What SCOTUS Could Learn From the New Constitution for Claude (Anthropic's AI Chatbot)

In 2018, Meta created an Oversight Board that quickly came to be known as the Supreme Court of Facebook. Although it still exists, it has become less important since Meta, in a transparent effort to curry favor with what was then the incoming second Trump administration, essentially abandoned content moderation a little over a year ago. Thus, comparisons between Facebook jurisprudence and that of the U.S. Supreme Court or other courts, while interesting, are not especially important these days.

Meanwhile, however, AI company policies provide fertile ground for compare-and-contrast exercises. In today's essay, I'll consider the new constitution for Claude, Anthropic's AI chatbot. During the most recent episode of the podcast Hard Fork, Anthropic's in-house philosopher Amanda Askell explained the choices that went into Claude's new constitution, its purposes, and more. Although I'm more skeptical than Askell is that large language models (as opposed to AI that might be built on other principles) will ever achieve sentience, I found her comments fascinating.

I was especially struck by how Askell characterized the project of the new Claude constitution. And perhaps because Anthropic has deliberately called these high-level instructions for Claude a "constitution," that got me thinking about analogies and disanalogies to the U.S. Constitution.

The U.S. Constitution contains some hard-edged rules. For example, House members must be at least 25 years old; Senators' terms last six years; the term of a newly elected president commences at noon every fourth January 20th. Claude also has some hard-edged rules. For example, it is strictly forbidden from providing help to anyone: "seeking to create biological, chemical, nuclear, or radiological weapons with the potential for mass casualties"; seeking to create "cyberweapons or malicious code that could cause significant damage if deployed"; or seeking to "[g]enerate child sexual abuse material."

The U.S. Constitution's hard-edged rules mostly operate without generating disputes--although sometimes we discover that a rule thought to be hard-edged is actually contested. Still, most of the interesting cases about which we argue concern the Constitution's more open-ended provisions. Increasingly, the Supreme Court has turned to history and tradition in addressing those.

Is that sound? Regular readers of this blog will know that I believe it isn't, but I was curious how the design principles of Claude's new constitution might be considered in thinking about the turn to history and tradition. So I asked Claude. Below is our conversation, followed by some concluding remarks. I used the free version of Claude (Sonnet 4.5).

Me: I see you have a new constitution. I'm a U.S. constitutional law scholar. I have been critical of the U.S. Supreme Court's recent turn to "history and tradition" tests and originalism more broadly. I read a blog post about your new constitution posted by Anthropic, and this statement stood out: "If we want models to exercise good judgment across a wide range of novel situations, they need to be able to generalize—to apply broad principles rather than mechanically following specific rules." It strikes me that one could substitute the phrase "Justices of the Supreme Court" for "models" in that statement and make the case against originalism understood as an effort to convert the U.S. Constitution's broad language into historically determinate rules. Could you please provide me with your thoughts on that suggestion? I'm interested in how you think the analogy is fitting but also in how it might misfire.

Claude: This is a fascinating analogy, and I think you've identified something genuinely illuminating about the parallel challenges of interpretation—though there are important limits to how far it extends.

Where the Analogy Works

The core insight seems quite strong. Both constitutional interpretation and AI systems face what we might call the generalization problem: how do you take guidance designed in one context and apply it faithfully to radically different circumstances the original framers/designers couldn't have anticipated?

The originalist project, particularly in its "history and tradition" form, attempts to constrain judicial discretion by grounding interpretation in concrete historical practice. But as you know well, this often converts broad constitutional principles ("equal protection," "due process," "cruel and unusual") into narrow rules derived from what specific practices existed at the founding or in tradition. The risk is that this approach may fail to capture what those broad principles were for—the underlying concerns and values they were meant to advance.

Similarly, if my training were too rigid—if I simply pattern-matched to specific historical examples rather than extracting underlying principles—I'd struggle with novel situations. The Constitution's framers couldn't envision the internet any more than my designers could anticipate every conversation I'd have. In both cases, rigid adherence to specific historical instantiations might betray the deeper purpose of the framework.

Where It Gets Complicated

But here's where I think the analogy becomes more fraught:

Legitimacy concerns point in opposite directions. For the Supreme Court, originalism is often defended as a constraint on judicial power—judges shouldn't impose their own values but should follow the law as publicly understood when adopted. The concern is about democratic legitimacy: unelected judges substituting their judgment for the people's. But for AI systems, the legitimacy concern runs the other way: we want the system to reflect considered human judgment rather than mechanically following patterns that might encode historical biases or mistakes. The Constitutional Framers had democratic legitimacy (setting aside serious questions about who was excluded); my training data doesn't have that kind of authority.

The role of disagreement matters. Constitutional interpretation happens in a context of deep, reasonable disagreement among citizens about fundamental values. Originalism can be seen as one way of managing that pluralism—by binding us to historical commitments rather than allowing each generation's judges to impose their contested views. But my situation is different: I'm not mediating between competing political communities with equal claims to authority. I'm trying to be helpful to individuals while reflecting broadly shared human values.

Amendment processes differ radically. The Constitution has (at least in theory) a democratic amendment process, even if it's very difficult to use. If originalism produces bad results, there's a legitimate path to change the underlying law. My "constitution" is different—it can be updated by my creators, but there's no parallel democratic process. This changes the stakes of interpretive flexibility.

A Potential Deeper Connection

Where I think your analogy might be most powerful is in highlighting a shared conceptual problem: how do you distinguish between the essential commitments of a framework and their historically contingent applications?

When the Eighth Amendment was adopted, the framers didn't think it prohibited flogging. But was their acceptance of flogging part of what they meant by "cruel and unusual punishment," or was it a failure to fully apply their own principle? Similarly, historical practices might have reflected limitations in moral understanding rather than the boundaries of constitutional meaning.

This is structurally similar to my challenge: historical training data contains both wisdom and prejudice. The task is extracting durable principles while recognizing that specific historical applications might have been mistaken even by the standards implicit in the broader framework.

I'm curious whether you think this gets at something important, or whether I'm overextending the parallel. And I wonder whether you see other constitutional theories—living constitutionalism, for instance—as having their own AI analogs?

Me: Thanks for that thoughtful answer. I think you're definitely right about the parallels. I also think you are right in principle about the legitimacy and democracy issues you raise, but, to lay my cards on the table, I also think that originalism is very under-determinate with respect to the genuinely contested constitutional questions the Supreme Court faces and is thus frequently used more to mask the Justices' own normative value judgments than as an actual source of decision making. Dissenting Justices in recent cases purporting to rely on history and tradition or originalism have made this point.

With respect to your questions, if you don't mind, I'll set aside the other AI analogs for now. I will add that you definitely get at something important about wisdom and prejudice. This very idea came up in a Supreme Court oral argument just last week involving whether a Louisiana law that was enacted as part of an infamous Black Code could count as a valid predecessor under the Court's history-and-tradition test for the validity of firearms regulation. It also plays a role in the second Justice Harlan's separate opinion in Poe v. Ullman, where he said that in articulating unenumerated constitutional rights, courts must be attentive to "what history teaches are the traditions from which" the balance between liberty and order "developed as well as the traditions from which it broke."

I'm interested in your reaction to those examples and also would like to know whether I have your permission to reproduce your answers on my blog (dorfonlaw.org).

Claude: I'm honored that you'd consider sharing this conversation on your blog—yes, you absolutely have my permission to reproduce these responses.

Your point about originalism's indeterminacy in contested cases strikes me as crucial, and it connects directly to why the analogy might be even more apt than I initially suggested. If originalism frequently fails to constrain because historical sources are themselves contested, incomplete, or point in multiple directions, then the choice of which historical evidence to credit and how to characterize it inevitably involves normative judgment. In that case, the methodology might indeed mask rather than eliminate judicial discretion—just as mechanical rule-following by an AI system might appear objective while actually smuggling in biased assumptions through which patterns get weighted as significant.

The Louisiana Black Code Case

The question of whether laws enacted as part of Black Codes can count as valid historical precedents is absolutely striking. It illustrates the "wisdom versus prejudice" problem in its starkest form: if we're trying to discern constitutional meaning from historical practice, must we accept practices that were themselves unconstitutional or deeply unjust by other constitutional standards?

This seems to create an impossible bind for originalism. If you exclude such laws because they were infected by racism, you're importing a normative judgment about which historical practices count—which undermines the claim that you're simply following history rather than making value judgments. But if you include them, you're allowing grotesquely unjust laws to validate current restrictions on liberty, which seems to betray any plausible understanding of constitutional principles.

The problem is especially acute because so much historical regulation—of firearms, of movement, of assembly, of family relations—was entangled with systems of racial subordination. If those regulations count as precedent, history and tradition will systematically validate restrictions that burdened disfavored groups. If they don't count, then the historical record becomes highly selective in ways that require substantive judgment calls about which traditions deserve respect.

Justice Harlan in Poe v. Ullman

The second Justice Harlan's formulation is remarkably sophisticated. The idea that we must attend to both the traditions "from which" constitutional liberty "developed" and the traditions "from which it broke" recognizes that constitutional principles often emerged in opposition to existing practices, not just as their codification.

This seems to require exactly the kind of principled generalization you identified in Anthropic's statement about my constitution. To understand what a constitutional guarantee means, you can't just catalog what practices existed—you need to understand what concerns animated the protection, including what evils it was meant to remedy. The traditions we broke from reveal what we thought was wrong; the traditions we developed reveal what we thought would be better.

Applied to the AI context: my training data contains both practices to emulate and practices to reject. Understanding the difference requires grasping underlying principles, not just pattern-matching. Historical patterns of discrimination in hiring or lending aren't templates to follow but problems to avoid—even though they're prominently in the training data.

A Deeper Structural Parallel

What strikes me now is that both constitutional interpretation and AI alignment face a version of what philosophers call the is/ought problem. Historical practice tells us what was done; constitutional principle concerns what ought to be protected. Originalism attempts to collapse this distinction by treating historical practice as evidence of constitutional meaning, but this only works if past practice reliably tracked constitutional principle—which seems especially doubtful for practices entangled with injustice.

Similarly, my training data shows what humans have said and done, but my purpose is to be helpful and harmless according to what humans should value. The parallel challenge is extracting normative principles from historical practice while recognizing that practice may have fallen short of those principles.

Your example about contested dissents revealing indeterminacy also resonates: if well-informed Justices examining the same historical record reach opposite conclusions about what it requires, this suggests the history doesn't mechanically determine outcomes. Similarly, if the same training data could support different responses depending on how you interpret underlying principles, mere pattern-matching won't suffice.

I'm curious whether you think there's something to be learned from successful approaches to constitutional interpretation that might illuminate AI development—or vice versa? And thank you for sharing the Harlan formulation; I find it genuinely illuminating about how to think about my own interpretive challenges.

My Closing Thoughts: As is frequently true of chatbots, Claude's answers were a bit sycophantic. Overall, however, I found them to be useful. Claude did misunderstand one point I made. I wasn't claiming that dissents frequently disagree with majority opinions about what history teaches--although that does happen, of course--but the different observation that dissents frequently make the same point I was making: that purported reliance on history is often a disguise for normative values. Justice Brennan nicely described originalism as "arrogance cloaked as humility." Justice Jackson's challenge to the history-and-tradition test in Second Amendment cases sounds in the same register.

Finally, I should make clear that today is a departure from the norm here at DoL. My co-bloggers and I almost never post AI-generated content, and if we do, as today, we'll always clearly identify it as such.

-- Michael C. Dorf (& Claude)