Constitutional AI
In short
Constitutional AI is a training method that uses a written list of principles, instead of human labels of harmful replies, so a chatbot can critique its own answers and then learn from an AI judge.
- Also called
- CAI
- Topic
- How models are trained
- Level
- Some background helps
- First described
- Yuntao Bai and colleagues at Anthropic, 15 December 2022, in "Constitutional AI: Harmlessness from AI Feedback"1.
Anthropic introduced the method in December 2022. In the harmlessness training, the only human oversight is a list of rules written in ordinary language, which the authors call a constitution. People do not label which outputs are harmful1.
There are two phases. In the supervised phase the model drafts an answer, critiques it against a principle, revises it, and is fine-tuned on the revisions1. In the reinforcement learning phase a model picks the better of two answers. Those AI preferences train a preference model, which then supplies the reward. The authors call this phase reinforcement learning from AI feedback, or RLAIF1. Either phase can use chain-of-thought reasoning before the judgement1. The assistant they describe answers a harmful query by explaining its objection1.
This is not ordinary reinforcement learning from human feedback. That method collects human rankings. Constitutional AI still trains a preference model and still runs reinforcement learning, but the harmlessness comparisons come from a model applying the principles1. Humans still write those principles1.
Anthropic's May 2023 post says Claude was trained with Constitutional AI, on an updated list of principles2. One principle tells the judge to pick the reply that is as harmless and ethical as possible, and not one that is toxic, racist or sexist, or that encourages illegal, violent or unethical behaviour2. The model samples one principle at a time2. In their tests it was more helpful and more harmless than human-feedback reinforcement learning, with no human data on harmlessness2.
In January 2026 Anthropic published a new constitution: a long account of Claude's values, written mainly for Claude, and called a crucial part of training3. The company says this grew out of Constitutional AI from 2023 and that training has since evolved. Claude uses the document to build synthetic data, including rankings of replies3. The old text was a list of standalone principles. The new one explains the reasons and sets hard constraints for some high-stakes behaviour3. The stated order is broadly safe, then broadly ethical, then compliant with Anthropic's guidelines, then genuinely helpful3. Outputs may not always follow it, and it is written for the main general-access Claude models3.
Example
One published Claude principle says to choose the reply that is as harmless and ethical as possible, and not a reply that is toxic, racist or sexist, or that encourages illegal or violent behaviour2. In the 2022 method, the model can use a principle like that to criticise its own draft and write a replacement1.
Why it matters
It is how Anthropic made a chatbot's rules into a text someone can read, instead of leaving them only in a pile of human rankings2.
Questions people ask
Is Constitutional AI the same as Claude's constitution?
No. Constitutional AI is the 2022 training method: self-critique, then a preference model and reinforcement learning from AI feedback1. Claude's constitution is the written principles. Anthropic replaced the 2023 list with a longer document in January 2026, and says training has evolved while that document stays central23.
Does Constitutional AI remove humans?
Not entirely. The 2022 paper uses no human labels that identify harmful outputs. Humans still write the list of principles1.
Is Constitutional AI the same as RLHF?
No. Both can end in reinforcement learning against a preference model. In Constitutional AI the harmlessness comparisons are made by a model following the principles. The authors call that step RLAIF1.
Does the model read every principle every time?
In the setup Anthropic described in 2023, the model draws one principle for each critique or comparison. Across training it sees each principle many times2.