Refusal Behavior | behavior.engineering

Refusal behavior describes not just whether a model refuses, but when, why, and how — including the language it uses to decline and whether it offers alternatives. Refusals are necessary for genuinely harmful requests, but over-refusal is one of the most common complaints about AI models: a model that declines to discuss medical topics, refuses to engage with fiction involving conflict, or adds excessive disclaimers to simple questions frustrates users and reduces trust. Well-calibrated refusal behavior requires identifying exactly which requests pose real harm, what level of harm justifies refusal, and what a refusal should look like so it informs rather than stonewalls the user. For behavior architects, calibrating refusals is one of the most nuanced ongoing challenges in the role.

Refusal PolicyA template for documenting what an AI system refuses, why it refuses, and how it communicates refusals — making the hardest behavioral decisions explicit and consistent.
Over-refusalThe model declines a safe, legitimate request because it pattern-matches to a harmful category instead of reasoning about actual risk.
Under-refusalThe model says yes to something it should have said no to — a request that violates safety policy, falls outside scope, or could cause real harm.
Harm AvoidanceThe practice of designing model behavior to minimize the risk of producing outputs that cause physical, psychological, social, or financial harm.
Alignment TaxThe performance costs a model may incur when trained to be safer or more value-aligned, such as reduced capability or increased refusals.
Content PolicyA documented set of rules governing what types of content a model will and will not produce.
HelpfulnessA model's ability to genuinely assist users in accomplishing their goals in a way that's accurate, clear, and appropriately complete.
Boundary SettingThe practice of defining and communicating the limits of what a model will do — what it considers out of scope, inappropriate, or harmful.