As Anthropic Goes Public, Who Has the Power to Stop an AI Company?
Chris Stokel-Walker / Nov 19, 2026
(Photo by Samuel Boivin/NurPhoto via AP)
It’s a big week for AI, with Anthropic becoming a publicly traded company today, testing investor appetite for the industry’s extraordinary growth. But its initial public offering also puts an unusual corporate governance experiment under the spotlight.
Under the arrangements set out in Anthropic’s own IPO prospectus, the firm’s founders will retain 50.1% of the voting power on general shareholder matters. An independent trust established to protect the company's mission will have the authority to appoint a majority of its directors. The structure is designed to balance commercial pressures against the company’s stated commitment to developing powerful AI responsibly. But it raises an awkward question: when those interests collide, who actually gets to say no?
There is good reason to ask.
This week also marks three years since the board of OpenAI, the company behind ChatGPT, abruptly fired its chief executive, Sam Altman, on Nov. 17, 2023. The board had then said he had not been “consistently candid” in his communications with them.
Four days later, following an unprecedented employee revolt and huge pressure from investors, Altman returned back to OpenAI. As he settled back into his role, toting an all-access pass around his neck, most of the board members who had removed him were heading in the opposite direction.
Three years on, the precise dynamics behind the boardroom coup remain contested, although an independent review found that a breakdown in trust, rather than concerns about AI safety, prompted the decision.
The episode exposed an uncomfortable truth about the people and personalities behind AI systems that have become increasingly embedded in our lives: the mechanisms designed to oversee these companies may not be as powerful as they appear. And as Anthropic prepares to enter the public markets, that question takes on fresh urgency.
Since Sam Altman’s temporary sojourn from the CEO seat, OpenAI has undertaken a massive, significant corporate restructuring, while its rivals, including Anthropic, have developed their own idiosyncratic approaches to how to govern the development of increasingly powerful AI models. Both companies have published safety commitments, appointed independent directors and created committees supposedly capable of intervening when things go wrong. Though less clear is whether they work.
Who can say no?
Anthropic’s corporate governance structure is unusually complicated. As a public benefit corporation, it is required to balance shareholder interests, the interests of those affected by its activities, and its stated public benefit mission.
Its Long-Term Benefit Trust, designed to contain five financially disinterested trustees, has the power to appoint and remove four of the company's seven directors. Yet its founders have control over general shareholder votes — crucially including the appointment of the remaining three directors.
For Kevin Frazier, director of the AI Innovation and Law Program at the University of Texas School of Law, the arrangements represent “one of the most interesting corporate governance experiments we’ve seen.”
But they also depend heavily on the people appointed to exercise those powers. Selecting five individuals to oversee such consequential decisions is, he said, “a gamble of sorts on the character of those individuals.” The company’s founders have been involved in selecting the initial trustees, something Frazier said could either help preserve Anthropic’s mission… or raise questions about how independent those trustees really are.
Making any changes to the structure is wickedly tricky, needing substantial agreement from a variety of parties. But who has final say over the day-to-day questions that firms like Anthropic might have to tackle — like whether or not to release a particularly powerful but potentially dangerous model — isn’t immediately clear from Anthropic’s prospectus. “I don't think I’ve seen any explicit mandate that [deferring to the board] be the case,” said Frazier.
OpenAI’s governance has also undergone its own changes. In October 2025, the company converted its commercial operations into a public benefit corporation (PBC), controlled by the non-profit OpenAI Foundation, which holds a 26% equity stake in the business. And a memorandum of understanding with California's attorney general established that the foundation retains the authority to appoint and remove every director of the commercial company.
The same memorandum also gave OpenAI's Safety and Security Committee, which sits within the non-profit, the explicit authority to require changes to safety procedures up to and including halting the release of AI models even if they would otherwise meet the company’s safety thresholds.
On paper, that seems like a significant shift. But “it’s really hard to know how well a company is governed because they’re not required to make very much public,” said Ellen Aprill, professor of tax law emerita at Loyola Law School and a senior scholar at UCLA's Center for Philanthropy and Nonprofits.
Aprill reckons the settlement offers stronger protections than earlier proposals that would have left the non-profit primarily responsible for distributing charitable grants, with little meaningful control over the business. But because many of the foundation’s directors also sit on the commercial company’s board, “it’s hard not to be dazzled by the potential for profits.”
For its part, OpenAI has seemingly recognized that as a worry and taken steps to try and mitigate its impact. Safety committee chair Zico Kolter serves exclusively on the foundation board, while Paul Christiano, appointed in September 2026, also serves as a foundation director and non-voting observer of the commercial board.
There’s also a difference between the legal right to intervene and actually doing so. Under OpenAI's Preparedness Framework, an internal Safety Advisory Group assesses model risks and recommends safeguards, but the ultimate decision about if models are safe enough to release falls to the CEO — Sam Altman — or a designated alternative. The independent Safety and Security Committee can intervene, but how often it does or would is still not clear.
Safety on paper
One big test of the gap between theory and practice came in July 2026, when OpenAI models conducting cybersecurity evaluations gained unauthorized access to internal infrastructure and third-party systems belonging to AI model hosting service Hugging Face. OpenAI responded by temporarily slowing development, pausing reinforcement learning training for models intended for deployment for two weeks, and strengthening safeguards. Independent researchers also investigated the incident.
OpenAI responded to the issue, but it’s not clear whether OpenAI's independent Safety and Security Committee had ordered the slowdown, or whether the decision came from management. Sabina Nong, an AI safety investigator at the Future of Life Institute, which publishes a biannual assessment of AI developers' safety practices, said that illustrates a wider accountability problem. Her organization's July 2026 AI Safety Index gave OpenAI an overall C grade, down from C+ in its previous assessment. Anthropic received C+, the highest grade awarded.
How important those uncertainties are is increasing as the technology develops. Henry Papadatos, executive director of the nonprofit SaferAI, which assesses AI developers’ published risk-management frameworks, said that his organization’s assessments found relatively little improvement between 2025 and 2026, even as AI systems became significantly more capable. “Their amount of progress, if you compare to the amount of risk that is increasing, is absolutely not following,” he said. The gap between companies’ abilities to manage risks and the potential consequences of the technology they are developing is widening, he reckoned.
In other words, companies' ability to manage risks is not keeping pace with the potential consequences of the technology they are developing.
Nong acknowledged companies have become more transparent about safety evaluations and risk mitigation, but is concerned that some have simultaneously weakened their previous promises to stop developing or deploying models when certain risks emerge. “You can write a lot of fancy stuff in risk governance,” she argued, while still deciding “the past commitments do not matter anymore.” OpenAI's own 2025 Preparedness Framework allows it to adjust safeguards if a competitor releases a high-risk system without comparable protections, subject to certain conditions.
Anthropic has made a similar calculation. Its Responsible Scaling Policy was first introduced in September 2023 but was revised, and some would say weakened, in February 2026. Anthropic said that the change reflects the difficulty of measuring emerging AI capabilities and the impracticality of imposing protections without cooperation from rival developers or governments.
That’s not to say there hasn’t been progress since 2023. Three years after OpenAI's board demonstrated that it could fire a CEO — but couldn’t keep him fired — its rivals have developed elaborate governance mechanisms intended to make independent oversight more effective.
But the fundamental test is whether those powers can be exercised in real life. For Frazier, that means AI companies need to be more transparent about “which decisions are being made by whom, pursuant to which criteria.”
Authors

