Axioms for AI Alignment from Human Feedback