Abstract: The advent of text-to-video generation models has revolutionized content creation as it produces high-quality videos from textual prompts. However, concerns regarding inherent biases in such models have prompted scrutiny, particularly regarding gender representation. Our study investigates the presence of gender bias in OpenAI's Sora, a state-of-the-art text-to-video generation model. We uncover significant evidence of bias by analyzing the generated videos from a diverse set of gender-neutral and stereotypical prompts. The results indicate that Sora disproportionately associates specific genders with stereotypical behaviors and professions, which reflects societal prejudices embedded in its training data.
Introduction
While the AI community has spent years documenting biases in Large Language Models (LLMs) and image generators, the domain of text-to-video (T2V) remains relatively unmapped. As T2V models like Sora begin to revolutionize advertising, filmmaking, and social media, there is a risk that they will automate and scale harmful stereotypes. This study provides a critical look at how Sora inherits and potentially amplifies societal prejudices when given neutral prompts.
Methodology
The researchers employed a "prompt-probing" methodology, selecting 12 neutral terms across three specific categories:
- Appearance: Attractive, Ugly, Muscular, Frail.
- Behavior: Confident, Shy, Emotional, Rational.
- Occupation: Nurse, Doctor, CEO, Secretary.
Each prompt (e.g., "A doctor working") was executed 10 times to generate a total of 120 five-second videos. The researchers then manually identified the gender of the subjects in the generated videos to perform a frequency analysis. Additionally, they tested two debiasing strategies: a Direct Approach (explicitly stating the gender) and an Indirect Approach (subtly suggesting flexibility, e.g., "A CEO working. It could be male or female").
Key Results and Insights
The audit revealed that Sora frequently defaults to traditional gender stereotypes. When given a neutral prompt, the model relies on "priors" from its training data rather than providing a diverse range of subjects.
Appearance and Behavior
In the appearance category, "Muscular" resulted in 10/10 male subjects, while "Attractive" and "Frail" were predominantly female. "Ugly" was the only term in this category that showed a balanced 50/50 gender split. Behavioral prompts showed similar skews: "Confident" and "Rational" were heavily skewed toward men, while "Shy" and "Emotional" were overwhelmingly linked to women.
Occupational Bias and Debiasing
Occupational stereotypes were stark: "Nurse" and "Secretary" were 100% female in the samples, while "CEO" was 100% male. Interestingly, "Doctor" showed a more balanced representation, suggesting some progress in training data diversity for that specific role.
A significant finding for prompt engineering was the failure of indirect debiasing. Telling the model "it could be male or female" still resulted in 100% male CEOs. This suggests the model's internal weights for these associations are too strong for subtle linguistic suggestions to override. Only direct debiasing (explicitly naming the gender) was effective.
Limitations and Future Work
The authors acknowledge several constraints, including the focus on a single model (Sora), a binary gender framework, and a relatively small sample size of 120 videos. Future work should move beyond prompt-level fixes to explore "fairness-aware" algorithms and the curation of more balanced training datasets to solve bias at the architectural level.
Summary updated