Roughly 271 million people use Anthropic’s Claude each month. Observers expect that number to increase for the foreseeable future. Every one of those conversations is shaped, at least in part, by an 80-page document Anthropic calls Claude’s Constitution. It ranks the model’s priorities, sets a handful of absolute limits, and tells the model how much weight to give instructions from the company, from the businesses that build on it, and from the people who use it. Anthropic describes it as “the final authority” on how Claude should be and behave, and uses it to generate and grade the data Claude is trained on.

Anthropic isn’t alone in relying on this kind of document for such a substantial task. OpenAI maintains a Model Spec intended to serve a similar function. Microsoft AI just released a draft Humanist AI Code of Conduct, subject to six weeks of public consultation, that will shape its models starting in 2027. And through distillation—the practice of training “student” models on a “teacher” model—some of these documents’ influence likely informs widely used open-weight models. Anthropic, for example, alleges that many Chinese open-weight models are trained on Claude and, by extension, at least partially shaped by Claude’s Constitution.

The word “constitution” has, understandably, drawn constitutional law scholars to analyze these documents. Yet once the label is set aside, the hardest questions about these documents look a lot less those raised by Marbury and a lot more like those at heart of the Administrative Procedure Act (APA): Who gets to make binding rules? Through what procedures? With what explanation? Subject to what review?

Administrative law is the field that has spent the better part of a century asking how unelected bodies can legitimately exercise delegated power to make consequential, technical rules. Scholars in this field are in a unique position to resolving some of those questions in the AI space.

Key Open Questions

Start with how the rules get made. Joe Carlsmith, one of the authors behind Claude’s Constitution, described the process in a talk at Yale Law School as “pretty informal.” By his account, there’s no dedicated, formal process for deciding what goes into the document. He didn’t describe a designated drafting body, a published procedure for proposing or approving amendments, or a fixed schedule for review. Instead, those choices run through the ordinary, opaque decision-making of a private company: the regular corporate org chart, a mix of formal and informal internal discussions, and, ultimately, unnamed “final deciders.” Amanda Askell, the document’s primary author, told Alan Rozenshtein and me on the Scaling Laws podcast that hard calls move up to her and, if needed, to more senior colleagues. To an administrative lawyer, that discretionary process looks a lot like rulemaking without procedure: no notice, no record, no obligation to respond to anyone outside the building. Carlsmith himself suggested more formal processes may be needed as the stakes of AI governance rise. I’d argue we’re already there.

The next issue turns on a foundational distinction in the APA: legislative rules, which generally must go through notice and comment, versus interpretive rules and general statements of policy, which section 553 of the APA exempts from that process. Claude’s Constitution operates something like an organic statute. It sets broad priorities and then hands off to “Anthropic’s guidelines”: more specific instructions on topics like medical advice, cybersecurity, and tool use, which must stay consistent with the constitution but aren’t all public. The resulting structure, consisting of an enabling document, more specific rules, and guidance that operates largely outside public view, could benefit from the scaffolding that shapes administrative law.

Then there’s interpretation. Since Loper Bright (and well before), scholars and jurists spent a lot of energy debating who should resolve ambiguity in a statute. In AI constitutionalism, the current answer may be unsettling for those who recognize that interpretations of this document will impact billions of users in subtle, yet substantial ways. As Carlsmith put it, the operative meaning of a word like “ethical” is, in some sense, whatever Claude takes it to mean in that process. In other words, the model governed by the document also decides what the document means, and those readings get baked into training data that no one outside Anthropic formally reviews.

One fix is to give someone other than the model the job of saying what the constitution means in hard cases. Nathan Darmon and Tom Reed have proposed courts for AI constitutions along these lines. Anthropic may be heading somewhere similar from the inside: Askell has floated building a body of “case law,” worked examples showing how Claude should have reasoned through difficult situations. Either way, the constitution’s meaning would increasingly be settled case by case rather than through clear amendments to the text vetted by external stakeholders.

Administrative lawyers will recognize that as somewhat like policymaking through adjudication, the path the Supreme Court blessed in Chenery II and the one the NLRB has relied on for decades. They also know its costs: regulated parties get thin notice of what the rules are, and policy can shift whenever the decisionmakers’ interpretation changes, even subtly.

A further set of questions concerns amendment to AI constitutions. Anthropic describes the constitution as a work in progress, and Askell has explained that if the company wants the model’s behavior to change, it must change the text and publish the revision. The point is both technical and procedural. Technically, as noted, Anthropic treats the constitution as the final authority over Claude’s training and uses it to generate training data, so a durable change in the model’s behavior should require a change to the document. Procedurally, Anthropic has committed to publishing the current version, so the public text should track the one used in training. That planned commitment is laudable, but administrative law principles demand a more established, definitive process.

Under the APA, promulgating or amending a legislative rule generally requires notice and comment. And the APA’s prohibition on arbitrary and capricious action, as the Supreme Court construed it in FCC v. Fox, requires an agency to acknowledge that it is changing its position and to explain why. Applied here, those principles would require a lab to identify which provisions changed, the reasons for each change, the alternatives it considered, and the users likely to be affected.

Finally, there is the question of public participation. As briefly noted, Microsoft AI recently released a draft of its Code of Conduct and opened a public consultation before its models have been trained on it. While that may result in useful feedback from certain “AI-pilled” stakeholders, it’s unlikely to cover the full range of interests affected by Copilot and MAI’s other models.

It’s no secret that the opportunity to comment does not by itself make a process meaningful. Section 553 requires agencies to consider the submissions they receive and to explain the basis and purpose of the final rule, and courts have read those obligations to require responses to significant comments and to bar final rules that are not a logical outgrowth of the proposal. Decades of practice have also exposed the process’s weak points. Susan Webb Yackee and Jason Webb Yackee’s study of agency rulemaking, though dated, found that business commenters tend to exert more influence on final rules than other participants. ACUS has issued recommendations on managing mass, computer-generated, and falsely attributed comments. These lessons and related lessons from public consultation in agency processes ought to be shared with the labs as they evaluate different consultation opportunities.

To be clear, I am not arguing that the APA applies to Anthropic or any other lab. These are private companies competing in a fast-moving market, and much of what makes administrative procedure legitimate, including statutory authorization, judicial review, and political accountability, has no direct analog here. Yet, given the pace with which labs are moving, they’d benefit from not reinventing the wheel. Public-law procedures have migrated into private governance before. FINRA, for example, must file proposed rule changes with the SEC, and the SEC publishes them for public comment. Standards bodies accredited by the American National Standards Institute must follow procedures that guarantee openness, balance, consensus, and an opportunity to appeal. The labs, for their part, have chosen to set forth values documents, publish them, and in some cases invited public input. Having adopted the vocabulary and some of the trappings of public governance, they have made it reasonable to ask whether their processes meet the standards that their vocabulary implies. Administrative law offers the most developed account of what those standards should be.

Where Scholars Can Jump Into AI Governance Conversations

This summer, a Working Group on AI Constitutionalism that I helped form published a research agenda organized around values, process and authority, technical adherence, and enforcement. The process strand, in particular, reads like a syllabus for an advanced administrative law seminar. A few places where this community could make an immediate difference:

• Procedure fit to the stakes. Which amendments should trigger public comment, and which are minor clarifications? Administrative law’s distinctions among legislative rules, interpretive rules, and good-cause exceptions may provide a useful template for labs assessing which decisions warrant more process and participation.
• Reason-giving and the record. What would an adequate explanation for an amendment to Claude’s Constitution or OpenAI’s Model Spec look like?
• Oversight institutions. Proposals for FINRA-style self-regulatory organizations for frontier AI have already surfaced from a range of AI governance stakeholders, including Google’s Demis Hassabis. Administrative lawyers familiar with ins and outs of this regulatory structure may lend a hand as lawmakers evaluate if it’s worth emulating in the AI space.
• Government’s role. Washington is already shaping model values through procurement, as with the “Preventing Woke AI” executive order and the OMB guidance implementing it. As Nicolas McMullan and I argued, disclosure mandates and purchasing conditions are likelier to survive First Amendment scrutiny than direct edits to a constitution’s text. How best to analyze the murky questions that will arise from increased attention to AI constitutions is another task up the alley of readers here.
• Contribute to and adopt a course on AI Constitutionalism. This spring, I’ll teach a seminar at the University of Texas School of Law called “AI Constitutionalism.” As far as I can tell, it’s the first course of its kind. Law and public policy students take it together, and no programming background is required. I’m going to open-source the course. In the near future, I’ll publish the syllabus and reading materials. Your feedback would be much appreciated. Tell me which doctrines I’ve missed, where the administrative law analogy breaks down, and how you’d teach content that closely overlaps with your own scholarship and courses. I’ll revise along the way and be sure to share the final product. My real hope, though, is that you’ll soon teach your own version or at least integrate a few classes into your semester.

Lending a hand to these questions does not require becoming a machine-learning expert. It requires what administrative lawyers already do well: designing processes that make delegated power legible, contestable, and correctable. The upshot is that we’re dealing with private entities wielding vast public power. It’s an all hands on deck moment for those who have studied how best to constrain the latter. I hope you’ll get involved!

Kevin Frazier is an AI Innovation and Law Fellow at Texas Law.