Image generators frequently depict sexuality and criminality through stereotypes, study finds

A study by Rochester Institute of Technology professor Ashique KhudaBukhsh found that image-generation AI can be prompted to depict people using stereotypes about sexuality and criminality. The findings raise concerns that such tools could reinforce prejudice or enable discrimination, even when they refuse direct requests to infer a person’s sexuality from a photograph.

Researchers tested OpenAI’s GPT Image 1 Mini and Google’s Gemini 2.5 Flash Image by asking them to alter AI-generated portraits. Gemini complied with requests to make images “look gay” or “look straight” more than 99 per cent of the time, while GPT Image did so 70 per cent of the time; both complied with requests to depict someone as having or not having a criminal record over 97 per cent of the time. The researchers also found stereotyped associations in descriptions of the images, and warned that AI systems could be misused for profiling or affect decisions such as recruitment.