Stanford University


Gillian Hadfield
Normative Infrastructure for AI Safety and Alignment

Gillian Hadfield: Normative Infrastructure for AI Safety and Alignment
March 4, 2024
Hybrid event
Show buttons

On Monday, March 4, 2024, Gillian Hadfield of the University of Toronto joined us to discuss new research, “Normative Infrastructure for AI Safety and Alignment.” Couldn’t attend the seminar? Watch the video below.

Questions from the seminar

  1. What is the basis of competition between the private regulators?  What is the basis of assessing the effectiveness of the private regulators?
  • In the regulatory markets framework, private regulators are licensed by governments and they are monitored/audited/regulated by government under that license to ensure they are achieving the outcomes set by government—such as a set metric for maximum fatal accidents in autonomous vehicles or ensuring that automated decisions do not unreasonably discriminate against protected classes (using various techniques, like those used by equal employment offices or courts to assess what is ‘unreasoanble’ discrimination). Figuring out methods for such regulation of the private regulators is a key design task and challenge, and a limit on the use of this method: if governments can’t effectively oversee private regulators, then this model is inappropriate. As for competition: governments require the ultimate target of regulation (e.g. the AI developer) to purchase the services of a licensed regulator and the key point here is that governments license multiple private regulators to compete to provide these services. The goal is to generate competition among licensed private regulators for the profits associated with providing this service (for a price) to regulated target and to harness that competition to provide incentives for investment and innovation in better methods (and technologies) for achieving the government’s regulatory goals, as well as efficiency in the provision of regulatory services.
  1. What about potential disincentive to innovate and the dampening effect due to the cost to startups and small/medium enterprises to pay a licensed regulator to certify their AI solution? Big Tech will have no problem paying of course.
  • I think this is a question about the validity of the government’s outcome requirements.  If those are legitimate and valued properly, then there just are costs of making sure that our AI systems do not pose the risks those requirements aim at. We currently have laws that impose inefficient costs of compliance that hamper startups. So the goal has to be to get to the most cost-effective way of achieving appropriate limits and controls. The regulatory markets proposal is aiming at that. It doesn’t deal with the capacity of government to impose illegitimate or excessive requirements—that’s a different question. But the regulatory markets proposal takes those goals as a given and asks: how can we achieve them with great cost-efficiency and efficacy?​​
  1. Curious if a more consumer-driven “oversight” mechanims is another model needed in this ecosystem, such as Consumer Reports, the Better Business Bureau, etc.
  • I think that’s a question about the democratic “what” rather than the regulatory markets “how”–it’s the same question we face today, whether we should have a greater role for consumer-driven standards setting. 
  1. Are there regulations for special cases?  In general, there is a possibility that many people will not be able to understand complex and well-regulated systems.
  • I think the understanding that ordinary people have of both the ‘what’ and the ‘how’ of regulation is important. One of the implications of the theoretical approach I sketched, based on my work with Barry Weingast, is that the ultimate test of a legal order is not an abstract or external normative requirement, but rather whether most ordinary people will look at our systems and be willing to participate in and cooperate with them. I think that’s a constraint on what we build and how we regulate. If systems and their regulation are so complex that ordinary people don’t understand them—or trust oversight—well enough, then I don’t think those systems and regulations do their job.​
  1. It seems that there is still a problem on how AI/ML is defined or who has the ultimate authority to decide if something really is using “true” AI/ML? 2. Also how is “personhood” defined please?
  • I think we ultimately don’t want to define and regulate “AI/ML” per se, but rather how systems and uses impact us. If a bank is using AI/ML to discriminate, it really doesn’t matter how AI/ML is defined: its just wrong to be discriminating. The question is whether the use of AI/ML makes it harder, or easier, to reduce discrimination. And if a powerful system exists that can transform how we do everything, as a general purpose technology, in ways that change how societies and economies—and their normative social orders—work, then again it doesn’t matter what our definition of AI/ML is.  I think we should focus on effects and impacts, not trying to define AI/ML. As for personhood: the idea here is akin to the idea of legal personhood that we accord corporations:  corporations are formally recognized as “legal persons” in the sense that they can own property, sue and be sued in court, made the subject of regulation. It’s not a philosophical definition of personhood; it’s a purely legal and instrumental definition.​


The great majority of AI safety and alignment efforts focus on identifying specific human values, preferences, or policies and finding ways either to embed those via AI training or finetuning or to impose them as standards on deployed systems. But all of these approaches are likely to be quite brittle and of short-lived success: human normative systems are complex, highly variable, and dynamic. In this talk I’ll present several ideas about how to build the normative infrastructure necessary for more robust AI alignment. These ideas include building the legal infrastructure need for agile governance, such as registration for frontier models, legal “personhood” for AI agents, and regulatory markets to recruit private sector innovation of regulatory technologies, and the technical infrastructure necessary to train AI systems to read and participate in our dynamic normative environments.

About Gillian Hadfield

Gillian Hadfield

Gillian Hadfield is the inaugural Schwartz Reisman Chair in Technology and Society, Professor of Law, Professor of Strategic Management at the University of Toronto, and holds a CIFAR AI Chair at the Vector Institute for Artificial Intelligence.  She is a Schmidt Sciences AI2050 Senior Fellow. She was the inaugural Director of the Schwartz Reisman Institute for Technology and Society from 2019 through 2023. Her research is focused on the study of human and machine normative systems; safety and governance for artificial intelligence (AI); and innovative design for legal and dispute resolution systems in advanced and developing market economies.  She has also long studied the markets for law, lawyers, and dispute resolution; and contract law and theory. She teaches Contracts and Governance of AI.

Prior to rejoining the University of Toronto in 2018, Professor Hadfield was the Richard L. and Antoinette Schamoi Kirtland Professor of Law and Professor of Economics at the University of Southern California from 2001 to 2018.  She began teaching at the University of California Berkeley and was previously on the University of Toronto Faculty of Law from 1995-2000. Her book Rules for a Flat World: Why Humans Invented Law and How to Reinvent It for a Complex Global Economy was published by Oxford University Press in 2017 and in paperback with an updated prologue focused on AI in 2020.

Professor Hadfield served as clerk to Chief Judge Patricia Wald on the U.S. Court of Appeals, D.C. Circuit.  She was the Daniel R. Fischel and Sylvia M. Neil Distinguished Visiting Professor of Law at the University of Chicago (Fall 2016), the Eli Goldston Visiting Professor (Spring 2012) and the Sidley Austin Visiting Professor (Winter 2010) at Harvard Law School, and the Justin W. D’Atri Visiting Professor of Law, Business and Society at Columbia Law School (Fall 2008.) She was a 2022 Fellow at the Rockefeller Foundation’s Bellagio Center in 2022, 2006-07, and 2010-11 fellow of the Center for Advanced Study in the Behavioral Sciences, and a National Fellow at the Hoover Institution in 1993. She also has held Olin Fellowships at Columbia Law School, Cornell Law School, and USC. She is past president of the Society for Institutional and Organizational Economics and the Canadian Law and Economics Association, a former director of the American Law and Economics Association and the Society for Institutional and Organizational Economics, and a member of the American Law Institute.  She has served on the editorial boards for the Annual Review of Law and Social Science, Law and Social Inquiry and the University of Toronto Law Journal and is a founding trustee of the Cooperative AI Foundation.

Professor Hadfield was a Senior Policy Advisor for OpenAI in San Francisco from 2018 to 2023, and is an advisor to courts, governments and several organizations and technology companies engaged in innovating new ways to make law and policy smarter, more accessible, and more responsive to technology and artificial intelligence, including the Hague Institute for Innovation of Law, LegalZoom, and Responsive Law. She is a member of the Canadian AI Advisory Council. She was a member of the World Economic Forum’s Future Council for Agile Governance and co-curated their Transformation Map for Justice and Legal Infrastructure; she previously served on the Forum’s Future Council for Technology, Values and Policy and Global Agenda Council for Justice; and was a member of the American Bar Association’s Commission on the Future of Legal Education, 

Stanford University