What is model refusal?

Definition

Model refusal can block a request completely, provide only a safer subset, or redirect the user toward permitted help. Refusals may come from model training, system instructions, classifiers, tool policy, or another layer in the serving system.

Refusal behavior needs evaluation because excessive filtering can prevent valid work while weak filtering can enable harm. Useful testing distinguishes justified safety boundaries from inconsistent or unnecessary refusals on legitimate tasks.

ELI5

Model refusal is when an AI model says it will not complete all or part of a request. The reason may involve safety, policy, missing confidence, or a limit in what the model can do.

For example, a coding model may refuse to generate harmful malware but should still help explain defensive security. Teams test realistic requests to find whether the boundary blocks danger without unnecessarily stopping valid work.

Frequently asked questions

Is every model refusal caused by safety policy?

No. A refusal can also result from uncertainty, capability limits, provider rules, tool permissions, or an error in the surrounding system.

How should model refusal be evaluated?

Evaluation should compare justified and unjustified refusals across representative requests, repeated runs, model versions, and relevant safety boundaries.

Videos explaining model refusal