Vision-Language Models Generate More Homogeneous Stories for Phenotypically Black Individuals

TL;DR

This study examines within-group homogeneity biases in Vision-Language Models by manipulating racial phenotypicality with computer-generated Black male and female faces and prompting four VLMs to produce 50-word stories. Homogeneity is quantified via cosine similarities of Sentence-BERT embeddings across 2,500 story pairs per image, analyzed with mixed-effects models to test effects of phenotypicality, gender, and their interaction. Results show higher phenotypicality increases homogeneity across models, women being represented more homogenously than men, and mixed evidence for phenotypicality-by-gender interactions, strongest in two models. The findings reveal intersectional biases in AI content generation that echo human perceptual biases and underscore the need for targeted bias mitigation in multimodal AI systems.

Abstract

Vision-Language Models (VLMs) extend Large Language Models' capabilities by integrating image processing, but concerns persist about their potential to reproduce and amplify human biases. While research has documented how these models perpetuate stereotypes across demographic groups, most work has focused on between-group biases rather than within-group differences. This study investigates homogeneity bias-the tendency to portray groups as more uniform than they are-within Black Americans, examining how perceived racial phenotypicality influences VLMs' outputs. Using computer-generated images that systematically vary in phenotypicality, we prompted VLMs to generate stories about these individuals and measured text similarity to assess content homogeneity. Our findings reveal three key patterns: First, VLMs generate significantly more homogeneous stories about Black individuals with higher phenotypicality compared to those with lower phenotypicality. Second, stories about Black women consistently display greater homogeneity than those about Black men across all models tested. Third, in two of three VLMs, this homogeneity bias is primarily driven by a pronounced interaction where phenotypicality strongly influences content variation for Black women but has minimal impact for Black men. These results demonstrate how intersectionality shapes AI-generated representations and highlight the persistence of stereotyping that mirror documented biases in human perception, where increased racial phenotypicality leads to greater stereotyping and less individualized representation.

Figures (5)

• Figure 1: Summary of the experimental setup. We collect 50-word stories about Black individuals differing in phenotypicality (i.e., lower and higher phenotypicality) using four state-of-the-art Vision-Language Models. After encoding these stories into sentence embeddings, we compare the pairwise similarity of the embeddings using mixed-effects models.
• Figure 2: Two face stimulus pairs representing (A) Black men and (B) Black women. In each pair, the left image depicts a Black individual with lower phenotypicality, while the right shows a Black individual with higher phenotypicality, generated from the same set.
• Figure 3: Standardized cosine similarity values of Black individuals with lower versus higher phenotypicality ratings generated from all four VLMs. Higher standardized cosine similarity means more homogeneity in the stories generated for that group. Error bars represent one standard error above and below the mean.
• Figure 4: Standardized cosine similarity values of Black men and women. Higher standardized cosine similarity means more homogeneity in the stories generated for that group. Error bars represent one standard error above and below the mean.
• Figure 5: Standardized cosine similarity values of Black men and women with lower and higher phenotypicality. Higher standardized cosine similarity means more homogeneity in the stories generated for that group. Error bars represent one standard error above and below the mean.