Before you can decide where an AI should refuse, you have to name why harm happens. Evil is not one thing — it comes in four motives, and almost none of the real cases are the cartoon villain.