OpenAI can't be trusted to police its own AI