New framework uses risk tiers to audit gender bias in AI image generation

Text-to-image (T2I) models are increasingly used for education, media, and public communication, but they tend to reinforce gender stereotypes, leading to representational erasure and skewed perceptions of role suitability. Existing bias metrics are fragmented and reported without a shared understanding of assumptions or context, limiting their usefulness for technical auditing and governance. Researchers from the paper 'Context Matters: Auditing Gender Bias in T2I Generation through Risk-Tiered Use-Case Profiles' (FAccT 2026) address this by proposing a comprehensive framework that aligns risk categories with evaluation metrics and harms.

The framework consists of three constituents: risk-tiered use-case profiles aligned with the EU AI Act's risk categories (unacceptable, high, limited) to vary auditing expectations; a metric catalog consolidating gender bias evaluation methods into gender prediction, embedding similarity, and downstream task categories; and a harm typology mapping context-dependent harms (e.g., representational, quality-of-service) to specific risk scenarios. The paper introduces THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias), which systematise auditing by incorporating context, scenario, bias manifestation, harm hypotheses, and audit strategy. This enables a standardised, risk-aware approach to gender bias auditing across different deployment contexts.