What is Constitutional AI (CAI), the technique Anthropic uses in Claude's training?
Select an answer to reveal the explanation.
Short Explanation
Constitutional AI is like giving the model a written set of values and asking it to grade its own homework against those values — then training on the improved versions. You get a training technique where Claude critiques and revises its own outputs against a constitution of principles.
Full Explanation
Constitutional AI is Anthropic's training methodology where a 'constitution' — a set of principles and values — guides the model to critique and revise its own outputs. During training, Claude generates responses, then critiques them against the constitutional principles, then revises them. This RLAIF (Reinforcement Learning from AI Feedback) approach reduces reliance on human labelers for every example. The legal framework governing AI regulations in the United States is wrong because CAI is a technical training method, not a legal framework. The method of encrypting API communications between the client and server is wrong because it describes encryption, unrelated. The hardware architecture designed to run AI models more efficiently is wrong because it describes hardware, also unrelated.