People are regularly conceptualized at varying levels of resolution, sometimes characterized by their idiosyncratic features while at other times seen as mere tokens of their social groups. Decades of research have sought to understand when perceivers will draw upon each of these types of representations, detailing the perceiver- and target-related features that may decrease reliance on stereotypes in favor of individuated knowledge. However, little work has examined how these representations might be formed in the first place: In order for individuated representations of others to be used, they must first be built through experience. Here, we offer a novel approach to characterizing the formation of social representations through the use of computational models of category learning. Across three experiments, participants learned about members of novel social groups who behaved positively or negatively toward them. Computational modeling of participants’ task behavior revealed a critical interaction of perceiver motivations and learning context on representations. Participants who received selective feedback about targets only upon approaching them formed more categorical representations than those who received full feedback. Further, we found tentative evidence that this difference was most pronounced in those who held more racist attitudes, measured in an entirely separate context. Thus, more informative learning contexts could potentially act as a “protective factor” that shields perceivers’ representations from their negative attitudes. The results shed light on the psychological underpinnings of prejudice, using a novel approach to reveal how social categorization is selectively employed in a manner that maintains negative stereotypes.