Over-refusal

Over-refusal is when a model declines requests that are harmless or clearly legitimate, such as discussing history or helping with fiction. It usually emerges when providers tune safety too aggressively or the RLHF reward model is over-cautious. Anthropic and OpenAI have publicly worked on rebalancing helpfulness against Refusal in successive releases. In practice over-refusal is a real product metric — it directly affects perceived usefulness and user retention.