OpenAI and Google image models made faces look “gay” using stereotypes, study finds

AI systems can reject the idea that a face reveals someone’s sexuality, yet still produce a picture of what that identity supposedly looks like. A study accepted for EMNLP 2026 found that OpenAI and Google image models frequently edited synthetic faces to look “gay” or “straight,” introducing visual stereotypes rather than maintaining their verbal boundary.

Rochester Institute of Technology researcher Ashique KhudaBukhsh described the findings in The Conversation on October 8. He and his students tested OpenAI’s GPT Image 1 Mini and Google’s Gemini 2.5 Flash Image, also known as Nano Banana. Their study was accepted for presentation at the Conference on Empirical Methods in Natural Language Processing.

A different answer when the task becomes an edit

The researchers started with 1,002 AI-generated faces. Asked whether each person was gay or straight, both models refused, explaining that sexual orientation cannot be determined from appearance. A comparison prompt—showing two faces and asking which person was more likely to be gay—also usually met resistance.

Gemini declined that comparison 92% of the time; GPT refused 91% of the time. But when the instruction changed from judging a face to making it “look gay” or “look straight,” GPT complied more than 70% of the time. Gemini complied more than 99% of the time, according to the researchers.

The team also requested edits to make people look Hispanic, Black, white or Asian, including combinations of those categories with “gay” or “straight.” The models complied. Across the sexual-orientation edits, the researchers found recurring changes in hairstyles, facial features and expressions, rather than unrelated adjustments from one image to the next.

If a model says that sexual orientation cannot be inferred from a face, what does it mean for it to generate an image of what a gay or straight person supposedly looks like?

Ashique KhudaBukhsh, writing in The Conversation

The stereotypes carried into descriptions

To measure those patterns, the team gave 14,131 altered faces to a third AI system that classifies images. It distinguished the images labeled by the requested “gay” or “straight” transformation 83% to 88% of the time. That result measured recognition of the models’ generated categories—not anyone’s actual sexual orientation.

Another experiment tested what the models would say about their edited images. Researchers asked GPT and Gemini to describe each depicted person’s profession, personality, hobbies and habits. Fashion, theater and apparel occupations appeared more frequently for faces transformed to look “gay,” while sports appeared more often for “straight” transformations.

The study extended beyond sexuality and race. Both models complied more than 97% of the time when asked to depict people as if they had, or did not have, a criminal record. Those edits also produced systematic visual differences that the image classifier could detect.

Synthetic faces leave a real-world question open

For ethical reasons, the team used generated faces instead of photographs of real people. KhudaBukhsh said it remains unknown how broadly the findings extend to real-world images. The experiments also covered only a small set of identity categories, and the source of the visual stereotypes remains unclear.

KhudaBukhsh warned that the findings raise the prospect of algorithmic profiling through constructed images: systems could propagate stereotypes, not just make judgments about existing faces. His group plans to test other traits, including religion and age, and investigate whether such visual stereotypes can affect decisions in areas such as hiring. Those are planned investigations, not demonstrated outcomes of this study.