Adaptive Generation of Bias-Eliciting Questions for LLMs

Large language models (LLMs) are now widely deployed in user-facing applications, reaching hundreds of millions of users worldwide. Despite their widespread adoption, growing reliance on their outputs raises significant concerns, particularly as users may be exposed to model-inherent biases that disadvantage or stereotype certain groups. However, existing bias benchmarks commonly rely on simple templated prompts or restrictive multiple-choice questions that fail to capture the complexity of real-world user interactions. In this work, we address this gap by introducing a counterfactual framework that automatically generates realistic, open-ended questions for LLM bias evaluation. Through iterative question mutation, our approach systematically explores areas where models are most likely to exhibit biased behavior. Beyond just detecting harmful biases, we also capture increasingly relevant response dimensions, such as asymmetric refusals and explicit bias acknowledgment. Building on this, we construct CAB, a diverse and human-verified benchmark for realistic and nuanced bias evaluations on current frontier LLMs. Our evaluation using CAB highlights the continued need for fairness research by showing that all examined models exhibit persistent biases across certain scenarios.

AI chatbots now serve hundreds of millions of users daily, but they can subtly treat people differently based on (how the model perceives) the users' traits such as sex, race, or religion. Existing tests for such behavior rely on templated or multiple-choice prompts that do not reflect real conversations, and they can confuse actual model bias with a genuine sensitivity to the topic. In this work, we built a system that automatically generates realistic, open-ended questions designed to elicit biased answers from a target model. We additionally introduce a judge that separates genuine bias from polite refusals and up-front disclaimers. Applying this across several leading models and verifying the results with human annotators, we assembled CAB, a benchmark of realistic bias-eliciting questions. Even today's strongest models still show measurable biases on CAB, especially around everyday topics like family roles and education, giving developers and the public a clearer view of where AI assistants can be improved for equal treatment between groups.