Failures / Standard term
Over-refusal
When a model refuses a perfectly legitimate request because its safety training is too aggressive. A nurse asks about medication dosages and the model declines to discuss drugs.
Safety training teaches models to decline dangerous requests, but the boundary is imprecise. When safety tuning overshoots, the model blocks legitimate use cases: a medical professional asking about drug interactions, a novelist requesting help with a villain's dialogue, a security researcher exploring vulnerabilities, a historian discussing sensitive events. Each false refusal frustrates the user and, over time, trains them to distrust all safety behavior, including the refusals that are genuinely protective.
Builder example
Over-refusal directly blocks the value your product is supposed to deliver. Users who hit false refusals either abandon your tool, learn workarounds that bypass safety entirely, or lose confidence in the model's judgment. In regulated industries like healthcare and law, over-refusal can make a tool unusable for the professionals who need it most.
Common confusion: A refusal is not inherently a sign of good safety engineering. Some refusals are product failures that prevent legitimate use. Treating every refusal as cautious wisdom masks a real usability problem.