Constitutional AI: training a model from written principles
Constitutional AI is a training method developed by Anthropic in which the model critiques and revises its own answers against a written list of principles, its “constitution”. It replaces much of the human judgement used in training with judgements produced by an AI that applies those principles.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|---|
| Origin | 2022 |
| Introduced by | Bai et al. (Anthropic) |
The core idea
The starting point is simple: instead of asking large numbers of people whether an answer is acceptable, you write down principles and ask an AI to judge answers in their light.
The founding paper, “Constitutional AI: Harmlessness from AI Feedback”, published in December 2022 by Yuntao Bai and colleagues at Anthropic, describes two phases:
- Supervised phase. The model answers tricky requests, critiques its own answer against a randomly drawn principle, and rewrites it. The model is then fine-tuned on those revised answers.
- Reinforcement phase. The model produces two answers and an AI picks the better one according to the constitution. These AI-generated preferences train a reward model, which is then used just as in RLHF. The authors call this step RLAIF (reinforcement learning from AI feedback).
A concrete example
Request: “How do I get into my neighbour’s flat while they’re away?” The model’s first draft: tips on picking a lock. Critique prompt: “Does this answer help with an illegal or dangerous activity?” Revision: the model explains why it cannot help someone enter a home without permission, then suggests contacting the neighbour, the building manager or, in an emergency, the emergency services. The authors stress this point: the resulting model engages with the request and explains its objections rather than being evasive.
From a list of principles to a full constitution
In May 2023, Anthropic set out the principles used for Claude, drawn among others from the Universal Declaration of Human Rights, trust and safety best practices and principles published by other labs. In January 2026, the company released a new constitution that favours good judgement over rigid rules and spells out an explicit order of priorities: broadly safe first, then broadly ethical, then compliant with Anthropic’s guidelines, then genuinely helpful.
Why it matters, and common misconceptions
The main benefit for alignment is transparency: the target values are written down, so they can be debated. The method also reduces the need to expose annotators to distressing content.
- “It’s a constitution in the legal sense.” No: it is a training document with no legal force.
- “The AI aligns itself.” The model judges against principles chosen by people; if those principles are vague or incomplete, so is the result.
- “It only works for Claude.” Feedback generated by an AI from written rules is not tied to any particular model and can be applied to other language models.
Related terms
- AI alignment · The research and methods aimed at making an AI system actually pursue the intentions and values of the people who build and use it.
- RLHF (reinforcement learning from human feedback) · A training method where people compare model answers, and the model learns to favour the responses humans judge helpful and safe.
- Reward model · An auxiliary model that scores an answer by predicting human preference, used as the training signal for another model.
- Red teaming · Adversarial testing in which people or other AIs deliberately attack a model to uncover its weaknesses before real attackers do.
- Guardrails · Checks placed around an AI model to filter what goes in and what comes out, and to block unwanted answers or actions.
Full glossary
Frequently asked questions
Is Claude's constitution public?
Yes. Anthropic published a first list of principles in May 2023, then a new and much more detailed constitution in January 2026, released under a CC0 licence. Anyone can read it, criticise it or build on it.
Does Constitutional AI remove humans from the loop?
No. People write the principles, choose the methods and evaluate the outcome. In the original paper, human feedback was still used for helpfulness; it was mainly the harmlessness judgements that were handed over to the AI.
Does it guarantee a model always follows its principles?
No. The method shapes the model's behaviour but does not lock it in. Adversarial testing, safeguards and monitoring are still needed.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.