How should AI systems behave, and who should decide?