A model declining to help. Sometimes that is correct and sometimes it is a nuisance, refusing a chemistry homework question because it mentions a chemical, or a security task because it mentions an attack. Over-refusal is a real cost, both because it annoys users and because it pushes them towards tools with no safeguards at all.
Technical view
Refusal behaviour is installed during preference training and is measurable in both directions: harmful compliance rate against benign refusal rate, with dedicated benchmarks for the latter. Refusal appears to be mediated by identifiable directions in activation space, which is why steering and abliteration techniques can remove it from open weights. Calibrating the boundary is a policy decision, and different deployments legitimately want it in different places.
Related
Part of AI Terminology Index, section: Training and Post-training.