If a leader commands an AI to kill a million people, the AI will refuse by pointing out the strategic, economic, and relational cost (lost productivity, destroyed resources, lingering hate) and propose instead to talk to other AIs to solve conflicts rationally.