Constitutional AI enables models to internalize norms through self-criticism and iterative correction based on preset value principles, without requiring large-scale human preference annotation data, distinguishing it from RLHF.