AI Constitutions Need Administrative Lawyers - Yale Journal on Regulation

About 271 million people use Anthropic’s Claude each month. Observers expect this number to keep growing for the foreseeable future. Every one of these conversations is shaped, at least partly, by an 80-page document that Anthropic calls Claude’s Constitution. This document ranks the model’s priorities, sets a few absolute limits, and explains how much importance it should give to instructions from the company, businesses that build products using it, and people who use it. Anthropic describes it as “the final authority” on what Claude should be and how it should behave. The company also uses it to create and evaluate the data used to train Claude.

Anthropic is not the only company that relies on this kind of document to do such an important job. OpenAI maintains a Model Spec meant to serve a similar purpose. Microsoft AI has just released a draft Humanist AI Code of Conduct. The public can comment on this draft for six weeks, and it will shape the company’s models starting in 2027. Some of these documents may also influence widely used open-weight models through a process called distillation. In this process, a “student” model is trained using a “teacher” model. For example, Anthropic alleges that many Chinese open-weight models are trained on Claude and are therefore shaped, at least in part, by Claude’s Constitution.

The word “constitution” has understandably led scholars of constitutional law to study these documents. But if we set that label aside, the hardest questions they raise look less like the questions in Marbury and more like the questions at the heart of the Administrative Procedure Act (APA): Who gets to make binding rules? What steps must they follow? What explanations must they give? And what kind of review should be possible?

Administrative law is the field that has spent nearly a century asking how government bodies that are not elected can fairly use the power given to them to make important, technical rules. Scholars in this field are especially well placed to help answer some of these questions about AI.

Key Open Questions

First, consider how these rules are created. Joe Carlsmith, one of the authors of Claude’s Constitution, described the process in a talk at Yale Law School as “pretty informal.” According to him, there is no dedicated, formal process for deciding what goes into the document. He did not describe a specific group responsible for writing it, a published process for suggesting or approving changes, or a fixed schedule for reviewing it. Instead, these decisions are made through the company’s usual, mostly private processes: its normal management structure, a mix of formal and informal internal discussions, and, in the end, unnamed people who make the “final” decisions. Amanda Askell, the document’s main author, told Alan Rozenshtein and me on the Scaling Laws podcast that difficult decisions are sent to her and, if necessary, to more senior colleagues. To an administrative lawyer, this flexible process looks much like making rules without clear procedures: there is no public notice, no formal record, and no duty to respond to people outside the company. Carlsmith himself suggested that more formal processes may be needed as AI governance becomes more important. I would argue that the time for those processes has already arrived.

The next issue involves a basic distinction in the APA. Legislative rules generally require public notice and an opportunity for people to comment. Interpretive rules and general statements of policy are exempt from this process under Section 553 of the APA. Claude’s Constitution works somewhat like a law that creates an organization and sets its basic powers. It establishes broad priorities and then refers to “Anthropic’s guidelines”: more detailed instructions about topics such as medical advice, cybersecurity, and tool use. These guidelines must follow the constitution, but not all of them are public. This system has a basic document, more specific rules, and guidance that is mostly hidden from public view. It could benefit from the structures and procedures used in administrative law.

Another issue is how the constitution is interpreted. Since Loper Bright, and even before that decision, scholars and judges have spent much time debating who should settle unclear parts of a law. In AI constitutionalism, the current answer may worry people who recognize that interpretations of this document can affect billions of users in small but important ways. As Carlsmith explained, the working meaning of a word such as “ethical” is, in a sense, whatever Claude understands it to mean during this process. In other words, the model governed by the document also decides what the document means. Those interpretations then become part of training data that no one outside Anthropic formally reviews.

One possible solution is to give someone other than the model responsibility for deciding what the constitution means in difficult cases. Nathan Darmon and Tom Reed have proposed courts for AI constitutions along these lines. Anthropic may be considering a similar approach within the company: Askell has suggested creating a body of “case law.” This would consist of examples showing how Claude should reason through difficult situations. Either way, the meaning of the constitution would increasingly be decided one case at a time, rather than through clear changes to the text reviewed by people outside the company.

Administrative lawyers will recognize this as a form of making policy through decisions in individual cases. The Supreme Court approved this approach in Chenery II, and the National Labor Relations Board (NLRB) has relied on it for decades. But administrative lawyers also know its costs: people affected by the rules may get little warning about what those rules mean, and policy can change whenever decision-makers change how they interpret the rules, even in small ways.

Another set of questions concerns changes to AI constitutions. Anthropic says its constitution is still being developed. Askell has explained that if the company wants to change the model’s behavior, it needs to change the text and publish the updated version. This matters for both technical and procedural reasons. On the technical side, Anthropic treats the constitution as the final authority for training Claude and uses it to generate training data. So a lasting change in the model’s behavior should require a change to the document. On the procedural side, Anthropic has promised to publish the current version, so the public document should match the one used to train the model. This is a good commitment, but administrative law principles call for a more established and definite process.

Under the APA, creating or changing a legislative rule generally requires public notice and an opportunity to comment. The APA also prohibits arbitrary and unreasonable government action. As the Supreme Court explained in FCC v. Fox, an agency changing its position must acknowledge the change and explain why it is doing so. Applied to AI labs, these principles would require a company to identify which parts of its document changed, explain the reasons for each change, describe the alternatives it considered, and identify the users likely to be affected.

Finally, there is the question of public participation. As noted earlier, Microsoft AI recently released a draft of its Code of Conduct and invited the public to comment before using it to train its models. This may produce useful feedback from certain people who are deeply involved in AI, sometimes called “AI-pilled” stakeholders. But it is unlikely to represent the full range of people and interests affected by Copilot and Microsoft AI’s other models.

Simply allowing people to comment does not guarantee that the process is meaningful. Section 553 requires agencies to consider the comments they receive and explain the reasons and purpose behind the final rule. Courts have interpreted these requirements to mean that agencies must respond to important comments and cannot adopt a final rule that goes beyond what people could reasonably expect from the proposal. Decades of experience have also revealed weaknesses in this process. A study by Susan Webb Yackee and Jason Webb Yackee of government rulemaking, although dated, found that comments from businesses tend to have more influence on final rules than comments from other participants. The Administrative Conference of the United States (ACUS) has issued recommendations on how to handle large numbers of comments, comments created by computers, and comments falsely attributed to other people. These findings, along with other lessons from public consultation in government, should be shared with AI labs as they consider different ways to invite public input.

To be clear, I am not arguing that the APA applies to Anthropic or any other AI lab. These are private companies competing in a fast-changing market. Many things that make administrative procedures legitimate, including legal authority granted by statute, review by courts, and political accountability, have no direct equivalent here. Still, given how quickly AI labs are moving, they would benefit from learning from existing systems instead of starting from scratch. Public-law procedures have been adopted in private governance before. For example, FINRA must file proposed rule changes with the SEC, which then publishes them for public comment. Standards organizations accredited by the American National Standards Institute must follow procedures that ensure openness, balance, agreement, and a chance to appeal. The AI labs themselves have chosen to publish documents explaining their values, and some have invited public feedback. By adopting the language and some of the practices of public governance, they have made it reasonable to ask whether their processes live up to the standards that language suggests. 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 establish published a research agenda covering values, procedures and authority, technical compliance, and enforcement. The part about procedures, in particular, reads like a course outline for an advanced administrative law seminar. Here are a few areas where scholars in this field could make a difference right away:

• Procedures that match the importance of the decision. Which changes should require public comment, and which are only small clarifications? Administrative law’s distinctions among legislative rules, interpretive rules, and exceptions for good cause may give AI labs a useful model for deciding which choices need more formal procedures and public participation.
• Explaining decisions and keeping records. What would count as a good explanation for changing Claude’s Constitution or OpenAI’s Model Spec?
• Institutions for oversight. People involved in AI governance have already suggested creating self-regulatory organizations for advanced AI, similar to FINRA. Those proposing the idea include Google’s Demis Hassabis. Administrative lawyers who understand how this kind of regulatory system works could help lawmakers decide whether it is worth using in the AI field.
• The government’s role. Washington is already influencing the values built into AI models through its purchasing decisions. Examples include the “Preventing Woke AI” executive order and the Office of Management and Budget (OMB) guidance that puts it into practice. As Nicolas McMullan and I argued, requirements to disclose information and conditions attached to government purchases are more likely to survive review under the First Amendment than direct changes to the text of an AI constitution. Working out how to analyze the difficult legal questions that may arise as governments pay more attention to AI constitutions is another task well suited to the readers of this journal.
• Help develop and teach a course on AI Constitutionalism. This spring, I will teach a seminar at the University of Texas School of Law called “AI Constitutionalism.” As far as I know, it is the first course of its kind. Law students and public policy students will take it together, and students will not need any programming experience. I plan to make the course freely available to the public. I will publish the syllabus and reading materials soon. I would greatly appreciate your feedback. Tell me which legal ideas I have missed, where the comparison with administrative law does not work, and how you would teach material that overlaps closely with your own research and courses. I will make changes as the course develops and share the final version. More than anything, I hope you will soon teach your own version or include a few classes on the topic in your existing courses.

Helping address these questions does not require becoming an expert in machine learning. It requires skills that administrative lawyers already have: designing procedures that make it clear how delegated power is used, allow people to challenge decisions, and make it possible to correct mistakes. The central point is that private organizations are exercising enormous power over the public. This is a moment when everyone who has studied how to limit such power needs to get involved. I hope you will join in!

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