What is computational social cognition?

Our lab combines computational modeling with behavioral and neuroimaging methods to study the formation and expression of prejudice, intergroup bias, attitudes, and trait impressions (Hackel & Amodio, 2018). Our modeling approach begins by developing a theoretical model about the cognitive processes that produce a response—for example, how a preference for one social group over another forms across repeated experiences with group members. This model is then expressed formally, as a computational model that generates expected patterns of outcomes. By testing the degree to which participants’ experimental task behavior fits the model, we can identify specific processes (e.g., forms of learning, decision biases, and interactions among processes) that drive aspects of social cognition. In much of our work, we pair this computational approach with a Memory Systems Model of social cognition, which identifies distinct systems of learning that operate, often in concert, to support impression formation and social behavior (Amodio, 2019; 2025).

Instrumental social learning

We have used computational modeling to examine an under-studied aspect of social cognition: how we form impressions and attitudes through direct social interaction. In a series of studies, we show that impressions may be formed through mechanisms of reward reinforcement learning in direct social interactions—that is, through the process of acting toward a person and learning from their feedback—a first in research on impression formation (Hackel et al., 2015).

We further find that when humans learn from another person’s feedback, they infer not just the reward value of an interaction partner (as would be suggested by conventional reinforcement learning models) but also the high-level trait characteristic associated with the feedback (e.g., one’s generosity). We show this through the combination of experimental task design, in which interaction partner’s reward feedback is orthogonal to their generosity, and the formalization of computational models that specify learning from reward, traits, or a combination of both. Across all studies, we show that while humans tend to rely more on inferences of traits over rewards when forming an impression, they typically infer both kinds of information simultaneously. Furthermore, using fMRI, we show that instrumental learning about rewards and traits involves distinct computations in the striatum—a neural structure that computes prediction errors—and that representations of reward and trait values are integrated in the vmPFC to guide future social decisions (Hackel et al., 2015).

In other research, we have studied how reward-based and trait-based inferences function differently in different contexts to flexibly guide social behavior (Hackel et al., 2022), and how trait-based reinforcement learning is enhanced for social agents but not unique to them (Hackel et al., 2020). Finally, in a set of behavioral experiments, we show that instrumental learning from individual members of a group contributes to higher-level representations of  group-level preferences (i.e., prejudices; Hackel et al., 2022). In this work, our

Instrumental learning of prejudice

We have also used computational modeling to understand how stereotypes and racial identity shape the way we learn about people, often in implicit and difficult to control ways. In one set of studies, we show that mere knowledge of a stereotype biases the way one learns about group members in direct interactions, via instrumental learning (Schultner et al., 2024a). Using computational modeling, we show that stereotypes act as priors, setting expectations for how a group member will act, and then also lead people to update their learning about individual people in terms of their group memberships (i.e., via separate group learning rates). We further show that this effect of stereotype exposure on learning is implicit, and even occurs when people explicitly try to ignore the stereotype! In other work, we show that this stereotype effect on learning is amplified when the stereotypes have moral content (Rösler et al., 2025), and we find similar effects on learning induced by a person’s race (Traast et al., 2024) or ethnicity (Traast et al., 2025).

Across studies, we find support for the same model of learning—in which stereotypes or race induce both a group-based prior and separate group-based learning rates—supporting the idea that stereotypes and racial identity bias the way we learn about people, leading to the formation of an internalized prejudice. In our current work, we are  testing interventions to block this effect and extending our models to explain effects of misinformation.

Observational learning of prejudice

We have also used a computational approach to understand how people acquire prejudice through  social learning—that is, by passively observing an intergroup interaction. In a series of experiments, we find that by merely observing interactions between a prejudiced actor and social group members, observers acquired the actor’s degree of prejudice and then later expressed it in their own intergroup behavior (Schultner et al., 2024b). Moreover, observers were unaware of the actors’ bias, and they (mis)attributed their acquired prejudiced attitudes  to the behavior of group members. Computational modeling revealed that observers formed their impressions based on two sources of information—the actors choices and the group members’ feedback—but confused these when expressing explicit preferences. These findings identify social learning as a potent mechanism of prejudice formation that operates implicitly and supports the transmission of intergroup bias.