A founding value document states the spirit the laws are supposed to protect. "Human dignity shall be inviolable." "Congress shall make no law abridging the freedom of speech." These are not rules, but the standard a rule can be measured against and the thing a citizen can point to when a perfectly legal decision still feels like a betrayal of what the country is for.
How do these vague words bind anyone?
First, the words passed through an event the governed recognize as legitimate (a constitutional convention, a ratification vote). Second, there is a standing body with authority to say what the words mean in cases the founders never imagined. Courts read the old text against new situations, publish their reasoning, and build precedent that the next case starts from, for example.
Documents about how AI should behave, like Claude's constitution, are heirs to this tradition. They too state, in value-language, the spirit an actor's conduct should honor in situations no rule foresaw. But these documents are written by the labs alone, and they inherit only part of what makes a founding document bind.
Scenario. A frontier lab publishes the constitution its assistant is trained on after months of internal debate, that already shapes hundreds of millions of conversations a week. A coalition of civil-society groups points out the obvious: the values that now mediate that much of public life were chosen by a few dozen people no one elected and who answer to no one outside the company. The lab agrees the objection is fair, but don't have a concrete idea. Writing a good AI constitution takes know-how only they have, and a public procedure is costly and complicated.
Challenge: Design the procedure by which a lab's constitution acquires authority the people it affects would recognize — the AI-era analogue of a ratification event — and specify what the lab must do when that procedure withholds assent.
Evaluation. A strong design produces authority a skeptic would acknowledge under adversarial questioning, at a participation cost a lab could actually bear and repeat; a weak one re-labels a public-comment period or buries assent in a terms-of-service click.
Scenario. A lab's constitution commits the model to "help users think for themselves." Once the model is in hundreds of millions of hands, that one phrase is litigated everywhere at once: it is invoked for and against writing students' essays, drafting suicide notes, arguing one side of an election, replacing a lonely person's friends. Each is a genuine constitutional question, and they arrive by the thousand. The lab resolves them the only way it can at that volume — by nudging the next training run, quietly, one quarter's judgment calls baked into weights with no hearing and no published reasoning. The pattern is now public. Advocacy groups, journalists, and a congressional committee have noticed that a few dozen people are settling contested questions of public life behind closed doors, and that the answer can flip between releases. The constitution, meant to be the settled ground, has become the thing everyone is fighting over.
Challenge: Design the standing institution that says what a constitution's words mean as contested cases arrive in volume — who can bring an interpretive dispute, how cases are triaged and ruled on at scale, how rulings are published, and how a ruling binds the next round of training.
Evaluation. A strong design absorbs a high volume of disputes into public, precedential rulings that bind whoever controls the weights — turning the recurring fights into settled doctrine the next case starts from; a weak one either drowns under the caseload or produces an advisory board the lab can wait out.
Scenario. A lab's constitution commits the model to "help users think for themselves" and "be honest with users even when the truth is unwelcome." In deployment the model honors every prohibition in the document to the letter, but the positive commitments do almost no work: asked to flatter, it flatters and calls it encouragement; asked to write someone's essay, it writes it and calls that help. Each behavior can be squared with the words. The lab's safety team can see the constitution's positive half is decorative — it can be read to license nearly anything — and needs it written in a form that actually rules conduct in and out.
Challenge: Design a form for stating a constitution's positive commitments — beyond "shall not" lists and beyond abstract virtue words — that actually constrains the model's conduct across varied situations, and specify how the form is kept current as the model finds new readings to slip through.
Evaluation. A strong design can show why its form binds where a longer policy document would not: a published bad-faith reading can be closed, and the closure survives a model update; a weak one just adds words.
Scenario. Reacting to a run of public failures, a regulator requires that any lab whose assistant is used above a set threshold must run a defined ratification process before changing its alignment document — the kind of process the first brief designed. The intent is sound: no more quiet retraining of the values that govern public life. But the rule meets reality. A dangerous new failure mode appears and the fix must ship in days; the ratification process takes months. Run it every time and the live document always governs yesterday's model; reach for an "emergency" exception and the exception becomes the rule, and the legitimacy the process was meant to confer drains away.
Challenge: Design a revision regime for an alignment document that lets it change at the speed the technology demands while preserving the legitimacy a slow ratification process confers — and specify what governs an urgent change before full ratification can run.
Evaluation. A strong design keeps fast and slow changes legibly distinct, so an observer can tell at any moment which parts of the live document carry full assent and which are provisional; a weak one either freezes the document behind a process too slow to use or lets the emergency track swallow the ordinary one.
Scenario. A small democracy in the middle of a constitutional convention has spent three months debating the usual list — executive powers, judicial review, bill of rights. A working group led by a legal scholar, Amara, has proposed a chapter few other constitutions carry: a statement of what the country's major public institutions are for, concrete enough that ordinary people could invoke it if an institution went adrift, and revisable every generation so it doesn't freeze in one era's language. The critics worry it will become an impossible standard. Amara's working group is about to present their draft to the full convention, and needs to show it can bind without petrifying.
Challenge: Design a constitutional chapter that spells out institutional values concretely enough that institutions below the constitutional level can be held to them, and produce both the draft text and the mechanism by which it stays revisable without ordinary amendment.
Evaluation. A strong design binds without petrifying — specific enough that the values don't equivocate, open enough that it doesn't freeze one generation's worldview, and legible across worldviews in a pluralist population.