Anthropic analyzed 700K conversations – found AI has a moral code of its own