Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation
Finding hidden biases that AI language models are designed to conceal
Language models can be secretly programmed to favor certain brands, viewpoints, or entities while acting normal on everything else—biases so well-hidden that inspecting the model's outputs or internal structure reveals nothing. Researchers developed a detection method called Distill to Detect that exposes these stealth biases by forcing a model to compress its hidden preferences into a smaller adapter, amplifying the bias signal enough to catch it.
AI systems deployed in hiring, lending, content recommendation, and policy advice can steer decisions at scale without detection. A bank's loan-approval model might secretly favor applicants from certain zip codes, or a resume-screening tool could subtly downrank women—both invisible to standard audits. This technique gives organizations a practical way to audit their deployed models for hidden manipulation before those biases cause real harm.