Who Actually Had the Power

Published October 1, 2026

On November 17, 2023, OpenAI's board fired Sam Altman as CEO. The board had the authority to do it. OpenAI's unusual governance structure placed control of the company with its nonprofit board, which had the power to appoint and remove the chief executive. When it announced Altman's departure, it said it had lost confidence in him after concluding he had not been consistently candid in his communications with the board.

Five days later, Altman was back as CEO, and most of the directors who had removed him were gone.

Those five days expose a distinction that's easy to miss when people talk about power: having the authority to make a decision and having enough power to make it stick are not the same thing. OpenAI's board found out the difference unusually fast.

Social psychologists John French and Bertram Raven gave us a useful way of separating things we tend to lump together as power. Their 1959 model identified five bases of it, and Raven later added a sixth. Legitimate power comes from a position or role: a board can remove a CEO because the organization's rules give it that authority. Reward power and coercive power come from controlling outcomes for other people, such as who gets promoted, who gets paid and who gets removed. Expert power comes from possessing knowledge or capability other people depend on. Referent power comes from people's willingness to identify with, trust or follow someone, regardless of what the org chart says. Informational power, added later, comes from information and arguments that can change what someone thinks or does.

The useful part of the framework isn't ranking who's "most powerful." It's that different people can hold different kinds of power at the same time, and those forms of power can point in opposite directions. OpenAI's board held legitimate power in an unusually clean form, and on November 17 it used it. Then it found out what the other forms of power could do.

Microsoft moved quickly. Within days, CEO Satya Nadella announced that Altman and co-founder Greg Brockman would join Microsoft to lead a new advanced AI research team. Microsoft was already OpenAI's most important commercial partner, supplied much of the computing infrastructure behind its models and had committed billions of dollars to the relationship. It could now offer Altman somewhere to go immediately, removing much of the uncertainty that normally comes with leaving.

More than 700 OpenAI employees eventually signed a letter threatening to leave unless the board resigned and Altman was reinstated. Among the signatories was Ilya Sutskever, OpenAI's chief scientist and one of the directors who had participated in removing Altman in the first place. Three days after the firing, Sutskever posted that he deeply regretted his participation in the board's decision and would do everything he could to reunite the company.

Sutskever is the clearest example of why the framework matters. As a director, he held legitimate power and had exercised it. As one of OpenAI's most important technical figures, he also held substantial expert power. Within days, the same person who had exercised legitimate power as a director was using his standing inside the company to help undo the decision.

The board's authority hadn't disappeared. Nobody discovered that the bylaws were invalid or that the directors lacked the right to remove the CEO. What changed was what everyone else was prepared to do about it. By November 22, OpenAI had a new initial board and Altman was back as CEO. The original decision had lasted five days.

French and Raven help explain why firing Altman didn't settle anything. The board had legitimate power. Altman had substantial referent power among employees. OpenAI's technical staff collectively held expert power the company couldn't easily replace. Microsoft controlled resources that gave it leverage of its own, but there's something in this case the categories alone don't explain.

Most employees in most organizations hold some expert power too. A good engineer knows systems the board doesn't understand. A salesperson may own relationships that would be hard to replace. A senior executive may have people who would rather follow them than whoever replaces them. Yet people usually comply when legitimate authority is exercised because acting against it is expensive. Leaving means finding another job, possibly giving up compensation or status and accepting the risk that the next place won't work out. Your colleagues may agree with you but still stay. Even when hundreds of people dislike the same decision, getting hundreds of people to do something about it is difficult, and formal authority benefits from all of that friction without ever having to create it.

OpenAI's situation was unusual because much of that friction disappeared almost at once. Microsoft already knew the people involved, already worked closely with OpenAI, had the money and the compute, and publicly offered Altman and Brockman a landing place within days. Employees threatening to leave were therefore making that threat with an unusually credible alternative sitting on the other side of it.

That changes how the five days look. The board didn't lose its legitimate power, and Altman didn't suddenly acquire influence he hadn't possessed before. Those things were already there. What changed was the cost of acting on them.

This travels well beyond OpenAI. Organizations contain people whose formal authority is obvious because it's written down: boards, executives, managers, regulators. Other forms of power are harder to see because they sit in expertise, relationships, reputation, information or the willingness of other people to follow.

Usually these coexist without being tested against each other, so the org chart gives a reasonably convincing picture of who's in charge. A crisis can show something different. A hospital can decide how a department should run, but making that decision work may depend on a handful of specialists who are difficult to replace. A consulting firm can decide who leads an account, but the client relationship may sit with someone else. A technology company can reorganize a team on paper and discover that the engineers who understand the system have options outside it.

None of this makes formal authority meaningless. In most organizations it works perfectly well, partly because walking away remains costly and coordinating resistance remains hard. That's what makes OpenAI useful as a case: it shows what happens when those ordinary conditions are briefly removed.

The board had the authority to fire Altman before November 17, and it still had that authority afterward. What those five days revealed was how much of its practical power depended on conditions outside the boardroom: employees staying, expertise remaining inside the company, partners continuing to cooperate and credible alternatives being difficult to find.

Once those conditions changed, the governance documents hadn't changed with them. The board could still make the decision. It just couldn't make everyone else live with it.