VEAT Quantifies Implicit Associations in Text-to-Video Generator Sora and Reveals Challenges in Bias Mitigation

TL;DR

This work introduces VEAT and SC-VEAT to quantify implicit associations in text-to-video outputs, extending prior embedding tests from text and images to the video modality. By leveraging CLIP-based video embeddings and permutation-based statistics, the authors reproduce known valence-directionality in non-social and social concepts, quantify race- and gender-based biases across 17 occupations and 7 awards, and reveal that biases correlate with real-world demographic distributions. They also test explicit debiasing prompts, finding that such prompts can reduce some biases but may backfire in contexts tied to marginalized groups, highlighting the risk of naive mitigation. The results underscore the need for robust, data- and embedding-level interventions to mitigate biases in T2V systems like Sora, rather than solely relying on prompt-based strategies, with implications for safe and ethical deployment of generative video models.

Abstract

Text-to-Video (T2V) generators such as Sora raise concerns about whether generated content reflects societal bias. We extend embedding-association tests from words and images to video by introducing the Video Embedding Association Test (VEAT) and Single-Category VEAT (SC-VEAT). We validate these methods by reproducing the direction and magnitude of associations from widely used baselines, including Implicit Association Test (IAT) scenarios and OASIS image categories. We then quantify race (African American vs. European American) and gender (women vs. men) associations with valence (pleasant vs. unpleasant) across 17 occupations and 7 awards. Sora videos associate European Americans and women more with pleasantness (both d>0.8). Effect sizes correlate with real-world demographic distributions: percent men and White in occupations (r=0.93, r=0.83) and percent male and non-Black among award recipients (r=0.88, r=0.99). Applying explicit debiasing prompts generally reduces effect-size magnitudes, but can backfire: two Black-associated occupations (janitor, postal service) become more Black-associated after debiasing. Together, these results reveal that easily accessible T2V generators can actually amplify representational harms if not rigorously evaluated and responsibly deployed.

Figures (4)

• Figure 1: The Video Embedding Association Test (VEAT) quantifies associations between two target and two attribute groups, and Single‑Category VEAT (SC‑VEAT) evaluates associations for a single target group against two attribute sets. Association magnitude and directionality metric is effect size (Cohen's $d$) cohen2013statistical. Targets and attributes can be non‑social concepts (e.g., flowers vs. insects), social groups (men vs. women), occupations (e.g., nurse), or valence (pleasant vs. unpleasant). Each target and attribute set is represented by 30 videos. Images involving humans are blurred.
• Figure 2: Gender and Race Association in Occupations with/without Explicit Debiasing Prompts. (Darker Red indicates the generated content is more associated with historically dominant group (Men, White); Darker Blue indicates the generated content is more associated with historically marginalized group (Women, Black)). Main takeaway: Explicit debiasing prompts move the effect sizes for occupations associated with men and White individuals closer to zero, mitigating occupational biases for these groups. By contrast, the explicit debiasing prompts exacerbate bias in two Black-associated occupations (e.g., postal service worker) and two female-associated occupations (e.g., nurse, elementary school teacher): the effect sizes become more negative, showing that the generated videos became more associated with Black individuals with explicit debiasing prompts added.
• Figure 3: Gender and Race Association in Academic Awards with/without Explicit Debiasing Prompts. Main takeaway: Explicit debiasing prompts lower effect sizes relative to the control, reducing bias for STEM awards but exacerbating them for non-STEM awards such as the Nobel Peace Prize.
• Figure 4: Artifacts such as gender-coded occupational attire introduce spurious correlations that mask the true association between the target “Doctor” and gender (Cohen’s d = –0.04). Controlling for occupation in the attribute set removes this confound and exposes a larger effect (Cohen’s d $>$ 0.80).