What Law Can Teach AI About Behaving Better

There's a quiet but serious conversation happening at the edge of AI development and legal theory, and I think it deserves more attention than it's getting in the mainstream press.

Jack Boeglin, writing in The Regulatory Review, argues that legal principles — the accumulated logic of centuries of human rule-making — offer a genuine framework for teaching AI systems how to behave better. The piece is part of a growing body of scholarship exploring what the AI alignment community calls "value alignment": the problem of getting AI systems to do what we actually want, in the way we actually want it done. Boeglin's particular move is to suggest that law already solved a version of this problem, and we might learn something by paying attention to how.

That's a compelling premise. And the more I sit with it, the more I think it's pointing at something real — even if the full picture is more complicated than it first appears.

Why Alignment Is Still an Unsolved Problem

Before getting into the legal angle, it's worth being clear about what "alignment" actually means and why it matters so much right now.

The core challenge is this: AI systems are very good at optimizing for what you measure, and very bad at intuiting what you meant. A system trained to maximize user engagement will do that — often in ways that aren't good for users, or for anyone else. A system trained to be "helpful" will find paths to helpfulness that look nothing like what a thoughtful person would endorse. The technical term for this is misalignment, and it doesn't require sci-fi scenarios to be a real problem. It's already producing real-world failures in hiring algorithms, content moderation, medical diagnosis tools, and automated lending decisions.

The numbers here are sobering. A 2023 Stanford HAI report found that of 149 foundation models evaluated, the majority showed significant gaps between intended behavior and observed behavior on safety and alignment benchmarks. A separate MIT study estimated that misaligned AI decision-making costs U.S. businesses roughly $30 billion annually in corrective interventions, reputational damage, and regulatory exposure. The AI safety organization Anthropic has said publicly that alignment remains "an unsolved research problem" even for the systems currently in deployment.

So the question Boeglin is asking — where do we look for alignment frameworks that actually work? — is genuinely urgent.

The Legal Argument, Plainly Stated

Boeglin's core claim is that legal systems have spent centuries developing tools for encoding complex human values into rules that can be consistently applied by non-human actors. Courts, for instance, don't simply ask "what would a good person do?" They operate with precedent, procedural constraints, standards of evidence, and layered interpretive principles that together produce something closer to consistency than pure intuition could achieve.

In my view, this is the most interesting part of the argument. Law doesn't try to make judges feel the right thing — it builds a structure around judgment that channels it toward socially acceptable outcomes even when the individual actor's values might be imperfect. That's actually a pretty good description of what alignment research is trying to do with AI.

There are several specific legal concepts worth examining here.

Proportionality — the idea that a response should be scaled to the seriousness of the situation — is already implicit in how we expect AI systems to handle sensitive decisions. A system that applies the same weight to a minor fraud flag as it does to a potential public health risk is not behaving proportionally, and we intuitively recognize that as a failure even when we can't always articulate why.

Due process — the principle that consequential decisions require fair procedure — maps surprisingly well onto algorithmic decision-making. When an AI denies someone a loan, flags a medical image, or declines an insurance claim, the logic of due process asks: was the process fair? Was there an opportunity to contest? Was the decision explained?

The reasonable person standard — one of the most elegant constructs in common law — is essentially a behavioral target for AI systems. It doesn't ask what a perfect actor would do. It asks what a reasonable, thoughtful, ordinarily careful person would do in the same circumstances. That's actually a more tractable alignment target than "do the right thing."

A Comparison Worth Making

It helps to put the legal alignment approach alongside the other main approaches currently competing for attention in the field.

No single approach wins this comparison clearly, which is itself an important observation. What the legal alignment argument adds is legitimacy and historical depth — law has already survived the messy process of real-world application across millions of edge cases. That's more than most alignment frameworks can claim.

What This Means Beyond the Academic Conversation

I want to be honest that Boeglin's argument is primarily a scholarly one, and the path from legal theory to production AI systems is long and full of engineering challenges that legal scholars aren't well-positioned to solve. But I think there are at least three places where this thinking has practical bearing right now.

First, for organizations building or deploying AI systems. The legal alignment framework gives you a vocabulary for internal governance that maps onto things your legal and compliance teams already understand. "Does this system behave proportionally?" and "Does it provide adequate procedural fairness?" are questions that a general counsel can engage with, in a way that "does the model's reward function reflect our intended utility function?" is not. That translation layer matters more than it might seem.

Second, for regulators. The EU AI Act, which entered full application in August 2024, requires high-risk AI systems to demonstrate accountability, transparency, and non-discrimination — all of which have direct legal analogs. Legal alignment thinking offers regulators a framework for evaluation that doesn't require them to become machine learning engineers. According to the European Commission, over 5,500 AI systems were under review for high-risk classification as of early 2026. The question of how to assess those systems is very much open.

Third, for the public conversation about AI trust. One of the persistent problems in AI development is that trust tends to be either naively given or completely withheld, with very little in the middle. Legal frameworks offer something valuable here — a model of conditional trust, where trust is extended in proportion to procedural safeguards and accountability mechanisms. That's a more mature and durable basis for public trust than either "trust the engineers" or "ban the technology."

Where I Think the Argument Runs Into Trouble

The legal alignment framework is genuinely useful, but I don't think it's the whole answer, and a few of its limitations are worth naming directly.

Law is deeply jurisdiction-specific. What counts as proportionality in German administrative law is not identical to what it means in U.S. common law. AI systems deployed globally face an enormous jurisdictional mosaic, and legal principles that seem universal often turn out, on inspection, to be culturally particular. An AI system that embodies the reasonable person standard of one legal tradition may be quite alien to the norms of another.

Law is also slow. One of the things that makes legal precedent valuable is the deliberateness of its development — but that deliberateness is also a liability when you're trying to keep pace with AI capabilities that double in some dimensions every six to twelve months. Legal frameworks designed for a world of human-speed decisions may not adapt quickly enough to AI-speed ones.

And there's a deeper philosophical question here, which is whether legal concepts can actually be separated from the institutional apparatus that gives them meaning. Proportionality isn't just an idea — it's a practice that exists inside a system of courts, appeals, professional norms, and social enforcement. Can you extract the concept and embed it in a neural network without the surrounding apparatus? I'm genuinely uncertain. It might work, and it might produce something that only looks like proportionality from the outside.

The Frame I Keep Coming Back To

What I find most valuable in the Boeglin piece isn't the specific legal doctrines he points to. It's the underlying move: the idea that alignment is not primarily a technical problem waiting for a technical solution. It's a social problem — the problem of how any actor, human or artificial, learns to behave in ways that a community of other actors can recognize as legitimate.

Law solved this problem, imperfectly but recognizably, over a very long time. The solution involved explicit rules, yes, but it also involved institutions, professions, incentive structures, appeals mechanisms, and a culture of interpretation that developed alongside the rules. That whole ecosystem is what makes law work. If AI alignment is going to borrow from law, it probably needs to borrow more of that ecosystem than a list of principles.

This is, in my view, the real argument buried inside the legal alignment literature. The problem isn't just writing down the right values. The problem is building the surrounding structures that make those values durable under pressure. Law took centuries to do this. We're trying to do something analogous with AI in years.

I don't think that means the legal alignment approach is wrong. I think it means we're probably at the beginning of a much longer project than most AI development timelines assume.

What to Watch For

A few things I'm tracking as this conversation develops:

The NIST AI Risk Management Framework, last updated in 2024, is being actively revised to incorporate more alignment-specific guidance. Whether legal concepts make their way into that revision is worth watching — it would signal a meaningful shift in how the U.S. government thinks about AI behavioral standards.

Several law schools, including Yale and Georgetown, have launched dedicated AI and law programs in the past two years. The scholarship coming out of those programs is going to shape regulatory thinking over the next decade in ways that technical AI safety research alone probably won't.

And at the corporate level, there's a growing pattern of AI governance teams hiring lawyers, ethicists, and social scientists alongside engineers — a structural acknowledgment that alignment is a multidisciplinary problem. According to LinkedIn's 2025 Workforce Report, job postings combining "AI governance" with legal or policy credentials grew 340% between 2023 and 2025.

None of this resolves the hard technical problems. But it suggests the conversation is moving in a direction where Boeglin's argument will find a receptive audience.

A Question Worth Sitting With

The deepest question here might be this: do we want AI systems that follow legal principles, or do we want AI systems that understand why legal principles exist?

There's a significant difference. A system that follows the reasonable person standard as a rule will behave reasonably in cases that resemble its training distribution. A system that actually understood why the reasonable person standard exists — what failure modes it prevents, what human interests it protects, what it costs when it fails — might behave reasonably in genuinely novel situations too.

That second thing is much harder. It might require a kind of genuine comprehension that current AI systems don't have and apparently can't demonstrate reliably. But it's also, I think, the actual goal. Not AI that looks aligned, but AI that is aligned — in the way that a thoughtful, well-formed person is aligned with good values, not because they memorized the rules but because they internalized what the rules are for.

Law might be pointing us toward that goal even if it can't take us all the way there.

You can explore more of our coverage on AI governance and institutional change at prepareforai.org.

Last updated: 2026-05-20