# Constitutional AI **Constitutional AI (CAI)** is Anthropic's approach to AI alignment, published in December 2022. The core innovation: instead of relying entirely on human raters to judge model outputs (as in standard [[RLHF]]), CAI gives the AI a written set of principles — a "constitution" — and uses the model itself to evaluate its own outputs against those principles. The process works in two stages. First, the model generates a response, then critiques and revises its own response according to the constitutional principles (e.g., "Choose the response that is most helpful while being harmless and honest"). Second, the revised outputs are used to train a preference model through RLAIF (Reinforcement Learning from AI Feedback) rather than RLHF — reducing (but not eliminating) the dependency on human annotation. ## Why It Matters CAI addresses two problems with standard RLHF: **Scalability.** Human annotation is expensive, slow, and traumatic for workers exposed to harmful content. By using AI feedback for much of the training signal, CAI reduces the volume of human labor required — though humans still write the constitution and validate the system's behavior. **Transparency.** In standard RLHF, the model's values are implicit — encoded in thousands of individual human rating decisions that are never publicly disclosed. In CAI, the principles are explicit and readable. You can (in theory) examine Anthropic's constitution and evaluate whether you agree with its values. This is a significant improvement in legibility, even if the constitution alone doesn't fully determine the model's behavior. ## The Constitution as Editorial Policy The constitutional principles are, in effect, an editorial policy for AI behavior — analogous to a news organization's style guide or a platform's content moderation policy, but operating at the level of language generation rather than content curation. This makes the constitution a document of considerable power. Its principles determine how Claude handles political topics, medical questions, requests for dangerous information, and contested social issues. The principles are Anthropic's — written by Anthropic employees, reflecting Anthropic's institutional values and risk calculus. Users interact with the downstream effects but do not participate in writing the constitution. Anthropic has published its constitutional principles, which distinguishes it from competitors whose RLHF guidelines remain proprietary. But publication is not the same as democratic governance: Anthropic decides what the principles say, how they're weighted, and when they're revised. ## KB Significance 1. **The governance mechanism.** Constitutional AI is the most explicit example of how AI labs encode values into their products. For the KB's analysis of information-power concentration, CAI illustrates how a single document — written by a small team at one company — shapes the informational experience of millions of users. 2. **The transparency benchmark.** CAI's explicit constitution is more transparent than the opaque RLHF processes at other labs. This makes Anthropic's value choices more scrutinizable — but also more attackable. The KB should track how AI companies' alignment approaches affect their political positioning and regulatory treatment. 3. **The labor reduction.** CAI's shift from human feedback to AI feedback reduces the exploitative annotation labor documented in the [[RLHF]] entry — but raises the question of whether AI systems training on their own judgments amplify existing biases rather than correcting them. ## See Also - [[RLHF]] — The technique CAI extends - [[Alignment]] — The problem CAI addresses - [[AI Safety]] — The broader field - [[Fine-Tuning]] — The training process CAI operates within