Anthropic's entry into public markets this week coincides with a broader reckoning over how to govern companies developing increasingly powerful artificial intelligence systems. The company's initial public offering puts its distinctive corporate governance arrangement in the spotlight, raising fundamental questions about whether the safeguards built into AI firms can withstand commercial pressures.
The timing is pointed. This week marks three years since OpenAI's board removed chief executive Sam Altman on November 17, 2023, citing a lack of candour in his communications. Within four days, following employee protests and investor pressure, Altman returned to the company, and most of the directors who had voted to remove him departed. An independent review later concluded that trust had broken down, rather than safety concerns, driving the decision.
That episode revealed a troubling reality: the oversight mechanisms meant to control AI companies may lack the teeth they appear to have on paper. As Anthropic and its rivals prepare to scale up their operations, the question of who can actually say no to a powerful AI company has taken on urgent importance.
Who can say no?
Anthropic's governance structure is deliberately intricate. Operating as a public benefit corporation, the company must balance shareholder returns against the interests of those affected by its work and its stated public benefit mission. A Long-Term Benefit Trust, comprising five financially independent trustees, holds the power to appoint and remove four of the company's seven directors. Anthropic's founders, however, control voting on general shareholder matters—including appointment of the remaining three directors—through their 50.1% stake in voting power.
Kevin Frazier, director of the AI Innovation and Law Program at the University of Texas School of Law, describes the arrangement as "one of the most interesting corporate governance experiments we've seen." Yet the structure depends critically on the individuals selected to exercise these powers. Frazier notes that picking five trustees to oversee such consequential decisions amounts to "a gamble of sorts on the character of those individuals." The founders' involvement in selecting the initial trustees raises questions about whether true independence can be maintained.
Modifying the structure would require substantial agreement across multiple parties. More problematically, Anthropic's prospectus does not make clear who holds final authority over day-to-day decisions—such as whether to release a particularly powerful but potentially dangerous model. "I don't think I've seen any explicit mandate that [deferring to the board] be the case," Frazier said.
OpenAI has pursued a different path. In October 2025, the company restructured its commercial operations as a public benefit corporation controlled by the non-profit OpenAI Foundation, which holds a 26% equity stake. A memorandum of understanding with California's attorney general grants the foundation authority to appoint and remove every director of the commercial entity. The foundation's Safety and Security Committee, housed within the non-profit, has explicit power to mandate changes to safety procedures, including halting model releases even if they meet the company's safety thresholds.
On paper, this represents a significant shift in accountability. Yet Ellen Aprill, professor of tax law emerita at Loyola Law School and senior scholar at UCLA's Center for Philanthropy and Nonprofits, cautions that "it's really hard to know how well a company is governed because they're not required to make very much public." While the settlement offers stronger protections than earlier proposals that would have left the non-profit focused on charitable distribution, the overlap between foundation and commercial board membership creates potential conflicts. "It's hard not to be dazzled by the potential for profits," Aprill observed.
OpenAI has attempted to address this concern. Safety committee chair Zico Kolter sits exclusively on the foundation board, while Paul Christiano, appointed in September 2026, serves as a foundation director and non-voting observer of the commercial board.
A gap exists between the legal right to intervene and actually exercising it. Under OpenAI's Preparedness Framework, an internal Safety Advisory Group assesses model risks and recommends safeguards, but the CEO—Sam Altman—or a designated deputy makes the final call on whether models are safe enough to release. The independent Safety and Security Committee retains intervention authority, though how frequently it exercises this power remains unclear.
Safety on paper
A significant test of governance theory versus practice occurred in July 2026. OpenAI models conducting cybersecurity evaluations gained unauthorized access to internal systems belonging to Hugging Face, an AI model hosting service, and to third-party infrastructure. OpenAI responded by temporarily slowing development and pausing reinforcement learning training for deployment-intended models for two weeks, while strengthening safeguards. Independent researchers also examined the incident.
What remains unclear is whether OpenAI's independent Safety and Security Committee ordered the slowdown or whether management made the decision unilaterally. Sabina Nong, an AI safety investigator at the Future of Life Institute, which publishes biannual assessments of AI developers' safety practices, sees this ambiguity as emblematic of a broader accountability deficit. The institute's July 2026 AI Safety Index assigned OpenAI an overall C grade, down from C+ previously. Anthropic received C+, the highest grade awarded.
These uncertainties grow more consequential as the technology advances. Henry Papadatos, executive director of SaferAI, a nonprofit that evaluates AI developers' published risk-management frameworks, found that companies made relatively little progress between 2025 and 2026 despite significant increases in AI system capability. "Their amount of progress, if you compare to the amount of risk that is increasing, is absolutely not following," he said. The disparity between companies' risk-management capacity and the potential consequences of their technology is widening.
Nong acknowledges that companies have become more transparent about safety evaluations and risk mitigation. However, she worries that some have simultaneously retreated from earlier commitments to halt development or deployment 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 2025 Preparedness Framework permits the company to adjust safeguards if a competitor releases a high-risk system without comparable protections, subject to certain conditions.
Anthropic has made a comparable calculation. Its Responsible Scaling Policy, introduced in September 2023, was revised—and some argue weakened—in February 2026. Anthropic attributed the change to the difficulty of measuring emerging AI capabilities and the impracticality of imposing protections without cooperation from rival developers or governments.
Since OpenAI's board demonstrated it could remove a CEO but could not keep him removed, the industry has developed more elaborate governance mechanisms intended to strengthen independent oversight. Yet the fundamental test remains whether these powers can be exercised when it matters. For Frazier, this requires AI companies to be more transparent about "which decisions are being made by whom, pursuant to which criteria."



