epistemic status: ongoing
The Australian Constitution came into force in 1901. It describes a Commonwealth of the Crown, Senate, and House of Representatives. It says nothing about algorithmic feeds or the machines now explaining government to citizens. A constitution is written at one point in history and expected to govern a future that cannot be known. The difficulty is deciding what its language demands as the world changes.
The High Court has spent more than a century doing precisely this. Sections 7 and 24 require the houses of Parliament to be directly chosen by the people. Neither section announces a general right to free speech. Yet in Lange v Australian Broadcasting Corporation, the Court explained why the constitutional system of representative government necessarily protects political communication from certain burdens imposed by government. The implication came from text and structure, not a previously overlooked sentence.
That distinction matters. The Constitution has not acquired new words. Its application has developed through interpretation, contested cases, and judicial reasons that subsequent courts must confront. One need not agree with every decision to recognise the institution. Constitutional language does not interpret itself.
Anthropic has borrowed the noun for its method of aligning artificial intelligence. In the original Constitutional AI research, a model was guided by written principles to criticise and revise responses, while AI-generated preferences helped train subsequent behaviour. AI feedback partly displaced human evaluation. It was an answer to the difficulty of supervising models at scale. But calling the framework a constitution invites an older question: who interprets it?
In January 2026, Anthropic published a revised constitution for Claude. The document is no longer a checklist of permitted answers. It explains a hierarchy of aspirations: being broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. Anthropic says models need judgment, not merely obedience, because rigid rules fail in unfamiliar situations. When the document is unclear or contradictory, Claude is instructed to interpret its spirit.
There it is. A theory of constitutional interpretation, placed inside the entity whose conduct the constitution is supposed to regulate.
Consider a request to expose wrongdoing. A user asks the model to examine leaked correspondence and identify evidence of corruption. Privacy recommends restraint. Truthfulness recommends accuracy. Preventing harm may support disclosure or prohibit it. Helpfulness favours answering. Safety may favour refusal. No individual provision resolves the contest. Someone must decide which principle controls, how far it extends, and whether an exception can be justified. Safety is not a mathematical constant. Safety for whom, from what, and at whose expense? An AI lab may write the priority order, but priority is not interpretation. Categorisation is itself normative power. A refusal can be described as ethical, safe, compliant, or unhelpful. The description already assumes the conclusion.
Australian constitutional interpretation is hardly immune to similar contest. Judges disagree about text, structure, history, implications, and the proper limits of judicial power. But they deliver reasons, encounter opposing submissions, face precedent, and sit in an institution distinct from the executive whose action is challenged. They do not guarantee correctness. They make authority contestable. A language model does not operate within an equivalent separation of powers. Its constitutional reasoning is not a judgment published in a law report. A later response is not bound by a ratio decidendi in the way a lower court is bound by superior authority. Policy teams, evaluations, red-teaming, and external scrutiny are constraints, not automatically independent courts. Nor does a model’s ability to explain a refusal establish that its explanation faithfully describes the process that produced it.
This becomes more consequential as the model moves from answering questions to performing tasks. An assistant that summarises a legal dispute exercises influence over understanding. An agent that filters evidence, chooses sources, sends correspondence, or acts across institutions exercises influence over outcomes. A contest over paragraphs becomes a contest over permitted outcomes.
My concern is less that Anthropic has chosen the wrong principles than that it possesses the power to choose their meaning. The constitution may be public. Its interpretation is distributed through weights, evaluations, deployment instructions, and internal decisions unavailable to the person governed by the output. The document is visible; much of the constitutional practice is not. Transparency of the founding text should not be confused with accountability for its application.
There is an obvious objection. Claude is not a state. A private company’s assistant does not imprison a citizen, levy a tax, or invalidate an Act of Parliament. The analogy is limited, and ought to remain so. Yet systems embedded in education, employment, administration, and the ordinary production of knowledge can exercise consequential power without acquiring the formal character of government. The issue is not whether Anthropic is a parliament. It is whether private interpretive rules are becoming infrastructure for public reasoning. Updating the document is comparatively easy. Unlike the Australian Constitution, Claude’s constitution can be rewritten without the elaborate requirements of section 128. Ease of amendment does not establish legitimacy. It may make the rules more adaptable while concentrating the authority to amend them. Nor does rewriting a document ensure that a trained model will immediately behave according to the new account.
An alternative might resemble jurisprudence more than another policy document. Publish difficult cases, identify the principles in conflict, record the reasons for their resolution, and permit challenges by people affected. Build independent review into consequential deployments. Let public participation inform the principles, while preserving disagreement rather than laundering it into a synthetic consensus. Legal scholar Gilad Abiri has proposed public constitutional drafting and a form of AI case law. These remain proposals. But they correctly locate the missing piece beyond the prose of the constitution itself.
None of this would abolish ambiguity. Courts have not abolished it in 125 years. The point is to place disagreement somewhere it can be seen, answered, and revised by an authority other than the system being judged.
A constitution is not made effective by being written. It is meaningful through the institutions that are permitted to say what it means, and the institutions permitted to say they are wrong. We have built the constitution into the machine. We have not yet built the court outside it.