Dario Amodei, CEO of Anthropic, recently called to “pace the frontier” of artificial intelligence. He has a valid point, but I think much of the debate following his call is focused on the wrong one. The most important question isn’t how much AI development should slow. It is who gets to decide when the risks have become serious enough that it should. That is not primarily a technology question. It is a governance question.
He also proposed placing independent evaluators inside leading AI companies, with access comparable to internal safety teams. That may prove more important than his proposal to slow frontier development.
I spent much of my career advising corporate boards and compensation committees on executive incentives and governance. One lesson from that experience is simple: incentives matter. The executives deciding whether to proceed with the next frontier model are the same people whose careers, compensation and competitive position benefit from proceeding. That doesn’t make them untrustworthy. It just makes them management.
Corporate governance exists in part because someone outside the operating chain needs both the information and authority to challenge consequential management decisions. That someone is the Board.
Give the Board Its Own Experts
This arrangement isn’t particularly exotic. Compensation committees routinely retain independent consultants who have extensive access to management and company information but report to the committee, not the executives whose compensation they evaluate. The consultant doesn’t run the company or set executive pay. The consultant makes sure directors receive independent advice before they make the decision.
Frontier AI could use the same governance principle. Major AI developers should have independent AI evaluators with unrestricted access to models, evaluation systems, relevant researchers and the development pipeline itself—including the use of today’s AI to help build tomorrow’s.
The evaluators should be selected by and report directly to a board-level AI Risk Committee. Their compensation and continued retention should be controlled by that committee, not management. They should be embedded enough to understand the technology, but independent enough to tell the board what management may not want it to hear.
Make Someone Own the Decision
Suppose an evaluator concludes that a developing model has crossed an established threshold for dangerous autonomous cyber capability or can accelerate AI research without adequate human supervision.
What happens next?
Routine findings can remain with management. Material findings go directly to the AI Risk Committee. Critical findings—particularly recommendations to stop or materially alter a major training program—go to the full board.
Management can disagree. Boards regularly hear disagreements between management and independent advisers.
But overriding a critical independent safety finding should require affirmative board approval.
And critical overrides should be publicly disclosed.
Documentation buried in confidential board minutes isn’t accountability. If directors decide to proceed despite a critical independent finding, investors, employees, customers and the public should know that the finding occurred and that the board chose to override it.
Directors care deeply about their reputations, but reputational accountability requires transparency.
A board presented with a documented independent finding of potentially catastrophic risk—and knowing that an override will become public—is in a very different position from executives privately deciding whether to continue.
The directors now own the decision.
Good governance does not eliminate risky decisions. It makes sure consequential decisions are made by people who have the information, authority—and reputational exposure—to own them.
The Road Already Traveled
We don’t need to invent this governance principle for artificial intelligence. For decades, boards have dealt with a fundamental problem: management possesses more information than directors and has powerful incentives concerning the decisions directors must oversee. The solution has been to give boards independent advisers, direct information channels and responsibility for consequential decisions.
AI raises the stakes enormously. The governance principle remains sound. Robert Frost wrote about the road less traveled. For AI safety, we might first try the road already traveled: independent expertise, informed directors, transparent decisions and accountability for those who make them.