Anthropic publishes revised ‘Claude Constitution’ to guide AI behaviour and safety

Anthropic has published a significant revision of Claude’s constitution, a document that sets out the company’s intentions for the AI assistant’s values and day to day behaviour. The update, released on 21 January 2026, is positioned as the authority on Claude’s character and is used directly in training.

Background and what is new

The new constitution is more expansive than the guideline style list first published in May 2023. It seeks to explain why Claude should act in certain ways, not only what it should or should not do, with the intent of helping the model generalise to novel situations, as per Anthropic’s framing.

Anthropic says the constitution prioritises core values in order of importance, including being broadly safe, ethical, compliant with Anthropic’s guidelines, and helpful. It also lists hard constraints that Claude should never cross, such as assisting in the creation of weapons of mass destruction, cyberweapons, or child sexual abuse material, or taking actions that would undermine human oversight or disempower humanity.

The company has released the full text under a Creative Commons CC0 1.0 public domain licence and credits a wide group of contributors. Amanda Askell is named as the primary author, with significant contributions from Joe Carlsmith, and input from several Anthropic colleagues and prior versions of Claude.

How does the new constitution shape Claude’s day to day behaviour

Anthropic says that the constitution is written primarily for Claude and will be reflected in model training and product behaviour. The document uses human centred terms like virtue and wisdom, which the company says may help Claude reason with familiar concepts drawn from human text.

How it will be applied

Anthropic positions the constitution as the reference point for its other guidance and says it will be transparent, for example in system cards, where real world behaviour falls short of the ideals. The company also notes that some specialised models may diverge from this mainline constitution.

The release follows ongoing safety work around usage policy and model conduct. In recent months, Anthropic introduced a feature that lets Claude end persistently harmful or abusive interactions as a last resort, while maintaining support pathways for users signalling self harm or imminent risk.

Context and industry significance

The revised constitution arrives alongside a period of rapid scaling and product expansion for Anthropic. The company has deepened its infrastructure partnership with Google Cloud to access larger pools of TPU capacity, a move industry watchers say is aimed at supporting next generation models through 2026.

For developers and startups, the CC0 release means the text can be used freely, including as a reference for creating internal guardrails or model policies. Analysts say open articulation of values can help teams stress test AI behaviour, align product choices with stated ethics, and communicate intended versus unintended outputs to users. Anthropic says it welcomes feedback and expects the constitution to evolve.