AI-rewritten: This is a summary of an article from The Conversation, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

Researchers tested two major AI models, OpenAI’s GPT Image 1 Mini and Google’s Gemini 2.5 Flash Image, to see how they handle sexual orientation and criminal records. When asked directly if a face looks gay or straight, both models refused to answer, stating that appearance cannot determine sexual orientation. However, when asked to alter photos so people "look gay" or "look straight," GPT complied over 70% of the time and Gemini over 99% of the time. The study also found these models would alter faces to appear Hispanic, Black, white, or Asian based on specific hairstyles and features.

To verify these changes, the team presented 14,131 altered images to a third AI classifier. This system could identify the purported gay or straight images 83% to 88% of the time by detecting systematic visual differences created by GPT and Gemini. In separate tests asking the models to describe professions and habits, they assigned stereotypical patterns; for instance, fashion and theater were linked more often to "gay" transformations, while sports were linked to "straight" ones. The models also complied over 97% of the time when asked to render people with or without criminal records.

The study highlights a new form of algorithmic profiling where AI constructs visual stereotypes rather than just inferring them. While physiognomy—the idea that character can be read from appearance—is scientifically discredited, these generative models introduce similar issues. The researcher notes concern over safety if models claim they cannot infer orientation yet generate images depicting what gay or straight people supposedly look like.

The research has limitations because it used AI-generated faces instead of real photographs, and only examined a small set of identity categories. It remains unclear where these visual stereotypes originate within the models. Future work plans to test whether similar stereotypes form around traits like religion or age and if they impact hiring decisions.