rationalization of one's own actions, Graham and Laurin & Jettinghoff propose a key extension, and Graham writes: “Processes like rationalization often occur at the collective levels of groups, societies, and cultures. Collective rationalization – whereby a group's collective action leads them to update or create new shared beliefs and desires – may be rational as well. Just as individual rationalization can extract useful information from non-rational sources like instincts and habits, communal processes can cohere moral intuitions and norms into the shared ‘useful fictions’ of shared moral narratives.” Graham's commentary focuses on ideology: A shared sense of meaning – an interconnected set of beliefs and values that give a sense of coherence, explanation and purpose to groups (Jost Reference Jost2006). Laurin & Jettinghoff's commentary focuses on system justification: Our tendency to assume that prevailing social structures are good. They agree that we seek not only to ascribe meaning to our individual actions, but also to collective cultural practices.
Laurin & Jettinghoff wonder, however, why we would be justified in not just inferring but actually adopting the beliefs and desires that we attribute to others. In other words, why drink the collective Kool-Aid? “If people were merely trying to identify beliefs and desires” that make sense of other's actions, “there would be no reason to expect system justifying rationales to predominate” over system-undermining ones. Why assume the system is good?
Different commentaries suggest different potential answers to this problem. Gelpi et al. appeal to cognitive efficiency: “Trusting the beliefs shared by one's social group – and ‘outsourcing’ one's own cognition to depend on knowledge held by others in their community – can also reduce the need to engage in cognitively effortful reasoning … (Sloman & Rabb Reference Sloman and Rabb2016).” In contrast, Laurin & Jettinghoff consider the possibility that we do not rationalize the system itself, but rather our own apparent acceptance of it. Perhaps, then, system justification is really justification of one's own complacency. Both of these explanations probably capture a good part of the truth, along with the possibility that the rationalization of collective behavior fosters social cohesion (Van Bavel et al. and Gelpi et al.).
But there is also a far simpler answer: We rationalize collective cultural practices, adopting their implied beliefs and values, because collective cultural practices are usually pretty good. Just as genes are subject to genetic evolution, collective practices are subject to cultural evolution (Boyd et al. Reference Boyd, Richerson and Henrich2011). Thus, if you observe that most people are doing something (or that the cultural system is organized in some way), that thing is more likely to be adaptive than maladaptive. This does not imply that the current norm is the best that could possibly exist, any more than we should suppose that a sparrow's wings are the best wings that a sparrow could possibly have. Rather, it implies that adhering to the norm is probably better than ignoring it, just as the sparrow is better off with her wings than without them.
Confronted with a claim like this – if most people do it, it's probably adaptive – our minds naturally leap to counterexamples. Wasn't it for good reason that our mothers scolded, “If everybody jumped off a bridge, would you?” After all, people do all sorts of terrible things. Most people eat too much, don't get enough exercise, are cognitively and socially biased, and so forth. There was a time and a place when most people thought the Earth was flat; when most people smoked cigarettes; and when most people thought it was acceptable to whip your child, your wife, and your slave. Human progress has always relied on those who say, “I can do better than what most people are doing,” and it always will.
These counterexamples are important and inescapable. But now imagine what it would mean to abandon our cultural inheritance entirely. Imagine a child who grows up without any inclination to think or act like others. Given soup, she would have no inclination to use a spoon just because her big sister does; rather, she would reason about whether a spoon works for her. The same goes for forks, manners, values, laws, school, sports, holidays, fashions, jobs, and so on. Shown a recipe for bread, she would not assume there was anything useful about the yeast, the kneading, the rise, or even the oven. She would just rely on her own knowledge and sound reasoning to deduce the proper manner of baking. Now, whose loaf would you rather eat: Hers or an obedient apprentice's?
Our collective cultural practices are a recipe for life. It is not a perfect recipe – in fact, you can almost certainly improve it with some tweaks. But it is unlikely that you will do better by ignoring the recipe altogether. More likely, any improvements you make will necessarily build upon the cumulative wisdom of generations of cooks that have come before you. (In Newton's terms, seeing further than giants requires not just extending your legs, but also climbing on their shoulders). Sometimes most people do the wrong thing, but these counterexamples miss the forest for the trees – nearly everything we do right was learned, in part, from those around us (Boyd et al. Reference Boyd, Richerson and Henrich2011). This accords with a basic impulse of conservatism in political theory and philosophy (e.g., Hayek Reference Hayek1973): Tradition may require revision or rejection, but it at least deserves respect.
Or, at least, it does if you are interested in fitness within the framework of biological and cultural evolution. As scientists attempting to understand the structure of the human mind – in particular, our propensity to rationalize both personal and collective actions – this is the kind of fitness we are interested in as a descriptive theory. But what was fit for the past may be unfit for the present, and what is fit for nature may be unfit for ethics.
R7.3. Spoiled Kool-Aid: Is rationalization outdated?
While accepting that rationalization may have an adaptive rationale, Weinberg & Weinberg, Pärnamets et al., and Graham ask: Is it the right thing to do? They offer three reasons to think it is not.
The first is familiar: At least some rationalization is self-serving, and it will generally lead us away from the truth. While it may have some adaptive benefits when it works properly (such as cohesion, as noted by Van Bavel et al.), Graham documents the pernicious effects when it goes awry, such as the justification of prejudice or even violence toward outgroups.
Graham goes further, however, by questioning even self-benefitting rationalization. He accepts that by adopting the values implied by instincts and norms, we are adopting values shaped by adaptive processes – biological and cultural evolution, respectively. These may be fit, he notes, but are they morally right? This is an important question; indeed, it is among the most important questions we face. If rationalization captures the conservative impulse to respect traditions, we must confront the fact that many human traditions are morally abhorrent. So, too, is blind respect for them.
Weinberg & Weinberg and Pärnamets et al. offer a final reason to distrust self-benefitting rationalization: While it may have been adaptive when it evolved, perhaps it is poorly suited to present circumstances. We evolved at a time when the world was relatively more static. Thus, norms, instincts, and habits shaped by the past were usually adaptive in the present. And, in turn, implicit beliefs and desires extracted from these sources were generally useful. Yet, today the world changes at a faster pace environmentally, technologically, and socially. This drives our instincts, norms, and habits out of date more quickly than ever before. (And while the commentaries do not mention it, it is also likely that our beliefs are truer, and capacity for reason keener, than ever before.) If natural selection struck a balance between non-rational systems and reasoning when rationalization evolved, perhaps it is the wrong balancing point today. This point is worth careful consideration. If indeed the balancing point of representational exchange is innate and shaped by biological natural selection, then it is hard to imagine that it is optimal today. Yet, as discussed in section R4, the balancing act between reason and intuition may be carried out online by metacontrol processes that are relatively more sensitive to present circumstances.
R7.4. Revision: The scope and challenge of representational exchange
This final group of commentaries offers two important revisions to the target article. First, they broaden the reach of the representational exchange to personal and collective self-understanding: The attempt to find meaning in our lives. The target article talked a lot about extracting information, but not at all about making meaning. Yet, the search for narrative, structure, and meaning in our lives is a perfect example of representational exchange. It may be a fiction to suppose that our individual or collective behavior already has meaning, as if organized by an intelligent agent with a specific plan in mind. But if it is a fiction, it is surely a useful one. We stand to learn a tremendous amount by assuming that the way each and all of us do things is, while not perfect, probably very good. Second, however, they sound a cautionary note even for self-benefitting rationalization. It may be adaptive, but is it good? They provide ample reasons for skepticism.
On their face, these commentaries seem to strike two discordant notes. On the one hand, representational exchange is at the heart of our search for meaning. On the other, it may often be a misguided search. Yet, in fact, these notes combine to play a familiar theme: the familiar struggle to balance reasoning against intuition – head against heart.
R8. The human problem
The Enlightenment bestowed a durable understanding of the human problem: To act rationally and, therefore, by reasoning. Summarizing this neoclassical spirit, Pinker (Reference Pinker2018, p. 8) writes: “If there's anything the Enlightenment thinkers had in common, it was an insistence that we energetically apply the standard of reason to understanding our world, and not fall back on generators of delusion like faith, dogma, revelation, authority, charisma, mysticism, divination, visions, gut feelings or the hermeneutic parsing of sacred texts.”
Although not our only understanding of the human problem, this is an influential one. It aligns with a strong tradition of psychological research that reifies reasoning, organizing it around two normative standards. First, one ought to establish beliefs by the rational application of Bayesian principles. Second, one ought to choose actions by expected value maximization. Against the background of these standards, the behavioral prescriptions of norms, instincts, and habits are defined as biases, mostly to be understood in light of their systematic deviations from reasoning. If they can be justified at all, it is either on the grounds that they are more computationally efficient (Griffiths et al. Reference Griffiths, Lieder and Goodman2015) or have indirect social benefits (Mercier & Sperber Reference Mercier and Sperber2011; Tedeschi et al. Reference Tedeschi, Schlenker and Bonoma1971; Von Hippel & Trivers Reference Von Hippel and Trivers2011). These ancillary benefits may compensate for what are otherwise strictly worse outcomes: mistaken beliefs, misguided desires, and maladaptive choices.
Thus, according to the strictest Enlightenment ideals, we must purge thought and action of the influence of non-rational systems. Pinker (Reference Pinker2018, pp. 8–9), again: “Thinkers such as Kant, Spinoza, Hobbs, Hume and Smith were inquisitive psychologists and all too aware of our irrational passions and foibles. They insisted that it was only by calling out the common sources of folly that we could hope to overcome them. The deliberate application of reason was necessary precisely because our common habits of thought are not particularly reasonable.” This is why we might characterize reasoning not as the human condition, but rather the human problem. It is a problem because, while reasoning is presumed to be superior, it also hard to do. If our highest human calling is self-improvement, then we must apply the force of will to achieve the purity of reasoned thinking.
In its purest form, this Enlightenment vision possesses the allure of simplicity, the veneer of logic, and the tragic myopia of self-certainty. Human reasoning is, after all, exquisitely fallible. Our beliefs are often incorrect and always incomplete. Our desires may internally inconsistent, interpersonally irreconcilable, and occasionally detrimental to our well-being. And our capacity for rational planning is severely constrained by time and cognitive bandwidth. None of this implies that reason is bad; to the contrary, it is indispensable. But it is fallible, and sometimes it can be improved by attending to non-rational adaptive mechanisms. Whether through the genetic inheritance of our instincts, the cultural inheritance of our norms, or the experiential inheritance of our habits, we are endowed with guides to action shaped by powerful forces of trial-and-error adaptation.
Thus, to act truly rationally cannot be to act only by reasoning. Rather, even if we accept the Enlightenment commitment to rational action – especially if we accept it – then the basic human problem is not to isolate reason from non-rational processes, but to improve it by considering them. This, of course, is Jason Bourne's problem: To glean from his programmed abilities whatever information serves his present interests. In this effort, rationalization is an essential tool. It is not a matter base self-enhancement or even one of elevated self-understanding. It is, at heart, a project of self-improvement – surely the highest human calling of all.
Target article
Rationalization is rational
Related commentaries (26)
Antecedent rationalization: Rationalization prior to action
Belief as a non-epistemic adaptive benefit
Cognitive dissonance processes serve an action-oriented adaptive function
Evidence for the rationalisation phenomenon is exaggerated
Ex ante coherence shifts
Hard domains, biased rationalizations, and unanswered empirical questions
Heroes of our own story: Self-image and rationalizing in thought experiments
Ideology, shared moral narratives, and the dark side of collective rationalization
Letting rationalizations out of the box
Means and ends of habitual action
Quantifying the prevalence and adaptiveness of behavioral rationalizations
Rational rationalization and System 2
Rationalization and self-sabotage
Rationalization and the status of folk psychology
Rationalization enables cooperation and cultural evolution
Rationalization in the pejorative sense: Cushman's account overlooks the scope and costs of rationalization
Rationalization is a suboptimal defense mechanism associated with clinical and forensic problems
Rationalization is irrational and self-serving, but useful
Rationalization is rare, reasoning is pervasive
Rationalization may improve predictability rather than accuracy
Rationalization of emotion is also rational
Rationalization: Why, when, and what for?
Rationalizations primarily serve reputation management, not decision making
The rationale of rationalization
The social function of rationalization: An identity perspective
What kind of rationalization is system justification?
Author response
Rationalization as representational exchange: Scope and mechanism