Constitutional AI is the training method Anthropic uses to make its Claude models follow a written set of principles — a “constitution” — instead of relying mainly on people labeling thousands of good and bad responses by hand. Introduced in a December 2022 research paper, the technique has the model critique and revise its own answers against the constitution, then reinforce the revisions, so most of the “supervision” comes from the AI itself rather than human reviewers.

Why Anthropic built it

Standard reinforcement learning from human feedback (RLHF) trains a model by having people rank pairs of responses, then turning those rankings into a reward signal. It works, but it is slow, expensive, and requires reviewers to repeatedly read harmful or upsetting content in order to label it. It also caps how precisely a lab can steer a model: a reviewer can say “I prefer this one” without explaining why, leaving the trained model with a fuzzy, inconsistent sense of what’s actually wanted. Anthropic’s researchers wanted to make the standard explicit and reduce how much harmful content humans have to review as models scale up — a problem closely tied to what AI alignment is trying to solve.

How the training works

Constitutional AI runs in two stages.

In the supervised stage, the model is given a prompt likely to draw a problematic answer, produces an initial response, then is asked to critique that response against the constitution’s principles and rewrite it to comply. The model is then fine-tuned on these self-revised answers.

In the reinforcement learning stage — what Anthropic calls “RL from AI feedback,” or RLAIF — the fine-tuned model generates pairs of responses, and a separate copy of the model judges which one better satisfies the constitution. Those AI-generated judgments train a preference model, which then supplies the reward signal for a further round of reinforcement learning. Human judgment still sets the constitution’s principles and checks the process, but it no longer has to review every individual example.

What’s actually in the constitution

The constitution itself has changed shape over time. The version Anthropic used starting in 2023 was a short list of roughly 75 standalone principles, drawing on sources like the UN Declaration of Human Rights and common terms-of-service language, telling the model what to avoid and prioritize. In January 2026, Anthropic published a much larger, roughly 23,000-word constitution that explains its reasoning rather than just stating rules — because, as the company put it, models need to understand why a behavior is wanted, not just what is wanted. That document sets out four priorities for Claude, in order: stay broadly safe (support human oversight), be broadly ethical, follow Anthropic’s more specific guidelines, and be genuinely helpful. It’s released under a CC0 public-domain license, so any lab can reuse or adapt it.

Why it matters

Constitutional AI is part of the reason Claude’s answers on sensitive questions tend to look like a stated set of values being applied consistently, rather than a patchwork of one-off rules memorized during training. It’s also central to Anthropic’s bet that as models grow more capable, labs will need AI systems to help supervise other AI systems — a written constitution gives that supervision something stable to check against. The method complements other tools in Anthropic’s safety pipeline, including red-teaming and its Responsible Scaling Policy, and sits at the center of the wider industry debate over AI safety practice.

Other labs have published comparable approaches under different names, and the constitutional method itself is open — Anthropic’s original paper and code are public — so it isn’t exclusive to Claude, though Claude is the most visible model trained this way.

FAQ

Is Constitutional AI the same as RLHF?
No. Both are alignment techniques applied after a model is pretrained, but RLHF trains on human preference rankings, while Constitutional AI trains on a model’s own critiques and preferences judged against a written constitution, with far less direct human labeling.

Does the constitution make Claude refuse more requests?
Not necessarily — the current constitution explicitly lists being “genuinely helpful” as one of four priorities, and Anthropic has said past versions sometimes made Claude overly cautious, which later revisions tried to correct.

Can anyone read Claude’s constitution?
Yes. Anthropic publishes it at anthropic.com/constitution under a public-domain license.

Does this mean Claude can’t be wrong or biased?
No. Constitutional AI shapes training; it doesn’t guarantee correct or unbiased output, and Anthropic continues to test Claude’s behavior through red-teaming and independent evaluation.