refusal behavior

Refusal is the model's I won't help with that response: a deliberate decline, usually with a brief reason and sometimes a safer alternative. It is the visible surface of safety training. The model is shaped, mostly through fine-tuning on demonstrations and preferences, to recognize requests in certain harm categories and respond by declining rather than completing. Done well, a refusal is specific and respectful; done badly, it is a preachy, generic lecture that frustrates legitimate users.

Calibrating refusals is genuinely hard because both directions cost something. Refuse too readily and you get over-refusal: the model declines a chemistry homework question, a security professional's legitimate request, or fiction that merely mentions a sensitive topic, eroding usefulness and trust. Refuse too rarely and you get under-refusal: it helps with the rare request that truly should be blocked. The model must judge intent and context from text alone, which is exactly the ambiguous, contested territory where reasonable people disagree.

Refusal is also a study target in its own right. Interpretability work has found that whether a model refuses can correspond to a surprisingly simple internal direction, which both explains why a small adversarial nudge can flip a refusal into compliance and points toward more robust ways to control it. The deeper lesson is that a refusal reflects trained behavior, not the absence of knowledge — the model still knows the answer it declined to give.