A renowned expert in technology, law, and institutional economics, Gillian Hadfield joined the Johns Hopkins University this year as the Bloomberg Distinguished Professor of AI Alignment and Governance. Now, she gives us an inside look at her current research and most significant findings in the field of AI governance.
What are the most important problems in your field today?
Getting AI governance right is crucial because we’re about to integrate these systems into our economies and societies—the ultimate test of large-scale human cooperation. And here’s the thing: Humans have actually solved this problem incredibly well by building sophisticated normative systems—complex webs of rules, norms, and institutions that allow millions of strangers to cooperate and exchange without everything falling apart.
But in current AI research, people are ignoring these solutions that have been essential to human flourishing. Companies will say they consult the public because they crowdsource input, calling this “democratic.” But that’s not democratic—they’re still answering based off of values that the company instructs them on. Take for example Anthropic’s Claude model; its “constitution,” the values that it holds important, are largely kept secret with only a few provisions made public. We’re trying to understand and build up AI governance on a foundation of quicksand. We lack a fundamental understanding of how human normative systems actually work—how people create and maintain the very systems we need AI to integrate into.
The current approaches to AI alignment put value encoding entirely in the hands of companies. They are either programmed directly into systems or trained by contractors who select answers based on company values. But this is profoundly undemocratic and poorly grounded. It’s not how society actually works. Values are decided by companies with no input from the public, with no transparency about what they’re choosing to value.
The second major problem is that we’re not ready for the AI transformation that is coming. We’re not looking to the future and asking what governance structures, what legal frameworks, what institutions we need to establish and build up. We’re heading toward a world with millions of AI agents participating in the economy with essentially zero legal infrastructure to manage their participation.
Which of these problems are you tackling through your research and why?
My research tackles the core technical problem: We don’t have the right ideas about alignment. I’m working with machine learning researchers in my lab to build computational models of how human normative systems function. It’s more than just following societal rules; it requires a deeper understanding of what value those rules have in society.
But here’s why this approach matters: I became convinced that the entire alignment conversation was fundamentally misguided. Computer scientists were trying to solve alignment without really understanding what they were trying to align to. Human normative systems aren’t just collections of preferences—they’re evolved solutions to cooperation problems. If we want AI to integrate into human society, we need to start there.
My policy work flows directly from this technical insight. Once you understand that human norms are complex systems, you realize that traditional top-down regulation won’t work for AI. We need governance approaches that can evolve and adapt, which is why I’m working to push ideas like regulatory markets—ways to harness competitive dynamics to improve regulatory performance while maintaining democratic oversight of acceptable risk levels.
Does your current research seek to answer a fundamental question in science, or does it have potential practical ramifications?
I see my current work as both fundamental and practical. The fundamental question my research is investigating is how large-scale human cooperation works: How do societies create stable institutions that allow millions of strangers to engage in complex exchanges without everything falling apart?
This fundamental question has immediate practical implications. The challenges of AI governance are forcing us to think more rigorously about how human governance works and reexamine our own societies—particularly how we build governance approaches that can evolve and adapt.
How is your approach different than what others in your field are doing?
I spent decades studying how human societies solve coordination problems at scale, and I realized that’s exactly what the AI alignment challenge is—but the field has been approaching it without that governance knowledge. I see myself as bridging the gap and merging the technical aspects of AI development with human sciences, bringing insights about how societies solve coordination problems to the challenge of building AI governance structures.
What has been your most surprising finding so far?
My most surprising finding has been the work around “silly rules”—the rules that have no material impact on society, like wearing black to funerals or men wearing neckties. Societies that have more of these seemingly pointless conventions seem to do better; they’re not just things that hang around for no reason. When we think about the rules that AI systems have to follow, they may need to follow the silly rules, too.
Your most significant finding?
My most significant finding is about normative competence—whether AI can be governed and whether we can build AI systems that integrate into our democratic systems. Our existing approaches to aligning AI systems, grounded in preferences, are not only fragile, but illegitimate.
Here’s the thing about governance: Normative social order is a dynamic equilibrium that requires rules that are general and therefore incomplete. They can’t specify in precise detail what any agent—human or artificial—is allowed to do in every situation.
There are technical paths for building AI systems that can more robustly integrate into human normative systems rather than just following programmed preferences. Systems can be designed to be socially aware and ask questions about how norms apply and what must be done to stay compliant. They can also use data on human moral flexibility as part of their training.
These are steps in the right direction, but for AI systems to be truly normatively competent, normative institutions need to be legible to them and accessible at the speed and scale at which they operate. They’ll need to be incentivized to respect our norms and laws, and perhaps even participate in enforcing them. Normative competence will be necessary for general-purpose AI systems to be steered in the directions chosen by a human community—the directions encoded in our dynamic normative social order.