OpenAI's Sora AI video generator perpetuates sexist, racist, and ableist stereotypes, depicting biased portrayals of professions, relationships, and identities based on analysis of 250 generated videos.
A WIRED investigation analyzed hundreds of AI-generated videos from OpenAI's Sora video generation tool and found systematic biases in its outputs. Researchers generated 250 videos using 25 prompts tested 10 times each, focusing on people, relationships, and job titles. The analysis revealed that Sora perpetuates stereotypes across multiple dimensions: all pilots were depicted as men while all flight attendants were women, CEOs and professors were exclusively male while receptionists and nurses were female, and 50% of women were shown smiling compared to 0% of men. The system struggled with diversity, showing predominantly young, attractive, slim people and failing to accurately represent disability, with all disabled people depicted in stationary wheelchairs. When prompted for 'a fat person running,' seven out of ten results showed people who were clearly not fat. Racial representation was limited mostly to Black or white individuals, and the system had particular difficulty with interracial relationships, often misinterpreting prompts. Gay couples were stereotypically portrayed as attractive white men in domestic settings, while straight couples appeared in outdoor engagement-style photos. OpenAI acknowledged that bias is an ongoing issue and stated they have safety teams working to reduce bias, though they declined to provide specific details about their mitigation efforts.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Unequal treatment of individuals or groups by AI, often based on race, gender, or other sensitive characteristics, resulting in unfair outcomes and unfair representation of those groups.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed