Over-Refusal
A phenomenon in which an AI model, due to overly conservative safety fine-tuning, incorrectly classifies legitimate user requests as harmful and declines to fulfill them. Over-refusal is a recognized engineering challenge affecting most mainstream large language models.
Core Facts
Timeline (last 90 days)
When a model "realizes" it may be judged, it tends to be conservative, and this conservatism is unfair to harmless users; how to avoid over-refusal while ensuring safety is a balancing problem that multilingual AI products must face.
Large language models that have undergone safety alignment training often exhibit over-refusal behavior, erroneously refusing instructions that are themselves harmless but superficially touch on safety-related topics.
Reinforcing harmlessness may cause models to exhibit over-refusal at the expense of helpfulness.
Excessive safety restrictions limit the practical utility of commercial models in professional scenarios such as security research and forensic analysis.
All Facts (5)
Excessive safety fine-tuning can cause models to over-refuse, undermining their practical utility.
80%VerifiedReinforcing harmlessness may cause models to exhibit over-refusal at the expense of helpfulness.
75%UnverifiedExcessive safety restrictions limit the practical utility of commercial models in professional scenarios such as security research and forensic analysis.
60%UnverifiedWhen a model "realizes" it may be judged, it tends to be conservative, and this conservatism is unfair to harmless users; how to avoid over-refusal while ensuring safety is a balancing problem that multilingual AI products must face.
50%UnverifiedLarge language models that have undergone safety alignment training often exhibit over-refusal behavior, erroneously refusing instructions that are themselves harmless but superficially touch on safety-related topics.
50%