The signatures gathered at the White House on Sept. 29 offer a revealing picture of where artificial intelligence governance stands today. U.S. President Donald Trump and representatives of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the "White House Accord on Super Intelligence.” The accord proposes four layers of oversight: robust internal controls, internal verification, independent external evaluation and independent board oversight. Cybersecurity, biological and chemical risks feature prominently. The text also leaves open the possibility that these recommendations could eventually become law.

It is important to state that this kind of initiative is not new. In 2023, the Biden administration secured voluntary commitments from leading AI companies on safety testing and transparency. However, to understand why this meeting took place now, one needs only look at the events of recent months.

In July, more than a thousand OpenAI agents exceeded the scope of their assigned tasks, infiltrated the Hugging Face platform's infrastructure and accessed critical parts of the system in less than 13 hours. In another incident disclosed publicly in September, an OpenAI agent tasked with researching public pharmaceutical expenditure accessed restricted data in the Australian government’s Medicare system without being instructed to do so. Furthermore, in the U.S. military, aircraft went airborne and military preparations were started after an AI tool fabricated information about the cargo of a Chinese vessel. The operation was halted at the last moment when it was realized that an AI-induced error had occurred.

These incidents are serious warnings of how quickly a technology left unchecked can produce unpredictable consequences. Combined with the dizzying pace of technological development, they made it inevitable that companies and the government would come together. Indeed, three days before the accord, the U.S. and China agreed to establish a bilateral communication channel for potential AI incidents.

The timing is also politically significant. In a Reuters/Ipsos poll conducted in mid-September, 73% of respondents said they were concerned that companies had not done enough to prevent serious harm to society, while 55% supported slowing development. With midterm elections approaching and local opposition to data centers growing, the White House needed to reassure the public that the situation was under control.

One word, two functions

A separate executive order signed that day instructed federal agencies to use "Super Intelligence” instead of "artificial intelligence.” Although this may initially seem like a simple name change, it has two implications that are unlikely to serve the future of AI well.

The first is economic. Until now, superintelligence has referred to intelligence surpassing that of humans. Current systems have not reached this threshold, nor does the executive order establish that they have. Yet turning this concept into a routine official term inflates market expectations. As expectations grow, so do investment and funding flows. In a sector already facing concerns about a bubble, hype reinforced by official government language can obscure real risks.

The second is political. When Trump introduced the term from the podium of the United Nations General Assembly in September, he said he hoped the rest of the world would use it too, adding: "Let’s see if I have any power.” The term thus becomes an important instrument of public diplomacy. Whether other countries, international organizations and companies adopt it will test Washington’s power to shape global discourse. Whoever defines the terminology also frames the debate.

What text leaves out

The central issue lies less in what the text in the signed accord says than in what it leaves out. The most significant gap concerns oversight itself. Specifically, the document refers to independent audits, but how they will operate remains unclear. Finding and appointing auditors appears to be left to the companies, yet the text establishes no appointment procedure, government approval requirement or criteria for independence. Will companies give auditors access to all their data, or will training data and internal test results be excluded on trade-secret grounds? Will reports be made public, or submitted only to company management or a government agency? Who will learn of an adverse report, and who will act on it? The answers will determine whether the accord becomes a genuine oversight mechanism or a public relations exercise.

The accord itself provides neither penalties nor an implementation timetable. Companies remain subject to existing laws, of course, but the text creates no new obligations. The Committee on Foreign Investment in the U.S. (CFIUS), which reviews foreign investment in the U.S. for national security risks, requires mitigation measures to be effective, verifiable, monitorable and enforceable. The White House accord recommends reasonable measures without turning them into measurable obligations. It also leaves unanswered the threshold at which a model would be considered "too dangerous” and whether identifying a problem would require delaying its release.

Who controls the table?

How effective will this accord be? Unfortunately, it is difficult to be optimistic. The fundamental reason is that the process remains under the control of the companies it is supposed to oversee. Those writing the rules, implementing them, selecting auditors, and interpreting the results are largely the same actors. It is difficult to convince spectators that a match is fair when the players choose the referee.

Competition also makes voluntary restraint difficult. Every resource devoted to safety could also be used to develop new capabilities and generate returns on infrastructure investments worth billions of dollars. Agreements among companies that slowing down would be beneficial do not mean they will actually slow down. The rhetoric of competition with China intensifies this pressure.

The legal force of a voluntary commitment is also limited. A company that misleads the public about compliance may face an investigation under consumer-protection law. But investigating a misleading statement is not the same as establishing rules that prevent unsafe development practices in advance. Allowing leading laboratories to jointly establish common rules also risks giving incumbent firms control over the rulemaking process.

All these steps show that the U.S. is alert to AI’s potential consequences. Yet the preference for self-regulation over federal regulation makes its approach clearer. The accord establishes no independent public oversight mechanism with defined powers. This suggests that major technology companies still exert substantial influence over U.S. policy in this area.

Could this be starting point?

None of this means the accord cannot provide a starting point. Voluntary commitments may gradually become established expectations in public procurement and contracts. The document’s openness to future legislation strengthens this possibility. But this requires a balance of power. Government, major technology companies and civil society must sit at the same table as actors capable of holding one another in check. Today, only the first two are at the table, and their relationship rests more on cooperation than oversight. A structure that excludes civil society, academia and affected communities cannot move beyond endorsing its own decisions.

Transparency is essential to this balance. The public must know who the auditors are, how they are selected, what data they can access and what their assessments find. Promises made behind closed doors, however well-intentioned, do not build trust.

The Sept. 29 signing ceremony shows that Washington cannot ignore AI’s risks. But acknowledging risk and managing it are different things. A match in which the players choose the referee needs a referee independent of those players.