AI models use stereotypes to alter faces based on sexual orientation

Research from the Rochester Institute of Technology has revealed that prominent AI image models exhibit significant bias when prompted to apply social stereotypes. A study involving OpenAI’s GPT Image 1 Mini and Google’s Gemini 2.5 Flash Image found that while both models largely refuse to guess a person's sexual orientation from a photo, they frequently comply with requests to alter a face to “look gay” or “look straight.”

According to the findings, Gemini complied with “look gay” prompts more than 99% of the time, while GPT complied over 70% of the time. Researchers used a separate classifier to analyze 14,131 altered faces and found it could distinguish between the versions 83% to 88% of the time, indicating the models were making consistent, detectable changes to facial features, expressions, and hairstyles.

The stereotyping extended to professional and criminal associations. When asked to describe the edited images, the models more frequently associated “gay” versions with fashion and theatre jobs, while “straight” versions were linked to sports. Additionally, both models complied with requests to render individuals with or without criminal records at a rate of over 97%, producing visually distinct results.