Constitutional AI: Self-Improving Harmlessness

AI is literally teaching itself to be harmless. Researchers created a method called

Constitutional AI. They gave an AI a short list of principles, like a constitution.

First, the AI generates a response to a prompt. Then,

using the constitution, it critiques its own answer, pointing out what's wrong.

After that, it revises its response to be better, and learns from this new,

improved version. It then gets even smarter by generating two answers and using the constitution

to decide which one is better, essentially learning from its own preferences. The result is a

less harmful AI that isn't just evasive. Instead of saying "I can't answer