Abstract
Recent decades have witnessed a sea change in thinking about emotion, which has gone from being seen as a disruptive force in human thought and action to being seen as an important source of situation- and goal-relevant information and evaluation, continuous with perception and cognition. Here I argue on philosophical and empirical grounds that the role of emotion in contributing to our ability to respond to reasons for action runs deeper still: The affective system is at the core of the process of evaluatively modeling situations, actions, and outcomes, which is the foundation upon which rational deliberation and action can be built. Taking up this perspective affords new approaches to long-standing problems in the theory of reason-based action.
Keywords
Introduction: Two Uses of “Reason”
In English, the term “reason” is used in both a broad and a narrow sense when explaining individual human action. In the broad, causal-explanatory sense, we give the reason why someone acts as she does by citing causal factors or conditions that played a role in the genesis of the action. In the narrow, putative-rationale sense, we focus on those factors that the agent implicitly or explicitly took to speak in favor of, or justify, the action.
Thus, if you answer my question “What was the reason why Mike lashed out at Nora—just when he needed her help?” by saying, “Mike was frustrated at failing to finish the project on time, and it got the best of him,” you are giving a causal-explanatory reason that does not show the behavior to have a rationale, even in Mike’s eyes. By contrast, if you answer the question, “Why was Nora’s response to Mike’s outburst so calm?” by saying, “She could see he was frustrated, and so tried to find a way to help that didn’t just upset him more,” you are giving her rationale: Nora, you are claiming, implicitly or explicitly took herself to have grounds for responding calmly to Mike, and acted accordingly. An agent’s rationale need not be a good reason—the question concerns how the agent saw her situation, options, and action, whether or not she was seeing things correctly.
Causal-explanatory and putative-rationale reasons are not mutually exclusive. Nora will not have acted “for the reason” of calming and helping Mike unless this rationale really did figure in the causation of her action. Otherwise, the putative rationale might be no more than a rationalization, hers or ours. Suppose that a personnel manager later asks Mike why he lashed out at Nora, and Mike says, “I needed to let her know that this was the last straw—she’s been obstructive all along.” This thought, however, might never have entered Mike’s mind—implicitly or explicitly—until the very moment he is asked to explain his behavior. Mike was greatly stressed by the impending deadline, and would have lashed out in the same way at anyone who approached him in the circumstances. A possible putative rationale, therefore, needs to play some appropriate causal-explanatory role if it is to be an actual rationale for the action.
The task of analyzing what it is for a rationale to explain a bit of behavior so that it can be shown to be an action done for a reason is a long-standing problem in philosophy, but the need for a solution extends well beyond philosophy. A clinical psychologist or psychiatrist must be able to determine whether the rationales offered by a patient explain his behavior or whether they are mere (perhaps self-deceptive) rationalizations. This distinction is also central to our thinking about intentional action—an agent’s professed rationale for acting, however plausible it might seem and however much the agent himself might be convinced of it, was not his actual intention if it played no role in bringing the action about. And people’s intentions are of course of central interest to us in interpersonal relations, self-scrutiny, ethics, law, and politics.
In this article, I will first identify four philosophical problems for understanding how it is possible for a putative rationale to play the sort of causal-explanatory role that makes an act be done for that reason. To tackle these problems we will look first to control theory, and the role of models in the regulation of behavior. This, in turn, will give us a starting point for a new perspective on how to understand reason-based guidance of behavior. In this perspective, the agent’s core evaluative system—the affect and reward system—plays a vital, organizing role. In the end, we will find a striking convergence of a priori philosophical considerations and a posteriori empirical developments in psychology, cognitive science, and affective neuroscience. What is novel in this frankly speculative synthesis is not the idea that affect can make some contribution to acting for reasons—it is well known, for example, that affect can furnish certain kinds of information or impetus to rational decision-making (Schwarz & Clore, 2003). Rather, it is the idea that the affective system is an indispensable part of rational decision-making and action, even when these take self-conscious, deliberative forms. 1
Four Problems About What It Is to Act for a Reason
People act for reasons, we believe—in matters big and small, and when self-conscious or spontaneous. Despite the ubiquity of acting for reasons, it has proven remarkably difficult to say what this involves. At least four basic, interrelated problems stand in the way. Let’s illustrate these problems using a series of hypothetical scenarios.
First scenario
While Ada has been out of town, a cat-sitter has been staying in her apartment. During that time, Bruce, her cat, has taken to lounging in the narrow front entryway. When Ada returns, preoccupied by the work that has accumulated in her absence, she sees Bruce lying there, and suddenly stops rushing around. An image passes through her mind of someone entering the hallway, not seeing Bruce, and treading heavily upon him—hurting Bruce and perhaps also stumbling and falling as a result. The idea of using the water spritzer she keeps in the kitchen—her normal way of deterring Bruce—comes immediately to mind. But just as she is about to head for the kitchen she stops again, since it occurs to her that Bruce might not readily associate the spritzing with his new lounging habit. She pauses, casting about for ideas, and the thought percolates up of simulating the accident she hopes to prevent. She could casually walk into the entryway, in full sight of Bruce but not looking at him, and, in passing, apply her foot firmly enough to Bruce’s tail to get the message across of the risk he is courting by lounging there. She’s at first a bit surprised by this thought, since she goes to great lengths not to cause Bruce any harm. But she thinks she can control her foot pressure to prevent any real harm and he really does need to learn this lesson as quickly as possible. So Ada proceeds to go through with the staged accident, and Bruce’s sharp protest and rapid exit suggest that he might have gotten the point. Whether or not we see Ada’s rationale for deliberately applying pressure with her foot to Bruce’s tail as adequate—whether or not it is a good or sufficient rationale for what she’s done—we can see that she did have this idea in mind, and that it played a predominant role in generating and guiding her action. So we say that she acted for that reason.
Second scenario
This scenario is the same as the first up to the point at which Ada, still standing in the entryway, is mentally casting about for a course of action. This time, Bruce has silently shifted position and now lies nearly at her feet. Ada, absorbed in thought, has failed to notice this, and as soon as she settles upon “staging” an accident, turns, and takes a first step in the direction of Bruce’s previous location, her foot lands squarely on Bruce’s tail, sending him scurrying away, pained, alarmed, and dissuaded from lounging in the entryway. Ada herself jumps back in surprise and alarm, almost toppling over.
The Deviant Causal Chain Problem
The second scenario shows the inadequacy of an account of acting for a reason according to which it suffices if the agent’s rationale plays a causal role in generating rationale-congruent behavior. Ada’s idea of applying pressure to Bruce’s tail to discourage his lounging in the entryway was indeed a chief cause of her turning and starting to walk in (what she took to be) Bruce’s direction. And this led directly to the act of stepping on Bruce’s tail, thereby dissuading him and achieving her aim. Yet we would not say that this step itself was taken under the idea of putting pressure on Bruce’s tail. As Donald Davidson argued in a seminal article (1963), examples like this draw attention to the many paths through which a rationale can cause a rationale-congruent act, not all of which count as acting for that reason. So this has come to be called the “deviant causal chain” problem.
The Agency-Without-Regress Problem
Some philosophers have argued that the chief deficiency of merely causal accounts of action is “leaving the agent out”—we see a succession of thoughts and behavior, but where is the agent herself? Christine Korsgaard has written,
[T]o will an end is not just to cause it, or even to allow an impulse in me to operate as its cause, but, so to speak, to consciously pick up the reins, and make myself the cause of the end. (2008, p. 59)
There is something right about this. While Ada’s stepping on Bruce’s tail in the second scenario was the causal result of her beliefs and aims, still, her actual stepping upon Bruce’s tail was in some sense not Ada’s doing—it surprised her, since she was not at the moment “steering” her foot onto his tail.
We cannot, however, rest content with this explanation. “Picking up the reins” would itself be an action done for a reason, and so we are at risk of launching a regress. Strange as it might sound, we must explain acting for a reason in terms of processes occurring within an individual that are not themselves actions done for a reason, but which can add up to the agent acting for a reason. This is the agency-without-regress problem.
The Problem of Nondeliberative Attunement to Reasons
For such agency to be possible, the action-constituting processes must meet three conditions: (a) they must somehow represent that reason (the “idea” under which the agent acts); (b) they must themselves be sensitive to this reason or what it requires; and (c) they must work together in an organized manner such that they constitute an exertion of controlled activity under the idea of that reason. Call these conditions for the attunement of her thought and action to the reason in question. The problem of regress tells us that we must be able to see how such attunement is effected without an additional, inner exercise of agency to do it.
But surely, some will protest, it is possible to initiate an action by other actions without launching a regress. One can focus on a problem, deliberate about how to solve it, decide which option one will adopt, and initiate action accordingly—all of which are mental actions—without need to posit an additional inner action or agent to organize and oversee all this. Indeed, this sequence of mental actions is the canonical form of “full-fledged” action for a reason.
That is true, yet potentially misleading. Suppose we say that an act is performed by an agent, and for her reasons, because it was the product of these mental acts of focusing, deliberating, deciding, and initiating. But then, why are these the agent’s actions, done for her reasons? To avoid a regress, we must say that such mental acts can be attributed to an agent as done by her, and for her reasons, even when they are continuous, first-order psychic processes involving no distinguished agential intervention. This means that she acts via this organized activity, not upon it.
Let us return to the first scenario, attending closely to the sequence of events. Ada has just returned to her apartment. Despite her distractions, she notices Bruce’s new lounging spot, stopping her in her tracks, and starting her thinking about the risks Bruce is running and what she can do to eliminate them. Her thinking draws selectively upon memory and imagination in ways oriented by her goal, yielding a series of thoughts about how to dissuade Bruce, though her thinking doesn’t settle upon one until it seems to her likely to be both effective and reasonably humane. Once she does see such a solution, her sense of the risk to Bruce gives urgency to the situation, and motivates her to act directly.
Ada thus portrayed seems recognizable as an agent acting for a reason, not a mere succession of psychic impulses. Yet each of these steps takes place nondeliberatively, without anyone “picking up the reins” or deliberating and deciding about whether to take them. Ada’s noticing Bruce’s new lounging spot and its possible risks arise from her perceptual system’s attentiveness to novelty and sensitivity to her values (i.e., Bruce’s welfare). By its nature, this kind of “steering” of attention cannot require agential intervention, since the anomalous situation must first be noticed before we can deliberate about it. Similarly, her values and goals prime memory recall, since relevant memories cannot be deliberatively called to mind unless the person is already aware of them. But how can values and goals implicitly guide attention, and prime memory and deliberation? Call this the problem of how such organized, nondeliberative attunement to reasons is possible.
The Rationalization Problem
Consider now a third scenario: This scenario is just like the second, except that Ada’s peripheral vision does pick up Bruce’s shift in position—the information enters her mind, though it is not consciously noted. Moreover, this information engages a strong, implicit motivation to punish Bruce. Bruce, it seems, has become very difficult in recent years, and Ada has accumulated a good deal of resentment, though without being aware of it. As she stands in the entryway, her mind casting about for ways of dissuading Bruce, her memory and imagination are primed to suggest actions that involve direct exertion of painful physical force upon Bruce. And once such ideas come to mind, her implicit anger also makes them seem more eligible that her usual, tamer ways of dissuading Bruce. Consciously, she would not allow herself to harm Bruce from malice, but her implicit anger nonetheless shaped her decision around a punitive goal, resulting in a tread that causes Bruce more than necessary pain, and surprises her conscious self.
Asked to explain why she stepped so hard on Bruce’s tail, Ada will say she had no idea this would happen—she was trying to minimize pain to Bruce, and the hard treading was an accident. And yet, it was a bit of goal-directed behavior, zeroing in on its target. Decades of research in social psychology attest to how easily our conscious selves can be mistaken about what we did and why we did it, and how readily we can construct a rationale post hoc (Haidt, 2001; Nisbett & Wilson, 1977). Ada, who takes herself always to have Bruce’s well-being uppermost in her mind, explains her behavior in a way consistent with this self-image. And yet her treading on his tail was done for a reason. Do we therefore deny that this was Ada’s reason, because she did not—and would not—consciously endorse it? That would identify Ada with only a fraction of her mind, and naively accord her self-narrative too much authority with respect to who she really is and why she really acts.
Once we recognize that conscious agency—even when the action in question is the explicit endorsement of a candidate act—is always exerted in league with the nondeliberative attunement supplied by implicit mind, we also must recognize that agents lack full insight into the source and nature of the reasons for which they act. But even if it is clear enough that Ada’s treading hard on Bruce’s tail was purposeful in the third scenario in a way that it was not in the second (i.e., the treading was her doing in an important sense, a manifestation of her capacity to act upon the world in light of her goals and not an accident that “happened to her”), still, how do we explain the working of such nondeliberative attunement to reasons? And crucially, how do we explain this in a way that still leaves something to the distinction between a genuine rationale as the idea “under which” an agent acts, and not a mere cause of the behavior?
Reasons and Regulation
The place to begin our answer, I believe, is control theory, and, especially, the theory of controlled movement. Control theory is concerned with the regulation of a system’s behavior in light of some target value, goal-to-be-realized, or task-to-be-achieved. A thermostat, for example, does not merely cause changes in the operation of the heating and cooling system, it regulates these changes in order to reach and maintain a target temperature. What is of special interest to us is the nature of the process by which this happens. A thermostat’s sensor detects whether room temperature is above or below the target value, and sends a feedback signal to the regulator, which in turn sends a command to the system—a feedforward signal—to activate cooling or heating until the sensor detects no discrepancy with the target value. The sensor, in effect, issues “error signals,” which the thermostat is designed to reduce. This enables thermostatic regulation to be “self-tending” in the sense of not requiring external intervention to achieve a fairly constant temperature in the room.
A control system of this kind is, however, not very smart. It is very likely to overshoot when heating or cooling, generating an error signal that leads to turning on cooling or heating, respectively, to compensate. This oscillation as the system fails to “settle” promptly on the target value wastes energy. The thermostatic regulator could thus be more efficient if it could use sensor input to “predict” when the temperature is nearing the target value, and gradually reduce the rate of heating or cooling to approach the target value smoothly. Such a thermostat would also be more effective, in the sense that the room would spend more time at the target value. A capacity to anticipate thus can enhance both efficacy and efficiency.
The thermostat would be yet more effective and efficient if it could also anticipate when the room’s occupants are likely to adjust the target temperature up or down. Then it could gradually initiate heating or cooling, respectively, ahead of time, preparing the room and reducing the load on the system. To gain such anticipatory capacity without requiring “supervision,” the thermostat will need to learn from experience about occupants’ patterns of use. Interestingly, it can use a feedforward/feedback, error-reduction mechanism to drive this learning, too. If, after 5 weekdays, the thermostat “predicts” that the temperature will be turned up at 7:00 am, then it will generate an error when this does not happen on Saturday and Sunday. Iterations of this learning mechanism will build a “model” of the typical week’s pattern of temperature settings by the room’s occupants, and the more accurate this model, the more effectively and efficiently the heating and cooling system will run.
Early in the evolution of contemporary control theory, Roger Conant and Ross Ashby (1970) proved what is called the “good regulator” theorem: under some very broad assumptions, a maximally effective and efficient regulator for a system must build a model of that system. A full model includes the actual and possible states of the environment and the system, and, relative to these states, the actions available to the system and their consequences for the system’s target values. Conant and Ashby did not hesitate to draw a “corollary” to their theorem: “the living brain, so far as it is going to be successful and efficient as a regulator for survival, must proceed, in learning, by the formation of a model (or models) of its environment” (1970, p. 89).
It has taken three or four decades, and the emergence of a host of new brain imagining techniques, but neuroscience is now amply confirming Conant and Ashby’s (1970) claim. We now have impressive evidence from direct neuronal recordings in mice, rats, and monkeys of the formation and utilization of models. Rats, for example, build up relational and grid-like spatial representations of their environment, which work in a coordinated way and are activated not only during exploration but in resting and sleep states, during which gaps in experience are interpolated and novel paths are simulated (Gupta et al., 2010; Ji & Wilson, 2007). When a rat is at a choice point in running a maze, activation in its “mental map” sweeps out alternate paths lying ahead, and neuronal firing rates reflect acquired expectations about the availability of rewards on those paths (Johnson & Redish, 2007; Johnson, van der Meer, & Redish, 2007).
Evolution appears to have produced a striking convergence between actual learning and decision-making processes in foraging mammals with a priori norms of rational learning and decision-making. With increasing probability as experience grows in range and diversity, intelligent animals and humans tend to approximate optimal learning and behavior relative to the information available, the options for actions they possess, and the goals they seek. Experimental confirmation comes from four sources. First, foraging animals and humans placed in a novel environment containing a diverse array of potential benefits, costs, and risks, tend to develop nearly optimal foraging patterns when allowed to explore the environment freely (Dugatkin, 2004; Kolling, Behrens, Mars, & Rushworth, 2012; Shettleworth et al., 1988). Second, detailed study of the learning exhibited by the perceptual system and motor system, and of reinforcement learning generally, shows that they exhibit many of the patterns predicted by Bayesian hierarchical causal modeling (Courville, Daw, & Touretzky, 2008). Third, neural recordings taken during animal learning and decision-making tasks, and similar brain-imaging studies with humans, show patterns of activation that closely approximate Bayesian updating of probabilities and rewards, with separate computation of expected value and risk, and corresponding effects on actual choice behavior (Schultz, 2010; Tobler, Dougherty, Dolan, & Schultz, 2006). And fourth, optimization models of motor control yield patterns of behavior strongly similar to actual movements (Todorov & Jordan, 2002). Perhaps it should not be surprising that millions upon millions of generations of natural selection for effectiveness and efficiency have yielded brains with a mental architecture of the kind that philosophers, mathematicians, cognitive scientists, and electrical engineers have reasoned to be optimal.
Regulation and Affect
What has any of this to do with emotion? The short answer is that the affect and reward system—affective system, for short—is the central locus of the learning processes, evaluative representations, and spatial mapping and simulation essential for the reasons-sensitive action guidance just described. The affective system has come to be understood as encompassing not only the “limbic system” (amygdala, thalamus, hypothalamus, hippocampus, and cingulate gyrus) but also other closely integrated areas involved in the processing of emotion and reward (ventromedial prefrontal cortex, orbitofrontal cortex, ventral striatum, insula, and basal ganglia). Two other closely integrated structures, the lateral prefrontal cortex (Wallis & Kennerley, 2010) and cerebellum (Ramnani, 2006), appear to play a key role in the linking of affect to attention and the planning, execution, and monitoring of action. The once influential “affect versus cognition” distinction is fading and is being replaced with the view that highly integrated cognitive-affective networks and “hubs” do the critical work of guiding intelligent interaction with the world to meet needs and fulfill goals (Pessoa, 2008)—that is, the task of a “good regulator.”
The terminology used for this system is of less importance in the present context than the fact that it is present in well-developed form in a number of our animal ancestors, and so does not require the operation of self-conscious, declarative cognition of the kind found in humans in order to attune attention, memory, inference, motivation, and choice to the individual’s needs and goals in normatively appropriate ways. Thus, this system can serve in humans as a core evaluative “hub” capable of implementing the nondeliberative attunement to reasons for action we needed to find in order to make possible conscious deliberation and action for reasons without a regress of agential control.
Why might it be that the affective system, broadly understood, plays this central role? Affect has long been associated with evaluation or appraisal (Deonna & Teroni, 2012; de Sousa, 1987; Ellsworth, 1991; Frijda, 1986, 1993; Goldie, 2000; Lazarus, 1991). Increasingly, affect is seen as a “common currency” for representing value in the brain, which can include evidential or epistemic value (e.g., uncertainty and risk) as well as practical or instrumental value (e.g., reward and punishment), and social value (e.g., trust and cooperation; see Behrens, Hunt, & Rushworth, 2009; Rolls & Grabenhorst, 2008; Storbeck & Clore, 2007). Formally speaking, affect has many of the features necessary for the representation of value: It varies in degree and intensity; it can take a wide range of intentional objects (e.g., one can be afraid of a situation, person, action, or risk); it possesses positive or negative valence and can produce such value-relevant behaviors as approach or avoidance, and acceptance or rejection; it exhibits a spectrum of forms that tend to be responsive to distinctive dimensions of assessment (e.g., fear is responsive to danger and risk; anger is responsive to personal contravention or insult; surprise is responsive to violated expectations; longer term moods like a sense of subjective well-being or ill-being are responsive to whether one is making progress in realizing one’s goals; Lawrence, Carver, & Scheier, 2002); and, finally, these varied forms tend to influence thought and action in ways reflective of these evaluative dimensions (e.g., fear focuses attention, primes danger-avoidant thought, and ramps up action-readiness; surprise modulates expectation and uncertainty; and subjective well-being influences orientation toward holding vs. changing one’s course of action).
The natural and social worlds present us with competing values—trust in others can enhance beneficial coordination but also render one vulnerable to exploitation; taking advantage of opportunities often goes with making risky choices; meeting future needs may require forestalling present appetition. The multidimensionality of the affective system enables it simultaneously to represent these competing values and associated risks, while also serving as a shared substrate in which they can be compared for magnitude, intensity, and urgency. Moreover, the high degree of connectedness of the affective system means that these evaluative representations can orchestrate attention, memory, imagination, and action-readiness in a manner that reflects these trade-offs—while also supporting “unsupervised” or “spontaneous” recalibration of these trade-offs as our evaluative expectations are continuously tested against the actual outcomes of action through feedforward/feedback learning.
Prefiguring and Underwriting: The Affective System and Consciousness
In a curious way, the features we have attributed to the affective system prefigure features that have recently been used to characterize the distinctive character of the conscious mind. Through simulation and evaluation, the affective system frees the mind from bondage to the current stimulus to permit the consideration of as-yet-unrealized possibilities. By receiving input from the various senses, interoception, memory, and association, the affective system can synthesize multimodal, “global” evaluations, and given its high connectedness, these global evaluations can be “widely broadcast” to guide action in a coordinated way (for discussion of comparable features of consciousness, see Baars, Franklin, & Ramsoy, 2013; Dehaene & Naccache, 2001; Dennett, 2001; Greenwald & Liu, 1985; Smallwood & Schooler, 2006). Perception, memory, association, evaluation, simulation, and decision in fact make a natural functional cluster, whether conscious or unconscious, since they all contribute essential components of intelligent control. Indeed, recent study of the primate and human brains’ “default system” (i.e., the network of nonconscious activation that increases as task demands are lessened) suggests that the principal function of this coordinated, largely nonconscious brain activity is prospection, the simulation and assessment of possible actions and outcomes (Buckner & Carroll, 2007; Seligman et al. (2013); Spreng, Mar, & Kim, 2009).
For a long time, it was plausible to assume that such features were the exclusive domain of consciousness—implicit responses were seen as the realm of stimulus-bound, “automatic” or “instinctive” responses, yielding cue-specific, stereotypic, temporally immediate, statistically untutored behavior. As we have seen, however, even in the case of reinforcement learning (Rescorla, 1988), the underlying processes appear to be representational, expectation-based, and sensitive to statistical patterns in experience. The speed of response of the affective system—beginning to encode incoming information evaluatively within 50–100 milliseconds—should not be confused with its simplicity, rigidity, or reflexivity (see also Moors, 2010; Snow, 2006).
This leaves us with an interesting thought. The broad affective system might be nature’s first version of the sort of representation-mediated, multimodal, unified prospective guidance of thought and action once thought to be peculiar to conscious deliberation and choice. The “feelings” or “intuitions” of the conscious mind, which play such a large role in how we think and act, but whose origins are often fairly opaque to us, might be summary outputs of the core evaluative system (see Bechara, Damasio, Tranel, & Damasio, 1997; Craig, 2009). The full statistical models generating these feelings would be much too complex to represent in consciousness, with its narrow “bandwidth” and limited working memory. But conscious representations of these models might not be necessary if conscious feeling can present the relevant information in a summary way—as degrees of confidence or uncertainty; strengths of desire or aversion; or senses of interest, surprise, disappointment, hope, frustration, and well-being (see, e.g., Carver, 2015; Craig, 2009; Lawrence et al., 2002). What we might deem a “gut feeling” or “intuition” that a particular idea is promising, a given person untrustworthy, or a recent event bad news, might be a great deal more than a visceral response or “animal instinct”—it might be a “felt” summation of complex, model-based information and simulations, beyond the conscious mind’s capacity to process on its own.
Statistical learning systems, however, have limitations of their own. Critically, updating does not generate new concepts or unprecedented hypotheses. The distinctive contribution of conscious thought thus might be to introduce novel ideas that can yield discontinuous revision of prior expectations and more or less dramatic changes in behavior. For example, while evolution will tend to favor animals whose “hypothesis space” is well-suited to their environment and capabilities, if rapid environmental change or other existential threats occur, and quite novel behaviors are required, then some break from the hold of incremental fine-tuning could be indispensable—an effortful, top-down process (for discussion of the role of consciousness in enabling novel thought and action, see Kiefer, 2012; Kunde, Reuss, & Kiesel, 2012; Mudrik, Faivre, & Koch, 2014). Such a gain in flexibility might be worth the metabolic costs of something as complex as human reflective, language-infused, deliberative consciousness—witness the ability of humans to adapt rapidly to the most diverse environments, while, sadly, many of our closest intelligent animal relatives are threatened with extinction through loss of habitat.
The Four Problems Revisited
The Problem of Nondeliberative Attunement to Reasons for Action
We have just seen the case for thinking that the affective system is suited for providing the nondeliberative attunement to reasons is necessary for agents to be responsive to reasons. First, given its location early in the perceptual stream, the affective system can orient attention and evaluate percepts for urgency and valence even before conscious thought has interpreted them. In addition, it can prime memory and potentiate inference in situationally and goal-relevant ways so that deliberation can be productive. Finally, it can continuously assess the success of behavior and adjust expectations and motivation accordingly, so that monitoring need not burden consciousness when things are going to plan, and can cue conscious deliberation when they are not.
The Deviant Causal Chain and Rationalization Problems
If we look at activity in a rat’s affect and reward system of the brain as it learns a maze and the distribution of valued outcomes, we see the emergence of patterns that appear to constitute a spatial and evaluative model of its environment. As the animal subsequently runs the maze, we see activation in these same areas appearing to function prospectively to simulate and assign expected value and risk to potential future trajectories. This affords us a “window” into which considerations are actually guiding the rat’s choice, and how. Because the rat’s evaluative model of the maze’s alternate pathways participates directly and dynamically in the control and monitoring of its motor behaviors, we see a representation of its “reason” for making the choice it does—the “idea,” so to speak, under which it acts. Indeed, in recent research it has been possible to construct in detail classical marginal utility functions and risk preferences for monkeys (Stauffer, Lak, & Schultz, 2014).
Suppose now that we could similarly peer into Ada’s mental activity in the different scenarios. In the first scenario, when Ada first encounters Bruce in his new lounging spot, her model of the entryway and its possibilities is updated to raise dramatically the expectation of risk to Bruce of certain pathways through it—a representation that reflects Ada’s general attentiveness to risk, and encodes her urgent concern over Bruce’s well-being. As she deliberates, we see activation of this model of the situation along with stored representations of risk-related information and her model of Bruce himself. This “test bed” permits empathic simulation of how various ways of discouraging Bruce might appear to him, and how this would affect their likely success. And as evaluations of these simulations are compared and she settles on “enacting” an accident, we see how this idea plays a forward role in generating both her behavior and the expectations that accompany it, and permit discrepancy-based feedback to guide subsequent action. Here, too, we see the idea “under which” the agent is acting.
Compare the second scenario, in which Bruce has moved close to Ada while she is forming her idea of what to do. Peering into her mind, we can see that her model of the situation is not updated to include Bruce’s new location, so this fact generates no expectation that her next step will land squarely on Bruce’s tail. Pain to Bruce is uniformly represented as a disvalue, and thus, even as Ada takes the step that treads hard on Bruce’s tail, such treading is represented having low probability and high negative expected value. Thus, her treading generates a strong “error signal,” attesting to the fact that this exertion of force “wasn’t her idea.” Even though Ada’s treading on Bruce’s tail was caused by, and achieved, her idea of dissuading him, Ada feels remorse, not satisfaction.
Now consider the third scenario. Even though Ada did not consciously notice Bruce’s change in position, it was registered nonconsciously, updating her internal model of the situation. Owing to her long-simmering anger at Bruce, this model assigned some positive value to delivering a harsh punishment to Bruce, and this primed memory and imagination to suggest responses that would cause Bruce pain, and made them more attractive as they came to mind. Such influence on thought and decision by unconscious evaluation—even when it is at odds with one’s conscious evaluation—is well-studied in the literature on “implicit bias” (Dasgupta, 2004). Because the same implicit evaluative model guides Ada’s motor behavior, we can see that her stepping hard on Bruce’s tail, while surprising to her conscious self, was the result of an implicit positive expected value. We thus have the distinction we need between Ada’s treading hard on Bruce’s tail being done by her and for a reason of hers (Scenario 3), and Ada’s treading hard being at odds with her fine-grained evaluative control of action on behalf of her reason (Scenario 2). Ada herself will offer the same explanation of her behavior in both cases—that the excess force exerted on Bruce’s tail was entirely accidental—yet we can see that this rationale actually captures why she acted as she did in the second scenario, but constitutes a post hoc rationalization in the third scenario.
The Agency-Without-Regress Problem
In the first scenario, I have argued, Ada acts in a manner that expresses and operates through her evaluative representation of the situation. We see this throughout the processes culminating in the action—in what she notices, what factors she considers, how she weighs them, and how she settles upon an option and carries it out. Does the fact that nondeliberative processes cued her attention, primed her memory, potentiated her imagination and empathy, supplied her “intuitive” decision-weights, and monitored her action, preempt her own proper agency? This question invites us to conceive of Ada as something apart from her evaluative point of view, launching a regress. And it would identify Ada with a fraction of herself incapable of acting for a reason on its own—and thus not with an agent at all.
We said at the outset that explanations in terms of putative rationales have a structure that distinguishes them from other causal explanations and makes them into an exercise of agency “under an idea” of something that the agent in some sense “saw” in the action. Equipped with the notion of affective model-based control of action, we can now see more clearly what this structure is, and why it enables action to be expressive of the agent’s evaluative standpoint and be attributable to her, even when the agent’s self-knowledge is incomplete. 2
Conclusion
The capacity of the affective system to encode and learn about values and risks, and thus to furnish well-attuned decision-weights in the models that guide action, helps us to begin to solve four crucial problems for explaining what it is to act for a reason. This would be no surprise to Aristotle (1999), who consistently emphasized that acquired, well-calibrated affect is indispensable to virtuous action, and gave the affective system a core role in answering a question he posed two millennia ago: How is it possible that ideas can produce action that is normatively appropriate to them—how, in other words, can reason take a practical form?
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.