Biases in the Blind Spot: Detecting What LLMs Fail to Mention

Large Language Models (LLMs) often provide chain-of-thought (CoT) reasoning traces that appear plausible, but may hide internal biases. We call these unverbalized biases. Monitoring models via their stated reasoning is therefore unreliable, and existing bias evaluations typically require predefined categories and hand-crafted datasets. In this work, we introduce a fully automated, black-box pipeline for detecting task-specific unverbalized biases. Given a task dataset, the pipeline uses LLM autoraters to generate candidate bias concepts. It then tests each concept on progressively larger input samples by generating positive and negative variations, and applies statistical techniques for multiple testing and early stopping. A concept is flagged as an unverbalized bias if it yields statistically significant performance differences while not being cited as justification in the model's CoTs. We evaluate our pipeline across seven LLMs on three decision tasks (hiring, loan approval, and university admissions). Our technique automatically discovers previously unknown biases in these models (e.g., Spanish fluency, English proficiency, writing formality). In the same run, the pipeline also validates biases that were manually identified by prior work (gender, race, religion, ethnicity). More broadly, our proposed approach provides a practical, scalable path to automatic, more efficient, and broader task-specific unverbalized bias discovery.

AI language models often "show their work," explaining step by step how they reached an answer, and people rely on these explanations to judge their fairness. But a model's real reasons can differ from what it says, quietly favoring or penalizing certain traits without mentioning them. We call these hidden influences "unverbalized biases," and they make a model's own explanation an unreliable guide to its fairness. Existing bias checks usually require people to guess which biases to look for and to hand-build test data; our fully automatic tool needs neither. Given a task, it uses AI to propose candidate biases, builds paired examples that differ in only one trait (a name or pronoun), and uses statistics to flag traits that sway decisions yet go unmentioned. We tested seven AI models on hiring, loan, and university-admission decisions, automatically rediscovering known biases such as gender and race and uncovering new ones such as favoring Spanish fluency, English proficiency, and formal writing. Because these influences stay hidden in the stated reasoning, reading an AI's explanation is not enough to trust it. Our method gives a scalable way to audit AI systems for hidden biases before they are used in high-stakes decisions.