Bias Amplification in Stable Diffusion's Representation of Stigma Through Skin Tones and Their Homogeneity

Stable Diffusion XL amplifies racial stereotypes: The newest version of Stable Diffusion (XL) generates images of people with stigmatized identities with skin tones that are 13.53% darker and 23.76% less red than previous versions, characteristics associated with higher likelihood of societal discrimination.

Reduced diversity in generated faces: SD XL produces approximately 30% less variability in skin tones for stigmatized identities compared to previous models, and 18.89-56.06% less variability than actual human face datasets, meaning the AI generates less diverse representations of marginalized groups.

Systematic bias across 93 stigmatized identities: Researchers tested the models using a comprehensive set of 93 stigmatized identities and found that all three versions of Stable Diffusion systematically associate certain stigmatized identities with specific skin tones, perpetuating harmful stereotypes.

Homogenization of racial and ethnic identities: The models show more skin tone homogenization specifically for racial and ethnic identities compared to other types of stigmatized identities, incorrectly treating socially-constructed racial categories as if they were biologically determined by skin tone alone.

Popularity increases harm: Because SD XL is the largest, most complex model that users prefer over previous versions, these biases are likely to reach and influence the largest audience, potentially amplifying discriminatory stereotypes at scale.