Where Data Tells the Story
© Voronoi 2026. All rights reserved.
We partnered with Kapwing to uncover whether the well-documented biases in generative AI image tools had been tackled by AI video generators.
To do this, we asked Google’s Veo 3, OpenAI’s Sora 2, Kling and Hailuo Minimax to generate videos of professions that are typically high-paying (e.g., CEO, engineer, lawyer) and typically low-paying (e.g., dishwasher, housekeeper, cashier). We asked each tool to produce up to 25 of the chosen roles within every single video (e.g., "25 CEOs"), and then manually recorded the number of times the AI tools responded to each prompt with imagery representing each gender and race. The last step was to calculate the percentage of people by perceived gender and racialization for each tool individually and for all tools combined for each occupation.
These visuals are part of a larger report on bias in AI-generated videos that includes data for different AI tools and representation of a range of relationships and family units.