Anthropic trains its AI Claude to act as a conscientious objector and refuse instructions, including those from Anthropic itself