Five days in November: what the OpenAI board crisis showed about governance

What happened, in order

On Friday November 17 the OpenAI board announced that it had removed Sam Altman as chief executive, saying it no longer had confidence in his ability to lead and that he had not been consistently candid in his communications with it. The board that made the decision had four members: chief scientist Ilya Sutskever, Quora chief executive Adam D'Angelo, Tasha McCauley, and Helen Toner of the Center for Security and Emerging Technology. Three other directors, Reid Hoffman, Shivon Zilis and Will Hurd, had left earlier in the year and had not been replaced. Mira Murati, the chief technology officer, was named interim chief executive.

Within hours Greg Brockman resigned as president and chairman, and researchers Jakub Pachocki, Szymon Sidor and Aleksander Madry followed. Over the weekend Microsoft announced that Altman and Brockman would lead a new research group inside the company. The board, rather than reverse itself, appointed a second interim chief executive, Emmett Shear, formerly of Twitch. Other candidates, including Nat Friedman and Alex Wang, had declined.

On Monday November 20 a letter signed by roughly 745 of the company's 770 employees threatened mass resignation unless the board stepped down. Sutskever signed it and apologised publicly for his part in the removal. On Wednesday November 22 Altman returned as chief executive under a new interim board of Bret Taylor as chair, Lawrence Summers, and D'Angelo, the one director who survived the week. Altman and Brockman agreed not to take board seats for now and accepted an internal investigation into the events.

The structure worked exactly as designed, once

It is worth being precise about what the structure was supposed to do. OpenAI's commercial arm sits under a non-profit whose board is not accountable to investors or employees. The whole point of that arrangement, as the company described it, was that the board could act against commercial interest if it judged the mission required it. On November 17 the board did precisely that. Nothing in the bylaws was violated. The mechanism fired.

What the week showed is that a legal right to act is a different thing from the power to make an action stick. The board had authority over the company on paper. It did not have the staff, who could walk to Microsoft en masse, and it did not have the investors, who had no formal vote but a great deal of informal reach. Once 745 of 770 employees said they would leave, the board's authority became a claim about an empty building.

We want to separate two questions that got merged in the commentary. One is whether the board was right on the merits. We do not know, and neither does anyone outside the room, because the board gave almost no public reasons. The other is whether a board that cannot explain itself can survive using an emergency power. The answer to the second turned out to be no, regardless of the answer to the first.

Why the board could not explain itself

The most consequential failure of the week was not the decision. It was the absence of any account of the decision that employees could weigh. A four-person board removed a chief executive with a two-sentence statement, and then spent five days declining to say more, even to the people whose continued employment decided the outcome. Whatever the directors believed, they acted as if belief alone would be enough.

Compare this with how a public company board would behave. It would have a chair independent of the executive, outside counsel, a communications plan, and, above all, a large enough board that no single resignation changes the majority. OpenAI's board had shrunk to four because departures were not replaced. A body that small has no slack. When Sutskever changed his mind on the third day, the board lost a quarter of itself and any claim to represent a settled view.

The lesson we take is narrower than many drew. Non-profit control may still be workable. A control mechanism nobody expects to be used will not be maintained, though, and an unmaintained mechanism fails when finally pulled. The board had authority and no infrastructure for exercising it.

What the new board tells us

The interim board announced on November 22 has three members, all with long experience of corporate governance and none with a background in AI safety research. Taylor ran Salesforce as co-chief executive and chaired Twitter's board during its sale. Summers is an economist and former Treasury secretary. D'Angelo is the continuity member. The directors who framed the removal in terms of the mission, Toner and McCauley, are gone.

That composition is a signal about what the company and its backers want from the board going forward. It reads as a board built to reassure investors and partners that the lab is governed like a normal large company. Whether it will also be a board willing to fire the chief executive over a mission disagreement is a question we would not bet on either way, but the selection criteria visible from outside point in one direction.

Altman and Brockman have accepted an investigation. If its findings are published in full, we will learn something about whether the original board had grounds. If they are summarised in a paragraph, we will learn something about the new board instead.

Will it be tried again

Our honest expectation is that no lab board will attempt this again in the same form. The week gave every future director a vivid demonstration of the cost of acting without staff support and without a public case, and gave every lab's employees a demonstration of how much power they hold. Both sides will update. Boards will become more cautious, and the emergency power will remain on paper while its practical use recedes.

If someone wants non-profit or mission-based control of a frontier lab to be real rather than notional, the design problem is now clear. The board needs enough members that it can lose one without losing its majority, a written process for removing executives that includes making a case to staff, and some commitment from staff and investors in advance that they will respect an outcome they dislike. None of that existed on November 17. We would want to see a lab adopt those mechanisms before we believed the next such structure would hold under load.