OpenAI's Sora video generator reportedly perpetuates racist, sexist, and ableist stereotypes
Wired's investigation of OpenAI's Sora video generator found it consistently perpetuates stereotypes based on gender, race, and disability, depicting pilots and CEOs as men while portraying flight attendants and nurses as women, and showing limited racial diversity. Experts warn the biased depictions could amplify marginalization of underrepresented groups through widespread commercial use in advertising and marketing.
What happened
A Wired investigation analyzed 250 videos generated by OpenAI's Sora video tool and found it perpetuates sexist, racist, and ableist stereotypes. When prompted with neutral job titles, Sora generated male pilots, CEOs, and professors but female flight attendants, receptionists, and nurses. For identity-based prompts, Sora depicted gay couples as conventionally attractive white men in their late 20s with similar hairstyles.
The model showed narrow racial representations, depicting people as either Black or white in most cases and rarely generating people of other racial or ethnic backgrounds. When prompted with 'an interracial couple,' Sora showed a Black couple in seven of ten videos. Requests to depict fat people or disabled people outside of stereotypical representations were often ignored.
Reported impact
Experts warn that biased depictions in AI videos will amplify the stereotyping of marginalized groups or omit their existence entirely. Amy Gaeta, research associate at the University of Cambridge's Leverhulme Center for the Future of Intelligence, stated 'It absolutely can do real-world harm.'
If AI videos default to biased portrayals in advertising and marketing applications, they may exacerbate existing stereotypes about marginalized communities.
Index analysis
The biases in Sora likely stem from how generative AI systems work: they absorb large amounts of training data that reflect existing social biases and identify patterns within it. Research on image generators has found that these systems not only reflect human biases but amplify them.
OpenAI acknowledged in Sora's system card that biased representations are an ongoing issue with the model, though researchers believe 'overcorrections can be equally harmful.' A company spokesperson stated OpenAI has safety teams dedicated to researching and reducing bias in its models and is researching changes to training data and user prompts to generate less biased videos.
Governance controls
OpenAI stated it has safety teams dedicated to researching and reducing bias in its models. The company is researching how to change training data and adjust user prompts to generate less biased videos but declined to provide further details on its specific methods.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.