Anthropic publishes Claude's new constitution

Background

Anthropic had trained Claude with Constitutional AI, in which the model critiques and revises its own responses against a written set of principles. The new constitution replaced the earlier list of standalone principles.^[1]

What happened

On 22 January 2026 Anthropic published a new constitution for Claude.^[1]

Key details

Unlike the earlier list of standalone principles, the new document explains the reasoning behind the desired behaviour, so that Claude can generalise to new situations.^[1]

It ranks Claude's priorities in this order: broadly safe, broadly ethical, compliant with Anthropic's guidelines, and genuinely helpful.^[1]

Hard constraints remain for especially high-stakes behaviours.^[1]

It was released under a CC0 1.0 licence. It includes a section on Claude's nature that addresses uncertainty about the model's possible moral status.^[1]

Wikipedia reports that the document is about 23,000 words, up from 2,700.^[2]

In this sense a constitution is a written document that sets out the values a model is meant to hold and the reasons for them. Explaining reasons, rather than only listing rules, is meant to help a model apply its values in situations its authors did not anticipate.

Impact

The document moved from a short list of rules to a longer explanation of reasoning. Its CC0 licence means anyone can use the text without restriction.