To create trust in AI, we need to plan what it should refuse to do.