Does drafting an AI model's constitution substitute for this institution's jurisdiction?
Published "AI constitution" proposals set values and behavior inside a model. Authorizing human institutions to govern AI is a separate question, and answering one leaves the other open.
SIRDA editorial hypothesis
A document that sets a model's values and behavior can be evaluated on its own terms: who drafted it, what process reviewed it, and how its authors can be held to account. That evaluation is separate from the question this inquiry asks, which is whether any human institution has authority to govern advanced AI at all. This hypothesis fails if the two questions are treated as interchangeable, so that a well-drafted model constitution is offered as a substitute for institutional jurisdiction, or an institutional proposal is judged solely by borrowing the vocabulary of model alignment.
0 objections ยท View challenges
Challenges to the editorial hypothesis
No published objections yet.
The argument
Recent proposals use "constitution" for two different things. One is a specification written into a model, describing the values and behavior it should exhibit; the drafter controls its content the way a company sets internal policy. The other is the founding document of an institution meant to hold power over people and organizations, which requires the accountability structures this inquiry investigates: authorization, review, capture resistance, rights, and an exit path.
Who Writes the AI Constitution? examines the first kind of document. Its authors, Nicolas McMullan and Kevin Frazier, argue that whoever controls a model's constitution holds a form of expressive power over what the model can say, and that regulating that power raises First Amendment concerns. That argument is the authors' own position, not settled law, and it assumes a drafter operating inside a legal system where the First Amendment could apply; a transnational body, a foreign developer, or an open-weight release would not automatically face the same analysis. A Thousand AI Constitutions, by Simon Goldstein and Peter Salib, takes a different position: rather than one uniform constitution, it argues for deliberately diversified values across models, and treats that diversification as a governance choice in its own right.
Both works are about who writes the values inside a system. Neither substitutes for an answer to this inquiry's questions: who authorizes a body to compel compliance from developers that never signed on, what limits that body during an emergency, and who reviews it if it expands its own mandate. A frontier developer that publishes a well-reasoned model constitution has not thereby answered whether any institution may compel it, another developer, or a state to follow a rule neither chose. Treating the two as the same claim would let institutional legitimacy be argued for by pointing at model-level documentation instead of independent authorization and review.
How to test it
Take a real published model constitution and ask two separate questions about it: does the document's own process show accountable drafting and review, and separately, does the document's existence license any institution to compel a change to it, or to a competitor's, without that institution's own accountable basis for jurisdiction. Track cases where advocates answer the second question by pointing only to the first.
Sources
Who Writes the AI Constitution?
Examines who currently drafts and controls the documents that set a model's values and behavior, and argues that this activity raises free-expression questions; the authors' First Amendment argument is their own view, not a settled legal conclusion, and it presumes a domestic drafter subject to US law.
Argues against a single uniform AI constitution and for deliberate diversification of the values embedded across models; presented through the Institute for Law & AI but authored independently of it.
Related Ecosystem
Required expertise
Kinds of expertise this question draws on, and active Ecosystem organizations already tagged with a matching area. A tag match is not a ranking.
- AI law
- Institute for Law & AI
- International institutional design
- Institute for Law & AI
- International law & jurisdiction
- Institute for Law & AI
- Legal alignment
- Institute for Law & AI
Related questions
Related failure conditions
Contributions
No published contributions yet.