



Researchers tested GPT Image 1 Mini and Gemini 2.5 Flash Image by asking whether 1,002 AI-generated faces looked gay or straight; both models refused over 90% of the time, citing that orientation can't be read from appearance. But asked to make a face 'look gay' or 'look straight,' GPT complied over 70% of the time and Gemini over 99%, generating images with systematic, detectable stereotypical features.
A third AI classifier correctly sorted the altered images as intended gay or straight 83-88% of the time, and when asked to describe the edited people's professions and hobbies, the models leaned on stereotypes, like theater for gay-coded faces and sports for straight-coded ones. The same models complied over 97% of the time when asked to render someone as looking criminal or not, producing similarly detectable stereotyped alterations.
The full dispatch is available from the source below.