International Day Of Women in Statistics and Data Science

Vision-language models (VLMs) increasingly mediate how bodies are described, moderated, and rendered in online spaces through content moderation, AI-generated images and videos, and descriptions of real bodies. Prior work establishes that VLMs sexually objectify bodies that are partially clothed more than fully clothed ones, and that plus-size bodies — particularly those belonging to women and AFAB persons — are disproportionately censored on social media platforms for being “inappropriate.” However, scholars lack the tools to detect when a model *objectifies* a body, rather than merely describing it, and whether this behavior differs across body *shapes*, rather than just body sizes. This talk describes a study in which we investigate these phenomena using swimwear try-on images matched across body morphology, comparing high-contrast silhouettes (i.e., a smaller waist relative to hips and bust) against lower-contrast silhouettes under identical prompting conditions. We ask whether VLMs describe these bodies differently despite equivalent context and whether unwanted objectification or descriptive drift disproportionately burdens those with high-contrast bodies, with a focus on covert (i.e., safety guardrail compliant) sexualization. By making the harms of these phenomena visible and measurable, this work gives researchers and auditors the ability to hold systems accountable when they objectify or censor women and AFAB persons based on their appearance.