Artificial Intelligence and Legal Alignment

In a conversation with The Regulatory Review, Jack Boeglin discusses the concept of “legal alignment” in artificial intelligence (AI) governance, the challenges of aligning AI systems with diverse legal regimes, and the economic and regulatory forces that shape where AI systems are deployed.

As AI systems become increasingly embedded in everyday life, scholars and policymakers have debated how best to ensure that these systems act consistently with human values. Although many approaches to AI alignment focus on technical solutions, Boeglin argues that legal norms and concepts offer a complementary—and often overlooked—framework for guiding AI behavior. At the same time, jurisdictions around the world have adopted divergent approaches to regulating AI, raising questions about whether developers will tailor their systems to local legal norms or decline to deploy them altogether. Boeglin concludes that the risks of misalignment could “truly be monumental,” and he urges the legal field to “rise to meet this moment” in governing the technology of the future.

Jack Boeglin is a Sharswood Fellow at the University of Pennsylvania Carey Law School, where he teaches classes on federal courts and the law of AI.

Boeglin previously worked as an appellate litigator and tech regulation practitioner at an international law firm in London and Washington, D.C. He has served as a law clerk for four judges: Justices Ruth Bader Ginsburg and Sonia Sotomayor of the U.S. Supreme Court; Chief Judge Srikanth Srinivasan of the U.S. Court of Appeals for the D.C. Circuit; and Judge Guido Calabresi of the U.S. Court of Appeals for the Second Circuit.

The Regulatory Review is pleased to share the following interview with Jack Boeglin.

The Regulatory Review: You have written about the “alignment problem” in artificial intelligence (AI) governance. What is it?

Boeglin: The “alignment problem” is the challenge of teaching AI systems to identify human objectives and carry them out consistently with human values. Put differently, how do we teach a computer program to understand what we want and how we want it to be done? To answer those questions, we will necessarily need to draw on a variety of disciplines—computer science, linguistics, ethics, and—I argue—law.

TRR: You have written that “legal alignment” could create “more legitimate and effective approaches to AI alignment.” What is legal alignment, and how could it solve the alignment problem?

Boeglin: “Legal alignment” is a strategy for solving the AI alignment problem that foregrounds norms and concepts drawn from the law. As every first-year law student knows, the law is intimately familiar with what can go wrong when trying to get someone else to carry out your instructions. Although AI is an emerging technology, the basic structure of the alignment problem is not new. Courts and legal scholars have long contemplated alignment-type problems in governing relationships among humans, as well as human relationships with prior forms of technology. Legal alignment seeks to leverage the lessons learned from those prior run-ins with “misalignment” to help address the AI alignment problem.

TRR: When envisioning legal alignment as a solution to AI governance, how do you account for differences in legal systems and regulatory approaches around the world?

Boeglin: One of the strengths of legal alignment is that it naturally incorporates the diversity of value systems around the world. Many competing alignment strategies would apply uniformly wherever AI is deployed, even though societal values and expectations might differ dramatically from place to place. Of course, there are legal systems that embrace values that some AI developers would be loath to endorse. Where that is true, a developer might reasonably choose not to align its AI systems to that jurisdiction’s laws, in the same way that some technology companies already refuse to offer their products in particular countries for fear of how they might be misused.

TRR: Could similar decisions also be driven by economic considerations—particularly in highly regulated environments such as the European Union—and what role does profit play in shaping whether companies deploy or align AI systems in particular markets?

Boeglin: The EU has indeed been very active in regulating “big tech,” and AI in particular. But the EU’s laws are primarily focused on setting out rules of conduct that AI developers must follow when developing AI systems. I view those laws as largely distinct from the project of legal alignment, which focuses on teaching AI systems to conform their conduct to the legal rules and norms that apply to ordinary people. But I agree with you that the question of whether the EU may eventually go too far in its pro-regulatory approach and cause U.S. technology companies to withhold their products from the European market is an important one. To date, big tech has been reticent to deny itself access to that market. Even where compliance is onerous, it has typically been deemed worth it given the massive financial benefit of offering products to the large and wealthy population of the EU. At some point, however, it is possible that the financial math could change.

TRR: Which stakeholders must be involved in addressing the alignment problem in AI governance?

Boeglin: Another comparative advantage of legal alignment is that it can piggyback on the stakeholder involvement that already takes place when formulating legal rules and norms. To be sure, law is far from perfect in this respect; it too often fails to be truly representative and to properly take account of all affected stakeholders. But it is superior to what is happening now, in which alignment strategies mostly reflect the personal views of workers at leading AI companies, rather than the result of a consultative stakeholder process.

TRR: What are the consequences if AI systems are not effectively aligned with legal norms and societal values?

Boeglin: No one can say for sure. So much will depend on how deeply these systems become embedded in the basic infrastructure of our governments, our workplaces, and our daily lives.  The trendlines, however, certainly suggest that AI is on its way to being integrated into every aspect of modern life. And if that happens, the risks of misalignment could truly be monumental. Already, we have seen what can happen when these systems are improperly aligned—as the recent spate of tort suits against leading AI labs, including over AI’s alleged role in contributing to teen suicide, attests.

TRR: What lessons have you drawn from your own legal career that inform how you think about legal alignment in AI governance?

Boeglin: When I was a law student a decade ago, I had the pleasure, and perhaps foresight, to take a seminar on “Artificial Intelligence, Robotics, and the Law” with Margot Kaminski and Jack Balkin. Back then, so much of what we talked about—digital companions, autonomous war machines, the end of white-collar work—seemed like science fiction. But what struck me as speculative then doesn’t seem so far-fetched now. Given just how quickly the technological landscape continues to change, it is essential that lawyers rise to meet this moment. Fortunately, our profession has centuries of experience in dealing with alignment problems. Our ability to draw on old solutions to help solve new problems could prove essential to governing the technology of the future.