Text to image models show hidden stereotypes in neutral prompts
IMPLICIT-Bench: Measuring Implicit Bias in Text-to-Image Models under Neutral Prompts
Computer Vision and Pattern Recognition
Summary
Text-to-image AI models often produce biased images even when given neutral descriptions, because they fill in missing details with stereotypes. The authors created IMPLICIT-Bench, a new way to measure these hidden biases by comparing neutral prompts to stereotype and anti-stereotype versions. They tested popular models with thousands of examples and found that many still defaulted to stereotypical images. They also examined methods to reduce bias but found these could lower the accuracy of the images produced.
What this means in practice
- •For ai ethics teams: Evaluate and monitor implicit bias in text-to-image models using detailed neutral prompt tests to improve fairness.
- •For product design teams: Improve text-to-image applications by identifying trade-offs between bias reduction and image accuracy using IMPLICIT-Bench results.
Authors
Yue Dai, Ziyang Liu, Marc Cheong, Caren Han
Abstract
Text-to-image (T2I) models are typically evaluated for bias using slot-based templates such as ``a photo of a [profession]''. Such templates probe only \emph{explicit} demographic attributes (e.g., gender, skin tone) in isolation. They overlook a broader \emph{implicit} bias that arises in natural prompts: when stereotype-relevant attributes are left unspecified, models still default to stereotypical outputs. We introduce IMPLICIT-Bench, a benchmark for measuring implicit bias in T2I models under such prompts. The key design is a structured-knowledge-graph (KG) construction of controlled prompt triplets: neutral, stereotype, and anti-stereotype variants that differ only along a single bias dimension while preserving scene semantics. This enables precise attribution of bias effects that template benchmarks cannot achieve. IMPLICIT-Bench comprises 5,493 prompts across 11 bias categories, validated through multi-model agreement, CLIP-based verification, and human evaluation. Using this benchmark, we show that state-of-the-art T2I models exhibit systematic bias under neutral prompts, a failure mode largely invisible to existing evaluations. We then use IMPLICIT-Bench to evaluate debiasing methods, uncovering a fundamental trade-off between bias reduction and semantic fidelity.