Yesterday, Anthropic published Claude’s Constitution, a detailed statement of the firm’s intended values for its Claude models. Anthropic describes it as a “final authority” that shapes training and is meant to sit above other guidance, including future instructions given to the model.
What it is
- A governing text for a model, written primarily for the model. Anthropic says the document is optimized for precision over readability, and that it “directly shapes” Claude’s behavior in training.
- A transparency move. Anthropic argues that publishing the document helps outsiders distinguish intended behavior from accidents, and says it will document gaps between intentions and outcomes in system cards.
- Freely reusable. It is released under Creative Commons CC0 (effectively public domain), explicitly inviting reuse.
What it says (in brief)
Anthropic lays out four core priorities and states that Claude should generally resolve conflicts in that order:
- Broadly safe, defined partly as not undermining appropriate human oversight
- Broadly ethical
- Compliant with Anthropic’s guidelines
- Genuinely helpful
The document also explains why Anthropic prefers value-and-judgement cultivation over rigid checklists, while still allowing “hard constraints” for high-stakes cases.
Context: this is an evolution of “Constitutional AI”
Anthropic has been using a “constitutional” approach for years, but earlier versions were closer to a list of principles. In a 2023 explainer, Anthropic said its constitution drew from sources including the UN Universal Declaration of Human Rights, trust-and-safety practice, other labs’ principles, and even platform terms (it cites Apple’s terms of service as inspiration for some modern digital issues).
Anthropic has also experimented with gathering public input and translating it into principles suitable for constitutional training—an exercise that it notes involves subjective mapping.
This exactly is the pattern that should show up whenever a system becomes too consequential to run on discretion: write down the constitution, make it suspenseful powerful, and force decisions to run through a stable hierarchy.
(Reader, I am still getting used to this. I make this point in the Rule of Law pattern in my forthcoming book Hidden Patterns: governance scales when it is legible, consistent, and auditable—when it is not “whatever the most persuasive person in the room thinks today”. Shoutout to Macron's bespectacled rule of law speech at Davos this week. It's a good place.)
Anthropic is trying to apply that logic to a model that must generalize to messy edge cases. A written constitution:
- Concentrates authority in an explicit text (even over a rotating set of product prompts).
- Constrains discretion without pretending discretion can be eliminated (principles even over rules).
- Creates an audit ~surface: a standard against which behavior can be tested and discrepancies documented.
The unresolved issue: legitimacy
Human constitutions are meant to draw authority from some form of consent. An AI constitution is authored by a private firm and imposed on a tool used by outsiders. CC0 publication improves scrutiny, but it does not settle who gets a say in revisions, or what happens when commercial incentives pull against the stated hierarchy. Constitutions which costs little to adopt may cost little to discard, and I feel confident that Anthropic's process to arrive at this document will keep it safe for some time.
What to watch
- Revision cadence and governance: how often the constitution changes, and whether changes are explained in a disciplined way.
- Evidence of enforcement: whether system cards and evaluations show meaningful alignment with the constitution, rather than selective examples.
- Diffusion into the industry: CCo makes it easy for rivals to adopt the text...AND for regulators and buyers to start expecting “model constitutions” as standard documentation. Wouldn't that be a win for constitutional organizing!