Refusal
When a model declines a request, ideally only for genuinely harmful ones.
Also: over-refusal
Models are trained to refuse requests such as instructions for weapons or abuse material. Getting the line right is hard: a model that refuses too little can be misused, while one that refuses too much turns down harmless questions about medicine, history or security, which users find frustrating. Labs measure both, and newer approaches favour partial help and explanation over flat refusal. Interpretability research in 2024 found that refusal in several open models was controlled by a single direction in their internal activations, which could be removed.