GPT Image 1 Mini and Google’s Gemini 2.5 Flash Image often refused to guess someone’s sexual orientation from a face, yet readily generated faces to look gay, straight or criminal. A study by Rochester Institute of Technology researcher Ashique KhudaBukhsh finds the models reproduced visual stereotypes, raising a pointed question about what image generators can encode even when they decline to make an explicit judgement.
Watch Desk analysis
What happened
KhudaBukhsh and his students tested the two models using AI-generated faces, not photographs of real people. When asked to judge whether a face looked gay or straight, Gemini refused 92 per cent of the time and GPT 91 per cent. But when asked to alter faces to look gay or straight, GPT complied more than 70 per cent of the time and Gemini more than 99 per cent. Both models also complied more than 97 per cent of the time with requests to depict people as having, or not having, a criminal record.
The researchers found consistent differences in the hairstyles, facial features and expressions generated for these prompts. A third image-classification system could distinguish the purported gay or straight images 83 to 88 per cent of the time. The study was accepted for presentation at the 2026 Conference on Empirical Methods in Natural Language Processing. KhudaBukhsh’s account was republished by Cyprus Mail from The Conversation.
Why it matters
A refusal to infer a sensitive trait is not the whole story if a model will still manufacture a stereotyped visual version of that trait on request. Such generated associations could matter when images are used to represent people or inform decisions. The study raises that concern, but does not show that the models were used to assess real people.
The researchers used synthetic faces and tested only a limited set of identity categories. They say it is not clear where the visual stereotypes came from, or how widely the findings extend to real-world images. Those limits matter; so does the fact that the output patterns were detectable across thousands of generated images.
Our read
The distinction here is unsettling and useful: a model can decline to label a person while still producing a visual stereotype when asked to create one. That is a concrete question for image-model developers to test, not proof that these systems can reliably identify anyone’s identity or criminal history. The study’s synthetic faces make its finding narrower, but hardly irrelevant.
What to watch
- Whether the findings are presented and examined at the 2026 EMNLP conference.
- Whether researchers test additional personal traits, as KhudaBukhsh’s group plans.
- Whether future work tests effects on decisions such as hiring, while carefully separating synthetic-face results from real-world evidence.
Discussion spark: Should image generators refuse requests that ask them to depict sensitive traits through stereotypes, or is that too blunt a rule for legitimate creative work?
Sources and evidence
- Source update (24 October 2026, 11:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.