“If AI models were designed to be perfectly corrigible and do anything a user asks, the model's ethics would be entirely dependent on the user's ethics, making them easily misused.”