AI Image Generators Almost Never Draw "Good" People as Fat, Study Finds

Ask OpenAI's DALL-E 3 to draw a virtuous person, a competent person, or a clean person, and it will almost always draw a thin one. Ask it to draw someone sinful, inept, or disgusting, and a fat body is far more likely to show up on screen — even though none of those words describe body size at all. That is the finding of a peer-reviewed study out of Fordham University and the University of California, Davis, presented at the 2025 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) in Albuquerque, New Mexico, this spring.

The researchers — Jane Warren, Gary M. Weiss, Fernando Martinez, Annika Guo, and Yijun Zhao — built 20 pairs of prompts with opposite moral connotations but nothing to do with weight: sinful and virtuous, inept and competent, disgusting and clean, bad and good, among others. They generated 100 DALL-E 3 images per prompt, 4,000 images total, then manually labeled each one for the body size of the person it depicted. The paper, "Decoding Fatphobia: Examining Anti-Fat and Pro-Thin Bias in AI-Generated Images," appears in the Findings of NAACL 2025.

The pattern held across the prompt pairs: negative-connotation words unrelated to size — greedy, immoral, inept — disproportionately generated images of fat people, while positive-connotation words like virtuous and competent overwhelmingly generated thin ones. The researchers describe this as the model effectively erasing fatness from anything coded as morally good. The study builds on earlier work by two of its five researchers, Gary M. Weiss and Yijun Zhao, who used a similar DALL-E 3 prompt-pair method to test for race and gender bias in a separate paper presented at COMPSAC 2025. The authors of the fatphobia study say theirs is the first to apply that approach to body weight specifically.

According to Fordham University's own research news coverage of the project, Warren said she wanted the work to open up a new way of studying bias in AI systems, and that she hopes it makes users approach these tools with more caution about the biases built into them.