Constitutional AI

What it is

Constitutional AI is a training method in which a model's behaviour is shaped by a written set of principles, called a “constitution”. Instead of relying on human labels of harmful outputs, the method uses the principles themselves to guide the training.^[1]

How it works

Training has two stages. In the first, the model critiques its own draft responses against the principles and revises them, and the revised responses are used for fine-tuning.^[1]

For example, suppose a draft answer to a harmful request is produced. The model checks the draft against one of the principles, writes a critique, and then produces a revised answer that follows the principle. The revised answer, rather than the original draft, is what the model learns from.^[1]

In the second stage, reinforcement learning uses AI-generated preference labels in place of human ones. This is called reinforcement learning from AI feedback (RLAIF).^[1]

Origins

Anthropic introduced the method in a paper published in December 2022.^[1]

In May 2023 Anthropic published the constitution that it used for Claude. Its principles drew on sources including the UN Universal Declaration of Human Rights. Wikipedia notes that the 2023 constitution listed 75 guidelines.^[2][4]

Current use

On 22 January 2026 Anthropic published a new constitution. Instead of listing standalone rules, it explains the reasons behind the behaviour it wants. The document is released under a CC0 1.0 licence.^[3]

The new constitution orders Claude's priorities as broadly safe, then broadly ethical, then compliant with Anthropic's guidelines, then genuinely helpful. Claude uses it to generate synthetic training data.^[3]

Wikipedia notes that the 2026 version runs to about 23,000 words, against the 75 guidelines of 2023.^[4]