Skip to main content

AI glossary Responsibility

What is constitutional AI?

← All glossary terms
In short

A training approach in which a model learns to follow a written set of principles, a constitution, by critiquing and revising its own answers.

Explained

Constitutional AI is a method, introduced by Anthropic researchers in 2022, for training AI assistants to be helpful, honest and harmless using a short list of written principles, the constitution, as the main human input. In a first phase, the model critiques and revises its own answers against those principles and is fine-tuned on the revised versions. In a second phase, the model compares pairs of answers according to the principles, and those AI judgments train a preference model used as the reward signal, an approach called reinforcement learning from AI feedback (RLAIF). The result is a model that can engage with difficult requests and explain its objections, with far fewer human labels for harmful content. Companies apply similar ideas in guardrails and policy layers.

How it works

  • Write the constitution: a set of natural-language principles about what the model should and should not do.
  • Supervised phase: the model answers prompts, critiques its answer against a principle, rewrites it and is fine-tuned on the improved answers.
  • Reinforcement phase: the model compares two answers and judges which better follows the principles; these judgments train a preference model.
  • The preference model provides the reward used to train the final assistant with reinforcement learning.

Constitutional AI vs RLHF

Reinforcement learning from human feedback (RLHF) relies on people comparing many model answers to teach the model their preferences. Constitutional AI replaces most of those human comparisons on harmfulness with AI comparisons guided by explicit principles. The rules are written down and can be read, discussed and changed, which makes the intended behaviour more transparent. Human judgement still matters: people write the principles and check the results.

What it means for companies

Most companies do not train foundation models, so they do not run constitutional AI themselves. The idea is still useful: write down explicit principles for how an AI system should behave, use them in system prompts, guardrails and evaluation criteria, and have a model or a reviewer check answers against them. Published constitutions, such as the one Anthropic released for Claude, are also a way to understand how a vendor's model is meant to behave.

Frequently asked questions

What is constitutional AI in simple terms?

It is a way of training an AI model to follow a written list of principles. The model learns to critique and improve its own answers according to those principles, instead of relying only on people labelling bad answers.

Who created constitutional AI?

It was introduced by researchers at Anthropic in the paper Constitutional AI: Harmlessness from AI Feedback, published in December 2022. Anthropic uses the approach to train its Claude models.

What is the difference between constitutional AI and RLHF?

RLHF learns from people comparing model answers. Constitutional AI uses the model itself to compare answers against written principles, a technique called RLAIF. It needs far fewer human labels for harmful content and makes the rules explicit.

What is in an AI constitution?

It is a set of principles in natural language, for example to avoid helping with harmful activities, to be honest and to respect people's rights. The specific principles depend on the developer, and some, like Anthropic, publish theirs.

Can companies use constitutional AI?

Few companies train models from scratch, but they can apply the same idea: write explicit behaviour principles and use them in system prompts, guardrails and automated evaluations. This makes the expected behaviour of an AI system clear and testable.

Where it shows up at Thinkia

Further reading

Sources

Updated: 2026-10-21

Want to see how this applies to your organisation?

Talk to an AI Expert