Interviews

Stavros Gadinis, Berkeley Law Professor and Author of Corporate Ordering – Interview Series

Antoine Tardif
Add Unite.AI to your preferred sources on Google
Stavros Gadinis, George R. Johnson Professor of Law at UC Berkeley School of Law
Stavros Gadinis, George R. Johnson Professor of Law at UC Berkeley School of Law and author of Corporate Ordering.
Key PointsCorporate ordering describes how firms write rules and monitor themselves before…
View more
  • Corporate ordering describes how firms write rules and monitor themselves before regulators act.
  • AI safety evaluations lack independence when management hires and can fire its own outside reviewers.
  • A safety framework that never delayed a launch is indistinguishable from public relations.

Stavros Gadinis is the George R. Johnson Professor of Law at UC Berkeley School of Law. His research examines corporate governance, financial regulation, and the internal systems companies use to make and monitor decisions. Before entering academia, he practiced corporate law in Europe.

His new book, Corporate Ordering: How Corporations Navigate Social Conflict, published by Cambridge University Press, examines how businesses develop governance systems when public rules are unsettled. Through cases involving ridesharing, climate sustainability, and artificial intelligence, it considers how companies make rules, exercise authority, and oversee their own conduct.

Your career spans corporate legal practice in Europe and research on corporate governance and financial regulation at Berkeley Law. How did that background shape your thinking, and what inspired you to write Corporate Ordering?

I began my career practicing corporate law in Europe, and then spent many years at Berkeley studying corporate governance and financial regulation, including how rules spread across borders when no single government is in charge. We tend to think of the business world and the regulatory world as opposites, one eager for quick decisions and large profits, the other committed to bureaucratic process and the public good. Practice and research both taught me that the line is much blurrier than that.

The book grew out of conversations with general counsel and senior executives about how they handle controversial decisions, the ones where employees, customers, communities and governments all want different things. What struck me is how closely their methods resemble those of a good regulator. They gather information by reaching out to the people affected, they explain their choices publicly, and they build mechanisms to check whether those choices hold up. These are not regulatory habits or corporate habits. They are the principles of sound decisionmaking in any institution that wields power over others. Corporate Ordering is my attempt to describe that machinery, show where it works, and show where it fails.

What is “corporate ordering,” and how does it help explain the way AI companies set rules for technologies that regulation has yet to catch up with?

Corporate ordering is the internal governance infrastructure companies build to make contested decisions before the law tells them what to do. It has three functions, which mirror the branches of a state. Rulemaking is how a company writes its own principles. Executive action is how it applies them to concrete decisions. Monitoring is how it checks whether it followed through, and how it finds out when it did not.

AI companies show all three at work. The frontier labs write constitutions, model specifications and safety frameworks, which is rulemaking. They decide whether a model is ready to launch, what it will refuse and who gets access, which is executive action. And they commission red-teaming, outside evaluations and incident reporting, which is monitoring. In the book I chart these functions across the major labs. The differences are real, but so is a striking convergence, as each firm discovers why it needs every piece.

I want to be clear about one thing, because it is often misread. Corporate ordering is not an argument against regulation. It describes the material regulation will be built from. When legislators act, they borrow concepts the companies developed first. The real question is whether that first draft was written carefully enough to deserve borrowing.

Your book examines ridesharing, climate sustainability, and AI governance. Which lessons from the first two areas are most useful for AI, and where does AI present fundamentally different challenges?

The lesson from ridesharing and climate is that corporations very often write the first draft of the law. Uber and Lyft developed insurance and safety reporting practices before states had rules for them, and much of corporate carbon accounting runs on the GHG Protocol, a standard business helped build. Regulators then condoned those practices, supplemented them or wrote them into law. Regulation rarely starts from a blank slate. And the more thoughtful companies are about their choices, the more they take affected parties into account, the more likely those choices are to survive into regulation.

AI differs in two ways. The first is speed. Regulators are often a step behind industry, but in AI the gap is far wider. Regulators will have to look to companies to develop standards and internal processes, and concentrate their own efforts on oversight, verification and enforcement. That makes their role different, not smaller. Verification is where public power matters most, because it is where the conflict of interest with management is strongest. The second difference is reach. AI touches nearly every part of daily life, and it has become a major investment for the nation as a whole. The systemic risk implications are as hard to miss as they are to fathom.

When an AI company selects and pays its own outside safety evaluators, what safeguards make that oversight meaningfully independent?

This conflict of interest is familiar from financial auditing, and corporate law already has tools to manage it. The first is who hires and fires. The relationship with the evaluator should sit with independent directors who have no other financial ties to the company, much as the audit committee, not the CEO or CFO, oversees the external auditor. Management should not be able to dismiss a reviewer whose findings are inconvenient.

The second is access and publication. The evaluator needs to see what it asks to see, and it needs the right to publish its findings, or at least a meaningful summary, without a company veto. Reputational risk is what gives a review its teeth. The third is common review standards, so that an evaluation means the same thing across firms and evaluators cannot compete by being lenient. The fourth is protection for whistleblowers, and even rewards when they bring actionable information, because people inside the company are often the first to know when an evaluation was steered.

Suppose an outside evaluator identifies a serious safety concern just before a major model launch. Who should receive that finding, and who should have the authority to delay or stop deployment?

The finding should go to the board directly, ideally to a committee of independent directors responsible for safety, and not only through the management team that is pushing for the launch. Management can and should respond, but it should not be the filter. The authority to delay or stop deployment should rest with the board, because this is exactly the kind of decision boards exist to make, one with large stakes, conflicting interests and no obvious right answer.

But the more important point is that the decision must be explicit and documented, and traceable to the people who made it. If the board goes ahead despite the concern, the record should show what it knew, what it weighed and why it judged the risk acceptable. That paper trail creates the conditions for liability, or at least reputational damage, and it makes people think twice. Courts are increasingly willing to recognize that design choices in software have consequences for users, and a company that can show a serious, documented effort stands in a very different position from one that cannot.

When safety concerns conflict with revenue targets or competitive pressure, what should a credible board decision-making process look like?

A credible process is an informed one, and it leaves a record. The board should see concrete information about the costs of delay, the nature and likelihood of the safety concern, and the options for mitigating it. It should ask whether the testing was adequate and, where it was not, commission more or invite outside review. Where real uncertainty remains, there are options between launching and shelving, such as a limited release, a staged rollout or a sandbox, that let the company learn more before it commits. And whatever the board decides, what the company then says publicly about the model’s safety must be accurate, which is where much of the current litigation risk lies.

Then the board should write down what it decided and why. Over time those records become a body of precedent, so the next hard call is measured against the last one, and anyone reviewing the company can see whether its stated principles actually bind. Microsoft has started to build this kind of decision record, and I think it is one of the most underrated tools in AI governance. Competitive pressure is real, and a board may reasonably decide to move ahead. What it cannot do is decide by default.

Your work proposes a “social business judgment” standard for evaluating board decisions. How would that apply to an AI company that postpones a profitable product launch because of potential public harm?

Under existing corporate law, a board that postpones a profitable launch is already well protected. Courts defer to directors’ business judgment, and accepting lower profits today to avoid harm tomorrow fits comfortably within it. The harder question runs the other way, whether the board took public harm seriously at all.

Social business judgment asks the board to evaluate potential public harm the way it evaluates any other business risk, through tangible, measurable and specific metrics wherever possible. The emphasis is on process and due diligence, not on the outcome. A board that sources the right information, hears from those affected and records its reasoning has earned deference, whether it postpones the launch or proceeds with it. A board that does none of this has not. So in the case you describe, the postponement itself is not the issue. The issue is whether the board can show how it reached that decision, and whether it would reach the same one again on the same facts.

AI companies publish safety frameworks and voluntary commitments, but readers rarely see how those policies affect individual decisions. What evidence would distinguish effective governance from a public relations exercise?

Very often, when companies publish frameworks, the public is left wondering whether the framework ever applies to specific decisions or remains rhetoric the company never follows through on. The answer is disclosure of a particular kind. Rather than more statements of principle, companies should report the specific choices they made and how the framework shaped them, which launches were delayed, which capabilities were held back, which risk thresholds were triggered and what followed.

The most telling evidence is a decision that cost the company something. A framework that has never changed an outcome, never delayed a launch and never been invoked against revenue is hard to distinguish from public relations. These disclosures also need to be specific enough to verify, and for the most consequential questions they should be verified by a third party with the standing to say so publicly.

Where does corporate self-governance reach its limits? Which decisions about acceptable AI risk require public rules or democratic oversight?

The most important limit has to do with failures to monitor. Even a well-intentioned, responsible company can face a major crisis, and at that moment conflicts of interest multiply. Employees may withhold information that would hurt their jobs or their bonuses. Executives may suppress bad news to avoid public criticism. Fear of lawsuits may keep the company from being open. In past corporate crises, the crucial information usually came from people leadership could not control, external auditors and reviewers, affected parties, or internal whistleblowers using independent reporting channels.

This is where public rules matter most. The law can add relatively little at the rulemaking stage, because statutes tend to borrow their key concepts from the companies. It can add more on disclosure, by reaching firms that will not disclose voluntarily. And it adds the most on verification, by requiring independent review, protecting whistleblowers and giving outside checks real authority.

The other limit is collective. The most a single company can do is stop itself. It cannot stop its competitors. Decisions about catastrophic risk, including whether to slow the frontier, require coordination across firms and governments, at home and internationally. One model I find promising is an international standard-setter along the lines of the ISSB in sustainability reporting, independent of governments but overseen by trustees that include them. Corporate ordering is necessary for this work, because the taming of these models happens inside the companies that build them, but it cannot carry the work alone.

If you were advising the board of a fast-growing AI company today, what three governance changes would you prioritize, and what evidence would you look for a year later to determine whether they worked?

First, outreach in rulemaking. Before adopting or revising its safety framework or usage policies, the company should consult the people affected (users, civil society, domain experts) and publish what it heard and how it responded. Second, regular responsibility reporting, an annual account of the hard decisions the board and management made under the framework, with enough detail to be checked. Third, an external monitor, hired by and reporting to independent directors, with full access and the right to publish.

A year later I would look for evidence that each one actually bit. For outreach, at least one rule that changed because of what the company heard. For reporting, disclosed decisions the company would rather not have discussed, including ones that cost it revenue or time. For the monitor, findings the company did not want, and a record of what the board did about them. If after a year everything reported is good news, the changes have not worked.

Thank you for the great interview, readers who wish to explore the book and purchase options should visit the official website for Corporate Ordering.

Antoine is a visionary leader and founding partner of Unite.AI, driven by an unwavering passion for shaping and promoting the future of AI and robotics. A serial entrepreneur, he believes that AI will be as disruptive to society as electricity, and is often caught raving about the potential of disruptive technologies and AGI.

As a futurist, he is dedicated to exploring how these innovations will shape our world. In addition, he is the founder of Securities.io, a platform focused on investing in cutting-edge technologies that are redefining the future and reshaping entire sectors.