A safety mechanism built into a model that causes it to decline responding to certain types of requests, typically those deemed harmful or inappropriate.