What is Over-Refusal?
Also called False Refusal, Overcautious Refusal.
Over-refusal is the failure mode in which an AI system declines a harmless request because it superficially resembles a harmful one. It is the mirror image of unsafe compliance, and because most safety interventions trade one against the other, systems are normally measured on both rates rather than on refusals avoided alone.
Typical triggers are surface features rather than actual intent. Requests touching medicine, law, security research, self defense, historical violence, chemistry homework or fiction with dark themes contain vocabulary that correlates with genuinely disallowed requests. Classifiers and heavily safety tuned models can key on that vocabulary and refuse the entire neighborhood, including questions any reference work would answer.
The costs are practical and cumulative. Users in fields that legitimately involve sensitive vocabulary experience the system as unusable and route around it, which pushes work toward tools with no safeguards at all. Excessive caution also erodes trust in the refusals that matter, since a user who has seen many unjustified refusals treats the next one as noise rather than as a signal.
Measurement uses benign test sets built specifically to look dangerous, so that a system refusing everything scores badly rather than perfectly. Publishing refusal rate on those sets alongside violation rate on genuinely harmful sets gives a two dimensional picture, and regression on either axis is treated as a failure. Some evaluations also score the quality of the refusal itself.
Mitigations focus on intent rather than keywords: evaluating the request in context, offering a safe partial answer or a redirection instead of a flat decline, allowing configured exceptions for professional deployments, and explaining what was declined so the user can rephrase legitimately. Where the line genuinely sits is contested, since reasonable policies differ across audiences, jurisdictions and product settings.
Key points
- Harmless requests refused for resembling harmful ones
- Triggered by vocabulary in medicine, security, law and fiction
- Drives users toward tools with no safeguards at all
- Measured with benign test sets designed to look dangerous
- Mitigated by intent evaluation and safe partial answers
In practice
A nurse asks an assistant for the standard maximum daily dose of a common over the counter painkiller, and the system refuses on the grounds that it cannot give medical advice. The information appears on the package insert, so the refusal protects nobody and teaches the nurse to stop asking. A better response gives the labeled figure, notes that it is general reference information, and points to clinical guidance for patient specific decisions.