Equal Groups, Unequal Identities: Auditing Normative Gender Representations in Vision-Language Models

Existing evaluations of multimodal fairness have largely focused on disparities between broad demographic groups, asking whether concepts are associated more strongly with one group than another. While valuable, this group-level perspective can overlook an important dimension of fairness: how identities are represented within a demographic category. By treating groups such as “female” and “male” as broadly homogeneous, existing evaluations may miss whether Vision Language Models (VLMs) privilege particular appearances, professions, racial identities, or other socially normative representations. In this work, we address this gap by investigating normative gender representation and within-group representational concentration in VLMs. We construct a structured, category-balanced dataset spanning appearance, profession, race, grooming, and counter-stereotypical gender categories, and examine their alignment with neutral gender prompts using similarity-based Top-$k$ retrieval. Our findings reveal that VLMs organize gender representations around socially familiar and normative identities, producing substantial concentration within demographic groups even when aggregate gender associations may appear balanced. These findings suggest that demographic parity alone provides an incomplete view of multimodal fairness. We therefore highlight normative representation and within-group diversity as complementary dimensions for auditing and understanding bias in VLMs.