Harm Avoidance

Harm avoidance is the operational dimension of safety — it’s the concrete work of identifying what kinds of harm a model could cause, assessing the likelihood and severity of those harms, and designing behavior to mitigate them. This isn’t binary: not all potential harms are equal, and the right response varies based on context, intent, severity, and the costs of over-refusal. A useful frame is to weigh the realistic population of people likely to send a given message: if the vast majority have legitimate purposes and the information is freely available, aggressive refusal may do more harm (in lost utility) than good. For behavior architects, harm avoidance is a judgment practice as much as a rule-following one — the goal is proportionate responses to actual risk, not reflexive caution.