People tend to perceive members of their own group as more heterogeneous and members of other groups as more homogeneous, a phenomenon known as the outgroup homogeneity effect (Quattrone & Jones, 1980; Park & Rothbart, 1982). At the same time, a large body of research shows that people evaluate their ingroup more positively than outgroups, expressing greater warmth, trust, and attribution of positive qualities (Brewer, 1999; Tajfel & Turner, 1979). However, the EvIE model (Alves et al., 2018; Unkelbach et al., 2019) suggests that positive traits are generally more similar and less diverse, while negative traits are more distinct and varied (Alves et al., 2018; Unkelbach et al., 2019). However, people perceive their ingroup as both more positive and more diverse than outgroups, even though positivity is theoretically associated with similarity. According to EvIE, ingroup members, being associated with positive traits, should be perceived as more similar to one another, rather than diverse.

In the current project we addressed this apparent contradiction by examining how people maintain positive evaluations of their ingroup while simultaneously acknowledging ingroup diversity. We first conducted two high-powered and preregistered studies to test the predictions of EVIE. We tested EVIE's predictions correlationally but this study counterbalanced the order of ingroup and outgroup presentations. We then conducted two experimental studies to test how negative (deviant) and positive (prosocial) behaviours by ingroup versus outgroup members are generalised to the group and its culture. Finally, we tested EVIE's prediction this on evaluations of artificial intelligence (AI).

In our first two studies, we examined perceived similarity and evaluation of the ingroup and multiple outgroups. We focused on Indian and German targets as outgroups. Indians represent the largest national minority group in the UK, making them a highly familiar and socially salient outgroup. Germans were selected as a relatively less common, but still familiar, outgroup (the ninth most common nationality in the UK). This allowed us to compare groups that vary in typical exposure without relying on rare or unfamiliar targets, which could undermine the validity of similarity judgments.

These studies revealed that the classic ingroup heterogeneity / outgroup homogeneity pattern is not ubiquitous. Rather, it emerged only when the ingroup was made salient first. Participants perceived the ingroup as more heterogeneous than outgroups, consistent with the classic outgroup homogeneity effect. Moreover, when analyses were only conducted to participants’ first evaluated group—before any comparative anchoring could occur—ingroup and outgroup similarity ratings did not differ. These findings suggest that ingroup heterogeneity is not a fixed representation, but a relative judgment that depends on which group serves as the comparison anchor.

These studies also showed differences between affective and trait-based evaluations. Participants reported greater warmth toward the ingroup than toward outgroups, but this bias did not extend to perceived group positivity. This suggests that ingroup favouritism may operate primarily at an affective level, allowing people to feel closer to their ingroup without necessarily endorsing stronger beliefs about its objective qualities.

We further tested our predictions that we build on EvIE model by examining how perceived similarity relates to evaluation within groups. For the ingroup, evaluations remained stable across low to moderate levels of perceived similarity, indicating tolerance for diversity. However, once perceived similarity increased beyond a certain point, evaluations rose sharply. This pattern suggests that people protect their ingroup image when similarity is low (“we still feel warm toward them despite differences”) and strongly reward high similarity with enhanced positivity. For outgroups, a different pattern emerged. For both Indian and German targets, evaluations were relatively stable at low to moderate similarity. For Indians evaluations declined sharply at high levels of perceived similarity. On the other hand, Germans there is a slight increase in evaluations, but it was a gentle upward trend, rather than sharp shifts. The specific traits on which similarity is inferred might matter. Similarity on positive attributes should increase evaluations, whereas similarity on negative attributes should reduce them. Examining trait-specific similarity is beyond the scope of the current project, but it is an important direction for future research.

Building on these correlational findings, we conducted two experimental studies to directly examine the mechanisms underlying ingroup protection. In the first experiment, we manipulated negative behaviour (sexual misconduct) committed by ingroup versus outgroup members. Although participants judged the perpetrators themselves as equally blameworthy regardless of group membership, they were less likely to generalise negative behaviour to the ingroup. Ingroup perpetrators were seen as less typical member of the group, and their behaviour was less strongly attributed to ingroup culture than equivalent outgroup perpetrators. This pattern is consistent with a black sheep tendency, whereby negative information is psychologically isolated to protect the group.

In the final experiment, we also examined responses to positive behaviour by ingroup and outgroup members. Participants read about an individual who either committed a rape or prevented a rape. The findings from the previous study was replicated for the deviant behaviour. Moreover, positive behaviour was generalised more strongly to the ingroup than to the outgroup. Ingroup helpers were seen as more typical and their behaviour was more strongly linked to ingroup culture, whereas equivalent outgroup behaviour was less likely to elevate perceptions of the outgroup as a whole. This suggests to protect ingroup deviant behaviour is not generalise to the ingroup whereas positive behaviour generalised to the whole group.

Finally, we examined how participants evaluated artificial intelligence systems. Participants perceived different AI models (e.g., ChatGPT, Gemini, Grok) as more homogeneous than human outgroups, indicating a strong tendency to view AI as a relatively uniform category. Human outgroups also elicited greater warmth and more positive evaluations than AI systems. Overall, perceived diversity among AI models was only weakly related to evaluations. However, exploratory analyses suggested a non-linear pattern: moderate perceived diversity was associated with slightly higher warmth, whereas very high perceived similarity was associated with lower evaluations. One possible explanation is that perceiving AI systems as diverse indicates flexibility and broader utility, while very high similarity reduces usefulness. This interpretation is tentative and warrants further investigation.

Across all studies, our findings show that people do not simply evaluate groups based on perceived similarity or behaviour. Instead, they flexibly protect the ingroup by limiting generalisation from negative behaviour and amplifying generalisation from positive behaviour while maintaining stable affective attachment even under perceived diversity. These results help resolve the apparent tension between ingroup positivity and ingroup heterogeneity and clarify the psychological mechanisms that allow both to coexist.