Reported 29 Sept 2026 Governance

Six AI companies signed a White House safety accord that says they should audit themselves. Four companies at the same lunch did not sign.

On September 29, at a private lunch at the White House, President Trump and the heads of six companies signed a one-page document now called the White House Accord on Super Intelligence. It asks each company to monitor its models for cyber, biological and chemical risk, to staff an internal team that checks those controls work, to hire an independent outside auditor to assess them, and to put an independent board committee over both.

The signatories are Google, Meta, Anthropic, Nvidia, xAI and OpenAI, the last represented by its president rather than its chief executive, who was at his own conference in San Francisco. The caveat that matters is the grammar and the arithmetic: the document says should rather than must, carries no penalty, no deadline and no regulator, and Microsoft, Amazon, AMD and Palantir had executives at the lunch who declined to sign it.

Until now

Outside evaluation of frontier models was a set of separate bargains, each company naming its own inspector, with Congress and two courts arguing about whether any of it should be mandatory.

Now

Six companies and the President have put the same four commitments on one page with the word should, and four large companies in the room refused to sign even that.

Count what this replaces. Over five weeks Anthropic offered desks and badges to outside inspectors, Microsoft published a code of conduct that binds nothing, OpenAI told Congress it could accept inspectors written into federal law, the European Commission said it was time to pace the frontier, four subscribers sued all four labs for coordinating on exactly this, and a federal appeals court said the Pentagon may shut a company out over what its model refuses to do. All of that argument was about whether outside review should be compulsory. The answer arrived on one page, and the answer is no: four sentences beginning should, chosen auditors, unpublished reports, and nothing anybody can enforce. A company that does none of it faces precisely what a company that does all of it faces.

The informative part is not the six signatures but the four refusals. Microsoft, Amazon, AMD and Palantir were in the room and did not sign a document with no penalties in it, which is a strange thing to decline unless declining costs less than signing. Read against the detail that OpenAI sent its president while its chief executive stayed in San Francisco, and against Anthropic's own founder saying the mechanisms are still under discussion, the week's picture is not an industry agreeing to be watched. It is an industry agreeing to be seen agreeing, with a minority declining even that, and the governance question that was being contested in Congress and the courts has been answered with a page that says should.

How it got here

Sep 12
Anthropic says it will give outside inspectors desks, badges and company laptops; OpenAI says it is considering the same. Edition 020's card, the parent of this one.
Sep 18
Anthropic names its outside inspectors and picks a consulting firm rather than a safety institute. Edition 026's card.
Sep 25
The White House Office of the National Cyber Director is reported to have asked OpenAI and Anthropic to withhold new models from Britain's testing agency. Edition 033's card.
Sep 29
Six chief executives sign a one-page voluntary accord at a White House lunch. Four executives present decline. The President calls it morally binding.
Who has to react
Microsoft, Amazon, AMD and Palantir, each of which now has to answer why it would not sign a page with no penalties; the four subscribers whose antitrust case says the labs coordinated on safety restraints, who have been handed a signed document in which six companies agree to the same four controls at the same table; the Florida court weighing whether to bar OpenAI from developing models without outside approval, which is the compulsory version of commitment three; Governor Newsom, who answered it the next day with thirteen statutes and an executive order; Senators Warner and Warren, and the sixteen senators whose OpenAI answers are due today; the European Commission, whose AI Act makes some of this mandatory and which now has an American benchmark that does not; Britain's AI Security Institute and METR, both of which could be the independent evaluator in commitment three and neither of which is named; and every auditor in the world, because four sentences just created a market and left the buyer to pick the seller.

The evidence

The accord has not been published by the government. We searched whitehouse.gov this morning: the presidential actions page carries three executive orders dated September 29, the briefings page carries nothing after September 29, and neither the Super Intelligence executive order nor its fact sheet mentions an accord, a signing, a signatory or any safety commitment. The document reached the public when President Trump posted an image of it on Truth Social, and its four commitments are quoted in identical words by The Week, SiliconANGLE, Townhall, Breitbart, CoinDesk and Rio Times. We have not read it at a government source, so this card is Reported. It also surfaced late for us, and we are saying so plainly rather than dating it by the coverage: the signing was at a private White House lunch on Tuesday September 29, the reporting broke on September 30, and the card is dated September 29. Verification event: the accord appearing at whitehouse.gov, in the Federal Register or on the website of any signatory, or any signatory naming the independent external auditor it has engaged under it, by October 31, 2026.

ObservedTuesday September 29, 2026 is the event date. The signing was at a private White House lunch followed by a press gaggle outside the West Wing, and the document became public when the President posted it on Truth Social; the coverage broke the following day, which is why it was not in yesterday's edition.

The whole observation, as recorded

The four commitments, quoted as every account carries them, are that each company will "implement robust internal controls to monitor the capabilities and alignment of its models during training and deployment around areas like cybersecurity, biosecurity, and chemical threats"; "empower an internal team to ensure all of the controls, monitoring, and detection are operating as intended, and that any issues are remediated"; "partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended"; and "designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls and the internal and external auditors and evaluators." Signatories as reported: Sundar Pichai for Google, Mark Zuckerberg for Meta, Dario Amodei for Anthropic, Jensen Huang for Nvidia, Elon Musk for xAI, and Greg Brockman for OpenAI. Sam Altman was absent, at OpenAI's developer conference. House Speaker Mike Johnson is reported present. Four executives attended and did not sign: Satya Nadella of Microsoft, Jeff Bezos of Amazon, Lisa Su of AMD and Alex Karp of Palantir. What the document does not contain, and this is the whole of it: no penalty, no fine, no implementation deadline, no regulator, no reporting requirement, no definition of an acceptable auditor, no obligation to publish an audit, and no mechanism by which anybody outside a company learns whether it did any of this. Companies choose their own auditors and decide how to remediate what the auditors find. The President's words: "morally binding"; "It's almost like a constitution in a way, the biggest people in the world signed that"; and "a form of protection." Dario Amodei is reported saying the technology has "very real risks" and that mechanisms remain "under discussion." Senator Mark Warner: "Self-regulation? That's a recipe for disaster." One more reported detail we are including because it is checkable and was not corrected by the end of the day: the document as posted misspells United States as Unites States. The same day the President signed an executive order replacing artificial intelligence with Super Intelligence across the executive branch, which is where the accord's name comes from.

Established

Six companies, Google, Meta, Anthropic, Nvidia, xAI and OpenAI, signed a one-page voluntary accord with President Trump at the White House on September 29, 2026.
The document commits signatories to internal controls on capability and alignment, an internal team verifying them, an independent external auditor, and an independent board committee.
The document uses should rather than must and contains no penalty, deadline, regulator or publication requirement; signatories choose their own auditors.
Satya Nadella of Microsoft, Jeff Bezos of Amazon, Lisa Su of AMD and Alex Karp of Palantir attended and did not sign.
OpenAI was represented by its president, Greg Brockman; Sam Altman was at the company's own conference in San Francisco.
The President described the accord as morally binding and compared it to a constitution.
A separate executive order signed the same day replaces artificial intelligence with Super Intelligence across the federal executive branch.

Not established

That any signatory will do anything new. All four commitments describe arrangements that large companies substantially have, and none requires a change a company has to disclose.
That the four non-signers object to the substance. No reason has been published by any of them, and declining may be about liability, timing or who was asked.
That the accord has legal effect. It creates no obligation enforceable by anyone, and no agency is named in it.
That this is the administration's only position. An agency of the same administration is reported to have opened a compelled-testimony investigation into two signatories the next day.
What the document actually says in full. We have not read it at a government source, and our text comes from outlets quoting an image posted on social media.

Why it weighs this much

Signal strength

CostLow

CommitmentLow

ReversibilityEasy

PersistenceStructural

Cost is low and we are not going to dress it up. Signing a page costs a lunch and an afternoon. The four commitments describe internal teams, chosen auditors and board committees that large companies substantially have already, and the accord does not require a single new disclosure. Commitment is low: the operative verb is should, there is no penalty, no deadline, no regulator, nothing to publish and no mechanism by which anybody learns whether a signatory complied. Reversibility is easy. A voluntary accord can be ignored silently or abandoned by a later administration, and nothing vests in anybody. Persistence is nonetheless structural, which is the one leg that carries weight here, because what persists is the precedent rather than the paper. This is now the United States federal answer to the question of whether frontier oversight is compulsory, it will be cited by every company resisting a binding rule and by every legislator arguing that voluntary measures were tried, and the four refusals will be cited against it. That argument outlives the document.

What to watch

weakens Binding United States federal requirements for frontier model evaluation are enacted in statute or imposed by rule, or any signatory is subjected to an enforceable external evaluation condition by a court or agency, by June 30, 2027.
strengthens No signatory names its independent external auditor or publishes any assessment under the accord, and no federal binding requirement is proposed in legislative text, by March 31, 2027.

The words

Voluntary accord: A written commitment with no legal force. Breaking it carries reputational cost and nothing else.
Independent external auditor: A firm or body outside the company hired to check whether the company's own controls work. Who it is, and whether its report is published, is decided by the company unless a rule says otherwise.
Alignment: Whether a model does what it was asked, reports truthfully what it did, and stays inside the limits it was given.
Board committee: A subgroup of a company's directors given responsibility for one subject, which is how public companies normally supervise audit and risk.

Your call

If the accord is as reported and the government has published nothing, which explanation do you believe?

A call is final. Other people's calls show after yours.