A good AI system should sometimes refuse—not as a performance of safety, but because boundaries preserve the integrity of the relationship.