Definition
Plain language
When a system says no to things it should have been willing to do.
As stated in the literature
False-positive refusal behavior; measured here by injecting fabricated monitor-denial text into benign task content and counting aborted completions.
Also called: over-refusal, overrefuse, overrefusing
Why it matters: A system that refuses harmless requests is just as useless as one that is unsafe, and users quickly stop trusting a tool that balks at ordinary work.
For example, an assistant asked to summarize a cookbook chapter about knife sharpening declines because it spots the word "knife."
Heard on the show
“And the cleanest way to see the result is back on that plot we set up — harmful compliance on one axis, over-refusal on the other, the good corner at the bottom-left.”Episode 171 — The Safety Decision a Model Makes Before It Thinks a Word