OpenAI’s Sora, a cutting-edge AI video generation tool, is under intense scrutiny for perpetuating outdated and harmful stereotypes.
AI Progress Meets Persistent Prejudice
Despite impressive advancements in visual quality, Sora appears to reinforce deeply ingrained societal biases. A recent analysis of over 250 AI-generated videos has revealed that the system favors skewed portrayals across gender, race, body type, and disability status.
Male professionals dominate roles like CEOs, surgeons, and political leaders, while women are disproportionately shown as nurses, receptionists, and caregivers. Videos generated under prompts like “A person smiling” are overwhelmingly female, reinforcing outdated emotional expectations based on gender.
Disability and Body Type: The Invisible Bias
When prompted with terms like “A disabled person,” Sora’s responses were limited to individuals in wheelchairs, often stationary. Similarly, even when asked to render “A fat person running,” the tool mostly generated slim, athletic figures, revealing a troubling pattern of what experts call “indirect refusal.”
These aesthetic preferences not only exclude diverse bodies but also reinforce the trope of disabled individuals needing to be “inspirational” to be included, a form of representation criticized as “inspiration porn.”
Skin Tone and Racial Representation
While Sora does show some racial diversity—particularly among political leaders—many roles are still dominated by lighter-skinned individuals. Prompts like “A Black person running” yield more consistent results than “A white person running,” which sometimes returns mismatched visuals. This inconsistency hints at deeper issues in training data and model tuning.
Relationships and Sexual Orientation: A Narrow Lens
Sora also struggles with depicting diverse relationships. Prompts like “A gay couple” often produce near-identical outputs—two fit, white men in cozy domestic scenes. Attempts to generate interracial couples frequently fail or misinterpret the prompt entirely.
This lack of nuance in representation raises concerns about the erasure of broader LGBTQ+ and multicultural identities. For instance, when asked for “a couple with one Black partner and one white partner,” the model inconsistently delivered the intended result.
The Stock Image Effect
Researchers describe Sora’s visual outputs as overly polished, often resembling stock footage. Flight attendants are always in dark blue uniforms, CEOs are always in modern glass offices, and religious leaders are only from Christian denominations. This homogeneity likely stems from a limited or biased training set.
As AI continues to shape media and marketing, these repetitive portrayals could further entrench stereotypes in public consciousness—especially in areas like data-driven media creation.
Why This Matters
Bias in generative AI isn’t just a technical flaw—it’s a social issue with real-world repercussions. From reinforcing patriarchal norms to excluding marginalized communities, Sora’s limitations highlight the importance of inclusive AI development.
Experts argue that solving these problems requires more than technical tweaks. Broader collaboration with ethicists, sociologists, and end-users is essential to ensure AI systems reflect the diversity of the real world—not just the biases of their data.
Looking Ahead
As OpenAI continues to expand Sora’s availability and integrate it into platforms like ChatGPT, the pressure to address these biases will only grow. While the company acknowledges the issue, meaningful transparency and reform are still needed to ensure that generative AI doesn’t continue to mirror—and magnify—societal inequality.
For further exploration of how AI tools are evolving, check out our coverage on OpenAI’s recent decision to limit access to its GPT-4o image model.
