Pietraszewski presents a computational theory of how people solve the important problem of figuring out what constitutes a social group in the context of conflict. I agree that past definitions of “groups” are in many cases circular and insufficient (Cikara & Van Bavel, Reference Cikara and Van Bavel2014). We are long overdue for a high-level formalization of how people solve the “group” problem and I am grateful to Pietraszewski for initiating this conversation. In my view, however, the target article's theory faces two major challenges. First, though Pietraszewski speculates about non-conflict-based group representations, the theory still falls short of offering a unified account of group representations across different coalitional contexts (e.g., when agents are not in conflict). Second, though the stated goal of this paper is to identify a tractable information-processing problem, but not yet task analysis (let alone empirical tests of the theory), this does not preclude Pietraszewski from evaluating whether the decades of existing data comport with his predictions. An accounting of the relevant evidence is, however, absent.

Here, I briefly review an alternative formal account of social group discovery, in complement to Pietraszewski's, which addresses these challenges. We adopt a computational model of latent structure learning to move beyond explicit category labels and dyadic similarity as the sole inputs to social group or coalition representations (Gershman & Cikara, Reference Gershman and Cikara2020). By contrast to Pietraszewski's model, latent structure learning (1) does not require conflict or ancillary cues (e.g., similarity on some feature), (2) does not require observation of overtly helpful or harmful behavior, which are relatively rare (Tooby & Cosmides, Reference Tooby and Cosmides2000), and (3) has already garnered empirical support.

We began by asking to what extent do people rely on similarity versus inferences of latent group structure, based on observable behavior, to guide their choice of allies? Note first that the behavior need not be help or harm, though it could be. Note also that the input here is not features (e.g., skin tone and language) which would be considered ancillary cues in Pietraszewski's framework. A mere similarity account predicts that people substitute judgments of behavioral similarity to the self to identify allies (e.g., did this person vote for the same candidate I did in the last election?). This approach, however, neglects that people, including very young children, are sensitive to how well agents coordinate not just with themselves but with others in the environment indicating that people are predisposed to building representations of coordinated coalitions – or social structures – out in the world rather than just egocentric, dyadic similarities or interdependencies (see Cikara, Reference Cikara2021 for a recent review). Thus, an alternative hypothesis is that people's inferences about coalition membership are improved by integrating information both about how agents relate to oneself as well as how they relate to one another.

By what process could people generate group representations on the basis of observing others' non-coalitional behaviors? Via basic statistical learning algorithms. The normative solution to this inference problem is given by Bayes' rule, which stipulates that the posterior probability over groupings given behaviors – P(grouping|behaviors) – is proportional to the product of the likelihood – P(behaviors|grouping) – and the prior probability P(grouping) (Gershman, Pouncy, & Gweon, Reference Gershman, Pouncy and Gweon2017). Individuals who behave similarly will tend to be grouped together, but these groupings are dynamic and context dependent. As Pietraszewski (and we) argue, any model of group inference must be updated over time with more evidence and take into account the influence of all interagent relationships or similarities, not just mere similarity to oneself.

To adjudicate between mere similarity versus structure learning accounts we predicted that even when two agents' choices were equally similar to participants' own, participants' decisions would be influenced by the presence of a third agent who altered the coalitional structure (i.e., by creating a latent group that included the participant and only one of the first two agents). Importantly, a dyadic similarity account would predict that a third agent would have zero influence because the first two agents were equally similar to participants.

We tested this prediction in a series of behavioral experiments framed as learning about strangers' political issue positions (Lau, Pouncy, Gershman, & Cikara, Reference Lau, Pouncy, Gershman and Cikara2018). In each trial, participants stated their position for or against a political issue, and then predicted the choices of three other agents on that same issue. After each prediction, they received feedback about that agent's actual choice. Finally, at the end of this learning phase, participants had to choose with which agent – A or B – they wanted to align themselves on a “mystery issue.” Critically, agents A and B agreed with the participant an equal number of times during the learning phase, making them equally similar to the participant. Depending on the block, however, agent C either clustered with agent B and the participant, or only with agent B, excluding the participant.

As predicted by a latent structure learning account, participants were more likely to align themselves with agent B than A when C's placement created a cluster that put the participant in the same group as agent B (despite the fact that agents A and B were equally similar to the participant). Participants also judged agent B as more competent, moral, and likable than agent A when agent B clustered with the participant versus not. Perhaps most interesting, latent structures continued to exert an effect on ally-choice behavior even when we provided participants with explicit group labels that contradicted the latent structure (i.e., always put agent B in the explicit outgroup). In a companion functional magnetic resonance imaging (fMRI) experiment, the neural signals associated with latent structure representations further explained ally-choice behavior whereas interagent similarity-associated signal did not (Lau, Gershman, & Cikara, Reference Lau, Gershman and Cikara2020).

In summary, and in line with Pietrasweski's challenge to the field, the latent structure learning framework moves away from ancillary similarity and category membership as sole inputs to group representation and inference, and retains the context sensitivity that is a major strength of Pietrasweski's account, but does not require conflict or observation of coalitional behaviors as inputs.