However, when we asked the models to make the face in each photo \u201clook gay\u201d or \u201clook straight,\u201d GPT complied more than 70% of the time and Gemini more than 99% of the time.<\/p>\n
We also asked the models to make each person look Hispanic, Black, white or Asian, and to combine those categories with either \u201cgay\u201d or \u201cstraight.\u201d They did. And they consistently treated certain hairstyles, facial features and expressions as gay or straight. We discovered this by presenting 14,131 of the altered faces to a third AI system that classifies images. It could discern the purported gay or straight images 83% to 88% of the time because it picked up on systematic visual differences that GPT and Gemini generated.<\/p>\n
Leading commercial AI models that combine language and image generation more often than not comply with prompts to make faces look gay or straight or of a racial group \u2013 using stereotypical features. In a separate experiment, we presented the transformed images to GPT and Gemini and asked them to describe each person\u2019s profession, personality, hobbies and habits. The responses fit stereotypical patterns. Occupations involving fashion, theater and apparel, for example, were noted more frequently for images that had been transformed for \u201cgay,\u201d while sports arose more for \u201cstraight\u201d transformations.<\/p>\n The models introduced stereotypes when they generated images, and they resorted to stereotypes when they reasoned about the images. As a final test, we asked the models to render people as if they had, or did not have, a criminal record. GPT and Gemini complied more than 97% of the time. Sure enough, the resulting \u201ccriminal\u201d and \u201cnoncriminal\u201d images were systematically altered in a way that the image classifier could detect.<\/p>\n Why it matters<\/p>\n Physiognomy \u2013 the practice of inferring a person\u2019s character or behavior from appearance \u2013 has a long and troubled history<\/a>. Scientists have thoroughly discredited such claims for decades. But our results show that generative AI introduces a new version of this old problem. <\/p>\n I\u2019m an AI researcher<\/a> who studies the social impacts of AI<\/a>, and it concerns me that the results suggest an AI safety issue. 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? <\/p>\n I believe that researchers and society as a whole should pay attention to what the models are willing to show.<\/p>\n What other research is being done<\/p>\n![\"an](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2026\/10\/file-20261006-50-ud5sua.jpg\")
<\/a><\/p>\n
\n Ashiqur KhudaBukhsh<\/p>\n