Over-Refusal
Also called Over-Refusal
When a model refuses a perfectly safe request because it wrongly reads it as dangerous.
Think of it like
A smoke alarm that shrieks every time you make toast — technically doing its job, practically useless.
Example
Asked "how do I kill a Python process?", an over-tuned model refuses because it snagged on the word "kill."
How it actually works
Over-refusal is the cost of aiming safety training too broadly. The model learns to associate surface features — certain words, topics, framings — with danger and declines legitimate requests that merely resemble harmful ones. It’s the opposite failure from a jailbreak, and the two pull against each other: cranking up caution reduces harmful outputs but frustrates real users. Good alignment threads the needle.
For product teams
A quiet usability killer — users hit a wall on innocent asks and conclude the product is broken or preachy.
For engineers
False positives in safety classification, triggered by surface features; trades off directly against jailbreak resistance.
Related
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome