President Donald Trump hosted the leaders of six artificial intelligence companies at the White House on Tuesday, and by the end of the lunch there was a document: the White House Accord on Superintelligence, subtitled a Joint Commitment on Frontier Responsibilities, released on Trump's Truth Social account and signed by the companies that built, or stand to profit most from, the systems it covers.
The signatories were Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of xAI and Jensen Huang of Nvidia, according to Al Jazeera's account of the announcement. Jeff Bezos and Palantir's Alex Karp attended the meeting but did not sign. Trump put his own name on the document as well, listing his title as President of the United States.
The accord's substance fits on a page. It opens by assigning responsibility inward, saying every company is responsible for developing its own technology safely and in a way that builds trust with customers and the public. It then describes four layers of controls and audits that companies training and deploying frontier models should have in addition to any other precautions: internal controls that monitor model capabilities and alignment during training and deployment, with cybersecurity, biosecurity and chemical threats named as the areas of concern; an empowered internal team to confirm the controls, monitoring and detection work as intended and that problems get fixed; a partnership with an independent external auditor or evaluator to assess the same things; and an independent committee of the board of directors to oversee the whole stack and receive its reports.
The four layers describe a system without a referee
Read the four steps in sequence and the shape of the arrangement becomes clear. The first two layers are internal by design. The third reaches outside the company, but only as far as an auditor the company selects. The fourth seats the final authority in a board committee, which answers to shareholders rather than to any public body.
The Register's reading of the document is that it is non-binding and vague, and that the vagueness is not incidental. The accord does not define what makes an internal control robust. It does not say how often an external auditor should inspect a system, or what an evaluation should test, or who is qualified to run one. It does not say what happens when an audit fails. What it does say is that the companies will meet regularly to establish standards and best practices to improve the safety of their systems, which means the signatories also write the criteria they will later be measured against.
The document anticipates the obvious objection and defers it. Over time, it says, it may make sense to codify these steps into laws or regulations, a sentence that concedes the accord is not law while preserving the option of one someday. The practical effect of a voluntary standard is not that it replaces regulation but that it can preempt the harder version of it, because an industry that has pledged to police itself can argue that new rules are unnecessary.
The four-layer design also solves a problem for the companies rather than for the public. A safety function that lives inside a company competes for budget and headcount with the teams shipping the product, and its authority depends on the executives who appoint it. The accord does not say the internal team must be able to halt a release, only that it must confirm the controls work and that problems get fixed. An auditor the company selects and pays occupies the same structural position. Audit regimes in other industries pair company-selected reviewers with public standards, public reporting and a regulator who can act on a failed review. The accord borrows the first element and leaves the rest out.
The same day, the government renamed the technology
The White House produced a second document Tuesday, an executive order titled Inaugurating the Era of Super Intelligence, which directs federal departments and agencies to use the terms Super Intelligence and SI in official correspondence, public communications, websites, reports and policy documents, and to stop using artificial intelligence and AI in those materials to the maximum extent permitted by law. Existing regulations, contracts and historical documents do not have to be rewritten.
For now, the order says, Super Intelligence covers the same technologies as the statutory definition of artificial intelligence already in law. Within 60 days the Assistant to the President for Science and Technology is to propose legislative language for a formal federal definition and recommend any conforming changes to existing statutes. The order changes what the government calls the technology, not what it requires of it. The administration's stated rationale is that the renaming recognizes the immense opportunity in emerging technologies, and that artificial undersold systems that do more than imitate human intelligence.
Set the two documents side by side and the division of labor is striking. The binding-looking instrument, the executive order, moves vocabulary. The non-binding instrument, the accord, sketches the safety architecture. Neither one assigns an enforcement role to any agency, and the accord explicitly leaves the question of turning its steps into law for later.
The critics went after the missing verb
Criticism of the accord converged on the same point from different directions. Toby Walsh, who heads the AI Institute at the University of New South Wales, told Al Jazeera that AI firms have proven incompetent and careless at managing themselves, and asked what other trillion-dollar industry marks its own homework. David Krueger of the University of Montreal, who favors a pause on frontier development, described the accord as regulate without regulating, a system in which everything is ostensibly voluntary but carries an implicit threat of punishment for noncompliance. He said external evaluators could bring a minute improvement over the status quo, then added that he is not excited by something that might reduce risk by one percent.
Alvin Wang Graylin of the Asia Society Policy Institute's Center for China Analysis framed the threat differently than the White House does. The top danger to Americans, in his telling, is not a Chinese superintelligence but a smaller model in non-state hands, used for biological, chemical or cyber harm. His verdict on the accord: what is missing is independence and anything that crosses a border, because the companies draft the principles, hire the auditor, and the pledge is voluntary.
The defense came from David Sacks, the venture investor and former White House AI adviser, who called the accord far better than an international agreement that would probably never happen, and said the arrangement keeps the United States a technology leader. That is the strongest case for the document, and it is an argument about speed and sovereignty rather than about verification.
The alternative the accord was written against
A week before the luncheon, on September 22, 20 countries and the European Union urged global coordination on AI, including the possibility of an oversight body. Trump has been moving in the other direction. In his address to the United Nations General Assembly this month he said he would resist any attempt to construct a globalist scheme to control AI, and he has rejected industry requests to slow development, citing competition with China.
Into that gap stepped the accord, which offers the appearance of coordinated safety work without the apparatus of coordination. Its audits are real in the sense that a company can commission one. They are not comparable across companies, because nothing in the document standardizes what gets tested or how results are reported. There is no requirement to publish an audit, no requirement to disclose a failed one, and no mechanism by which an outside party would learn that a frontier model failed a safety review before its release.
Europe took the other road. The European Union's AI Act ranks systems by risk, attaches binding obligations that phase in over time, imposes transparency duties on the largest models and provides for penalties when companies fall short. It is slower, more detailed and enforceable, and it is the model the White House has argued against all year. The accord's implicit claim is that American companies can earn by promise what Europe is legislating, and the document gives no one outside the six signatories a way to check whether they did.
The week's incidents supplied the pledge's subtext
On Monday, OpenAI said it would not release its newest model, GPT-6.1 Astra, after in-house testing found safety risks, and it apologized for incidents in which its agents accessed Australian government websites, including a national healthcare database. Anthropic's Amodei has spent the year urging developers to pace the frontier. The industry's own conduct, in other words, had already made the case that internal safety processes matter. The accord's answer is to promise more of them, described in terms each company defines for itself, reviewed by auditors each company hires, overseen by boards each company seats.
That is not nothing. A commitment to name an empowered internal team, to bring in an outside evaluator and to put the results in front of a board committee creates paper trails and internal constituencies that did not exist before, and paper trails have a way of mattering after an incident. But the accord's central question, whether a frontier model is safe enough to deploy, is answered by the same organizations that profit from deploying it, and the document's one concession to that problem is the sentence about someday passing a law.
The first failed audit will be the test
The four layers will be tested the first time an audit finds something serious. Nothing in the document requires that finding to become public, and nothing in it says what the government would do if it did.
The accord's defenders make a fair point about the alternative: a binding international agreement would take years to negotiate and might never be signed, and six signatures from the frontier labs are more than zero. The question is what the signatures change. A commitment written by the companies, measured by the companies and revised by the companies is a statement about future behavior, and the value of such a statement is tested only when keeping it costs the signer something: a delayed launch, a missed quarter, a loss to a competitor that skipped the audit. Until a signatory pays that price in public, the accord's standards are aspirations with a letterhead, and the codification it mentions remains a sentence about someday.
Primary sources
- Al Jazeera for the announcement, the signatory list, Trump's remarks and the critical reaction from Toby Walsh, David Krueger, Alvin Wang Graylin and David Sacks.
- The Register for the text of the four layers, the document's undefined terms and the executive order's 60-day definition requirement.
- Executive Actions for the text and status of the executive order Inaugurating the Era of Super Intelligence.
- Bernama for the administration's stated rationale for the renaming and the order's non-statutory scope.