Algorithmic ingredients of bounded prosociality: Affective signals, empathy, and episodic simulation

Abstract

Humans learn, use tools, communicate, and often act prosocially (i.e., benefit others). Prosociality may or may not be the most uniquely human of the four, but it is likely the most deservingly bounded (e.g., qualified with “often”). We are often prosocial, but when we are not, the alternatives encompass everything from neglect to war. As “(…) the glory and scum of the universe” (Pascal, 1658, p. 42), humans inadvertently shape all of life’s facets with these opposing capacities. How do these seemingly incompatible capacities coexist? Functional accounts suggest that we are overwhelmingly prosocial toward members of our ingroup, while treating outgroup members with indifference or even hostility—especially if they are deemed a threat. Add flexible coalitional cognition that eagerly carves the world into “Us” versus “Them”, and we begin to see how the same person who teaches trigonometry to some could trigger trebuchets against others. In combination, this view provides an account for why we relate to others the way we do, along with a coarse outline of how. The outline is coarse because it does not specify steps that may plausibly lie between distinguishing friend from foe and giving or taking resources accordingly. For example, we frequently witness or imagine others’ experiences, which evokes emotions that influence how we relate to those others. In principle, we would expect any species-defining function, such as bounded prosociality, to be reflected in how we use the tools at our disposal, including imagination and affect. This dissertation aims to unpack the algorithmic ingredients of prosociality by probing three interrelated candidate tools—affective learning signals, empathy, and episodic simulation—in brain and behavior. Part 1 (Vollberg & Cikara, under review) asks how humans may learn to sometimes aggress despite being predominantly prosocial and harm-averse. When scientists think of learning in organisms and machines, they typically think of updating from differences between the value of expected and actual outcomes (prediction errors). The underlying valuations are most tractable when based on external rewards (e.g., money, food, or points). But (aggressive) behaviors can also provide internal rewards, which, indexed via affect, may constitute a separable and complementary learning signal. Participants persisted in behaviors that felt better than expected, especially when those behaviors were aggressive. When participants were given the opportunity to empathize with members of their opponent’s group, they de-escalated aggressive behavior that had felt better than expected, but only if they liked the opponent’s group. Part 2 (Vollberg et al., 2021) asks whether group-dependent differences in empathy reflect differences in how detailed we imagine the scene surrounding another’s experience (Vollberg & Cikara, 2018). Memory researchers have long developed manipulations that are thought to upregulate episodic simulation and scene imagery. Borrowing those manipulations to probe effects on empathy showed increased empathy across the board, but left differences between liked and disliked groups unchanged. Part 3 (Vollberg et al., in preparation) asks whether empathy can be attributed specifically to how people represent a given scene, separated from the person in that scene, and whether such representations link to prosocial behavior beyond empathy ratings. Functional MRI revealed that measures of brain activation in response to familiar places were just as good or better than their person-related counterpart at predicting both how bad participants felt with the target of a misfortune and how much money they would give to that target. Moreover, scene-related measures of brain activation were more pronounced for liked targets, regardless of how familiar participants were with the scene. Taken together, these findings suggest that we recruit more detailed scene memories when imagining the misfortunes of liked others; that scene representation influences empathy and prosocial behavior; and that group-dependent empathy influences how we learn from (i.e., de-escalate) aggressive behavior that felt better than expected. Using a range of approaches on large samples, the work presented here thus begins to uncover how our affective and cognitive architecture may serve the computational goal of bounded prosociality.