Stable Diffusion Exhibited Biases for Prompts Featuring Professions

Researchers analyzed over 5,000 images generated by Stable Diffusion, an AI text-to-image model developed by Stability AI, and found systematic bias in occupational and demographic representations. The system was trained on the LAION-5B dataset containing over 5 billion images scraped from the internet. When prompted with job titles, Stable Diffusion consistently generated images that amplified existing stereotypes beyond real-world statistics. For example, while 34% of US judges are women, only 3% of AI-generated 'judge' images depicted women. For 'fast-food worker' prompts, 70% of generated images showed people with darker skin tones, despite 70% of actual fast-food workers being White. The bias extended to criminal contexts, with over 80% of 'inmate' images showing people with darker skin, though people of color comprise less than half the US prison population. Research by Sasha Luccioni at Hugging Face and Bloomberg's analysis revealed these patterns across multiple occupations. The system is being integrated into commercial platforms including Shutterstock, Adobe, and Canva, potentially spreading biased representations. Stability AI acknowledged the biases stem from training data but has not yet deployed improved models to address the issue.