Normative Conflicts and Shallow AI Alignment