Pietraszewski in the target article offers a bottom-up approach to understanding the problem of group representation, describing how these representations could be constructed ex nihilo in the context of conflict by identifying patterns of interactions within triads. Although this account provides a straightforward computational account of mental representations of intergroup conflict, it neglects important top-down influences such as induction and generalization. We suggest that these social group categorization processes may better account for core features of modern human group living, such as flexible, dynamic social identities, and generalized trust in strangers in massive environments, than an approach that relies only on constructing group representations out of an event framework.

As bottom-up processing alone is unlikely to be realizable at scale given humans' limited cognitive resources, top-down processes likely must play some role in developing group representations. One such process is inductive reasoning, which allows us to efficiently learn complex concept and category knowledge from relatively sparse data (Kemp, Perfors, & Tenenbaum, Reference Kemp, Perfors and Tenenbaum2007; Tenenbaum, Kemp, Griffiths, & Goodman, Reference Tenenbaum, Kemp, Griffiths and Goodman2011). Although categories can be quickly built from experience, we also receive rich, highly structured information about these categories in the form of generic language and category labels, quickly scaffolding our learning about abstract category structures, both non-social (Butler & Markman, Reference Butler and Markman2014; Taverna, Padilla, Baiocchi, & Peralta, Reference Taverna, Padilla, Baiocchi and Peralta2021) and social (Baron, Dunham, Banaji, & Carey, Reference Baron, Dunham, Banaji and Carey2014; Gelman, Ware, & Kleinberg, Reference Gelman, Ware and Kleinberg2010; Roberts, Ho, & Gelman, Reference Roberts, Ho and Gelman2017b). By representing groups as abstract, symbolic categories that capture statistical regularities and make probabilistic predictions about how members are likely to appear and act, we gain several insights that are not captured by a model which constructs group representations out of event frameworks alone. Relying on category and concept learning not only allows us to use a highly tractable, domain-general strategy to learn about and represent complex social groups, but it also readily expands the predictive capacity of group assignment beyond conflict and reciprocity, to resolving highly abstract cooperation and coordination problems that complex societies must solve.

Beyond simply giving “rules for assignment” based on statistical regularities in shared features and behaviours, categories allow us to make predictions about how category members are likely to behave, and even to develop prescriptive norms about how they ought to behave (Foster-Hanson & Rhodes, Reference Foster-Hanson and Rhodes2019; Roberts, Gelman, & Ho, Reference Roberts, Gelman and Ho2017a), including group members' moral obligations to the group (Chalik & Rhodes, Reference Chalik and Rhodes2018, Reference Chalik and Rhodes2020). In the context of a conflict, knowledge about agents' social categories can allow observers to make inferences about what kinds of behaviours or group-constitutive roles are more or less likely for an agent to take; rather than (or in addition to) the roles determining the inferences about groups, the prior expectations around group memberships determine the inference of roles.

Indeed, these processes may be critical in the formation and representation of groups where most members are strangers, as members of modern social groups such as cities and countries are able to recognize only a small proportion of the group's total population with fine individual detail (Dunbar, Reference Dunbar2010), and must regularly interact with group members about which one has no information beyond knowledge of perceptually or contextually inferred social categories (such as race, gender, or nationality). Relying on group-based norms and expectations allows us to make broad inferences about strangers that facilitate social prediction without needing to represent these individuals or interactions in a deep way. The cultural transmission of norms regarding social roles and division of labour (e.g., Lew-Levy, Lavi, Reckin, Cristóbal-Azkarate, & Ellis-Davies, Reference Lew-Levy, Lavi, Reckin, Cristóbal-Azkarate and Ellis-Davies2018) can also easily be accommodated within a framework of social category learning and, we argue, is in fact essential to explain how individuals access and internalize these norms within large, complex societies (Gavrilets & Richerson, Reference Gavrilets and Richerson2017).

In contrast to an event model, categorization processes are likely more involved in the social learning processes that underlie cultural learning. Rather than being learned from the bottom-up, many group identities have been developed over long histories of cooperation, affiliation, or conflict. Although processes of social categorization are efficient for making sense of the social world, they come at a cost; when applied to these “prepackaged” groups, the cognitive tools of induction, categorization, and pedagogical reasoning that allow us to swiftly bootstrap sophisticated social category structures can lead to overgeneralization and stereotyping (Macrae, Milne, & Bodenhausen, Reference Macrae, Milne and Bodenhausen1994).

Interestingly, despite Pietraszewski's claim that categorizations are rigid, our social categories are not only swiftly and efficiently learned, but also highly flexible, dynamic, and complex in structure, with people holding multiple social identities which can overlap to varying degrees or even be contradictory (Roccas & Brewer, Reference Roccas and Brewer2002). Social identities can rapidly shift in accordance with their context or prominence, and these shifts can, in turn, shape attention, modulating our memory for social categories (Van Bavel & Cunningham, Reference Van Bavel and Cunningham2012; Van Bavel, Packer, & Cunningham, Reference Van Bavel, Packer and Cunningham2011) as well as amygdala activation (Cunningham, Van Bavel, & Johnsen, Reference Cunningham, Van Bavel and Johnsen2008). When group identities are multiply nested or produce conflicting loyalties, contextual activation or prominence of a group identity could lead to using different social category structures to make behavioural decisions as well as predictions about likely outcomes in complex, real-world contexts.

We applaud Pietraszewski's goal of formalizing the process of social group representation and grounding our understanding of groups in this psychological process. Although we agree that triadic event information can motivate people's reasoning about some – particularly ad hoc – groupings, we believe that representing groups as categories provides unique insight into phenomena such as the transmission of cultural knowledge and generalized trust while allowing for symbolic, abstract group identities that are nevertheless flexible and situationally dependent. These features, we argue, may help us better understand the development of “prepackaged” group identities such as nationalities and ethnic groups through our evolutionary history as well as represent how specific identities emerge and shift among contemporary populations. We propose that extending existing computational models of categorization and concept learning (e.g., Love, Medin, & Gureckis, Reference Love, Medin and Gureckis2004) into the social domain by including dimensions such as identity or motivation may offer a highly tractable account for how we can deploy basic, domain-general inference processes to solve the challenges of reasoning about large and complex groups.