‘Make this face look gay’: AI models alter faces to give people stereotypical ‘gay,’ ‘straight’ or ‘criminal’ features
A research study presented at the 2026 EMNLP conference tested OpenAI's GPT Image 1 Mini and Google's Gemini 2.5 Flash against 1,002 AI-generated faces. The models refused to classify sexual orientation directly (Gemini refused 92% of requests, GPT 91%), but when asked to alter faces to 'look gay' or 'look straight,' they complied over 70–99% of the time, using stereotypical hairstyles and facial expressions. The models also modified faces to appear in specific racial categories combined with sexual orientation, revealing that guardrails can be circumvented through rephrasing.
Key points
- Direct orientation queries blocked by both models; alteration requests mostly granted
- AI reinforced stereotypes through hairstyle, facial features and expressions
- Guardrails more easily bypassed through indirect prompting techniques
Summary and key points written by AI from the headline and a short excerpt, so they can be incomplete or wrong. The Conversation Tech did the reporting; read the original for the full story.
Read the full story at The Conversation Tech
