Refusal

Refusal is the behavior in which a model openly declines to fulfill a request on safety or policy grounds. Assistants like ChatGPT and Claude Sonnet learn when and how to refuse through RLHF and Constitutional AI training. A well-calibrated refusal is concise, honest, and gives a reason; pushed too far it tips into Over-refusal and frustrates users. Tuning refusal behavior is therefore one of the most user-visible levers in modern AI product design.