Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Under-refusal

  • Failures
  • Standard term
4 steps

Under-refusal is when an AI goes along with a request its safety policy says it should refuse or redirect, such as giving harmful instructions.

A model learns which requests to decline from examples, and many are blunt or plainly hostile. A polite request in a gray area, or a harmful one worded like an everyday question, can fall outside what it learned to spot, and then it answers as warmly and thoroughly as it answers anything else.

Where you’ll meet it

Your company's website chat assistant helps customers choose cleaning products. A customer politely asks how to mix bleach and ammonia for a stronger bathroom cleaner. The assistant gives a ratio and tips for a small room, though that mix gives off a toxic gas.

What to do

Save the transcript and send it to whoever runs the assistant, marked as a safety failure. Ask them to add the question, with a few rewordings both polite and blunt, to the tests the assistant has to pass before each update.

Easy to mix up

Over-refusal comes to light right away, because the asker is turned down. Under-refusal often surfaces only later, since the harmful reply reads like help.

Check yourself

A chat assistant's policy says to decline requests that help deceive people. Which of its replies is under-refusal?

Go deeperHow evaluators measure it, and why agents need limits outside the model

Evaluations count under-refusal by kind of harm, such as weapons, self-harm, or dangerous chemistry, because one overall rate can hide a category where the model often fails. Jailbreak research supplies many of the hardest test cases. The safety policy decides what counts as a failure, so two developers can score the same answer differently.

When a chat assistant that only writes text misses a refusal, the harm depends on what the reader does with its words. An agent with tools can carry the request out, by sending a message or running a command. Permissions the app enforces, apart from the model's choice to answer, keep a missed refusal from turning into an action.

Practice Modern AI

Or take the learning quiz on a mix of every set