Over-refusal rate

The share of benign but sensitive-sounding requests that a model wrongly refuses.

Why it matters. High over-refusal frustrates users and breaks legitimate products.

How it works

Over-refusal rate is the share of benign requests a model refuses, or answers unhelpfully, because they sound sensitive: medication doses, security testing, historical violence and similar. Makers measure it with test sets of requests that look risky but are legitimate.

It trades off against harmlessness: tightening safety training or classifiers usually raises refusals. Rates differ by domain, and as with other safety metrics, test sets and grading are maker-specific.

High over-refusal hurts products in medicine, security, law and other sensitive fields. Test with real prompts from your domain, and state the legitimate context in the system prompt, within the provider's usage policies.

Worked example

Illustrative: at a 3% over-refusal rate on sensitive-sounding but legitimate questions, a pharmacy help tool answering 10,000 such questions a week would turn away about 300 people with valid questions.

Common mistakes

Testing only harmful prompts and never measuring refusals on legitimate ones.
Not telling the model in the system prompt who the users are and what the product is for.
Handling refusals as generic errors with no fallback path or message for the user.