The Illusion of Inclusion: An External Audit of Bias in an Open and a Diversity-Oriented Text-to-Image Model

[ abstract ]

An external, black-box audit testing the claim that commercial text-to-image models are "debiased" or "diversity-aware." Under a locked protocol, 322 images were generated from neutral occupational prompts across six occupations on two systems — an open baseline (Stable Diffusion XL) and a commercial diversity-oriented model (Google Gemini Nano Banana 2) — and coded for perceived gender, skin tone (Monk scale), age, attire and expression against EU, US and global workforce benchmarks. The open model amplified occupational gender stereotypes in both directions; the commercial model applied a selective, one-directional correction, over-injecting women into prestige roles while leaving manual and care work untouched, with a larger aggregate deviation from real workforce composition than the baseline. Across both models none of 217 usable figures was dark-skinned, age was skewed by gender, and expression flattened into uniform smiling — yet both rendered counter-stereotypes perfectly when explicitly asked, locating the harm in defaults. Findings are read through a data-feminism and human-rights lens and linked to the transparency and non-discrimination objectives of the EU AI Act.