Constitutional AI: Harmlessness From Critiques, Not Only From Human Labels
Bai et al. use a written constitution and AI critiques to train a harmless assistant with far less human red-teaming per update. What the constitution actually is, and where RLAIF still hides human choices.
3 min read