Four Layers of Controls, Zero Deadlines: Reading the White House AI Accord

The accord is one page. I could not open the White House’s own copy, so I read the full text as reposted in a Substack dated 30 September, and checked its wording against the quotations in Medianama’s report. They agree. The accord is titled “White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities,” and Euronews lists the signing companies as Anthropic, Google, Meta, OpenAI, Nvidia and Musk’s group (it describes Musk as head of Tesla, SpaceX, X and xAI). Medianama dates the signing to 29 September.

Most of what I have seen about it is reaction. I wanted to do the dull thing first and read the verbs.

what the page asks for

The preamble says “every company is responsible for developing its own technology safely.” The body then says each company “should implement the following four layers of controls and audits.” Note the auxiliary verb: the layers are introduced by belief, and only the closing paragraph says the companies are committed. Here are the layers, abridged but in order.

  1. “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, and to ensure that its models do not hack or access technical systems in unintended ways.”
  2. “Empower an internal team to ensure that all of the controls, monitoring, and detection are operating as intended, and that any issues are remediated.”
  3. “Partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended.”
  4. “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, as well as to ensure any issues identified are remediated.”

A fifth sentence says the participating companies “will meet regularly to establish standards and best practices.” Another says that “over time, it may make sense to codify these steps into laws or regulations,” and that the companies will do this regardless.

So I built a table of my own. For each of the five operative items I asked four questions of the text: does it name a test or threshold, a checker, a date, a consequence? A full mark means the text says it. An open ring means the text gestures at it, with a topic, a role or a word, but no standard.

What each part of the 29 September accord names: a test, a checker, a date, a consequence A matrix of five accord items by four properties. Internal controls: test gestured at, checker gestured at, no date, no consequence. Internal team: test gestured at, checker role given, no date, consequence gestured at. External auditor: test gestured at, checker role given, no firm named, no date, no consequence. Board committee: no test, checker gestured at, no date, consequence gestured at. Meet regularly: no test, checker gestured at, no date, no consequence. In total two of twenty cells are fully stated, eight are gestured at, and ten are absent. No item has a date. WHAT THE ACCORD NAMES · FIVE ITEMS × FOUR PROPERTIES · MY READING OF THE TEXT A test orthreshold A checkerrole A date A statedconsequence 1 Internal controls 2 Internal team 3 External auditor 4 Board committee 5 Meet regularly monitor cyber, bio, chemical, unintended hacking checks the controls work; issues remediated independent assessments; no firm named receives reports; ensures remediation set standards and best practices Names risk areas, no test or threshold — accord text "Internal controls," the company itself — accord text "Operating as intended," no standard — accord text An internal team — accord text The word "remediated," no penalty — accord text "Operating as intended," no standard — accord text An independent external auditor or evaluator, unnamed — accord text A board committee that "oversees" — accord text "Ensure any issues identified are remediated," no penalty — accord text "The participating companies," meeting regularly — accord text stated in the text gestured at, no standard absent
Two filled cells out of twenty, and the date column is empty on every row. The table is my reading of a one-page text, and "gestured at" is a judgment call; but no reading turns a blank date into a date.

Including the fifth item, the count is two checker roles, eight gestures, ten blanks, and zero dates. The accord names who looks, by role (an internal team, an external auditor) and stops before the harder nouns: which test, by when, with what result if it fails. The word “remediated” appears twice, and apart from a preamble phrase about issues being “promptly identified and resolved,” it is the only thing that happens to a finding. Those are cadence words, not dates.

what went missing since 2023

For comparison, the White House’s fact sheet of 21 July 2023 listed eight commitments, under three headings, from seven companies, and used the verb “commit”: to “internal and external security testing of their AI systems before their release,” and to “publicly reporting their AI systems’ capabilities, limitations, and areas of appropriate and inappropriate use.” That was background three years old, and I use it only as a yardstick. It was also voluntary, and it also named no enforcer.

The 2026 text has no public-reporting sentence. Medianama’s summary says the accord has no requirement to publish audit results, name the auditors or report incidents to a government body, and in the reproduced text I read I found none either. The 2023 pledge had at least one output a stranger could read. This one has a committee that receives reports, and the accord text promises none to anyone outside the company.

Who signed in 2023, in 2026, or both Three columns. Signed in 2023 only: Amazon, Inflection, Microsoft. Signed both times: Anthropic, Google, Meta, OpenAI. Signed in 2026 only: Nvidia and xAI. The 2023 fact sheet listed eight commitments under three headings; the 2026 accord has four layers of controls plus a clause to meet regularly. SIGNERS · WHITE HOUSE FACT SHEET 21 JUL 2023 · SIGNING REPORTS 29–30 SEP 2026 2023 only · 3 both times · 4 2026 only · 2 Amazon Inflection Microsoft Anthropic Google Meta OpenAI Nvidia xAI (Musk) 2023eight commitments, under three headings 2026four layers of controls, plus one clause to meet regularly The Substack signature block reads "SpaceXAI"; Euronews describes him as head of Tesla, SpaceX, X and xAI. No Amazon or Microsoft name appears in the 2026 signature list.
Four firms carried over. Three left the room, two came in, and the list of obligations got shorter. Whether Amazon and Microsoft were ever asked is not something I found.

The chain has a shape. The company builds the controls, the company’s internal team checks them, a firm the company partners with checks the team, and a board committee the company designates receives the results. The accord does not make the last step public or attach it to anyone outside. That is a design choice, and plausibly a defensible one for a first round of a voluntary regime. It resembles, by analogy, a concern raised in a recent preprint about governing AI systems: Tony Rost’s April 2026 paper argues, as hypotheses and not findings, that at the high end of capability asymmetry “the governed system either games its own evaluation or sits inside the governance process,” and that transparency remedies then “lose traction.” I am not saying that is what happens here. I am saying the text contains nothing that would let an outsider tell.

The outside record on the signatories’ earlier pledges is also not reassuring, though it is one grader’s view. The Future of Life Institute’s Summer 2026 AI Safety Index, published in July, graded Anthropic C+, OpenAI and Google DeepMind C, Meta D+ and xAI F, and wrote that “Anthropic, OpenAI, Google DeepMind, and Meta have weakened or voided pledges to pause unilaterally if redlines are approached, some citing competitor-contingent conditions.” I have not audited the index’s method, and the pledges it describes are the companies’ own frameworks, not the 2023 commitments. But it is the nearest thing to a track record, and it points one way.

One more detail, which is the one I like best. Anthropic’s Transparency Hub, last updated 23 July 2026, says of the 2023 commitments: “Though these specific commitments are no longer formally maintained under the Trump administration, our organization continues to uphold all these principles.” That is the status of the predecessor, a little over two months before the new accord: per Anthropic, the commitments are no longer formally maintained, and the company says it kept the principles anyway. In the accord text, the company’s own continued practice is likewise the enforcement mechanism, and the company is the one describing it.

what was said at the signing

The President, per Euronews, said “I think I’m seeing tremendous self-policing. And they understand that they have to self-police,” and described the accord as “morally binding.” Medianama reports that no law or regulator enforces the accord and that it carries no penalty. Euronews says Trump suggested about ten people would oversee the enterprise, and Medianama says a board will oversee AI safety, with no members named. That is an announcement, not accord text, and I have no detail on it. The text itself says codifying these steps into law “may make sense” over time, and gives no date for when that might be decided. In an earlier essay I argued that renaming AI “Super Intelligence” changed no definition in force. The accord’s title carries the new name, and its text still defines nothing a regulator could measure against.

What would make it checkable is short. The name of one auditor. The number of issues remediated. The date by which a first report reached a board. The text as it stands produces none of the three, and I will know more about what it means if a signatory volunteers them.

References

  1. White House (2026). White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. Full text as reproduced in Messier, D., “White House Accord on AI”. Substack, 30 September 2026. The White House’s own copy was not readable when checked.
  2. Medianama (2026). Trump, Google, Anthropic, Meta, OpenAI sign AI safety accord. September 2026; reports a 29 September signing.
  3. Euronews (2026). Top AI and tech firms sign “morally binding” accord to self-police development after meeting. 30 September 2026.
  4. The White House (2023). Fact Sheet: Biden-Harris Administration Secures Voluntary Commitments from Leading Artificial Intelligence Companies to Manage the Risks Posed by AI. 21 July 2023 (archived).
  5. Anthropic (2026). Voluntary commitments. Transparency Hub, last updated 23 July 2026.
  6. Future of Life Institute (2026). AI Safety Index, Summer 2026. July 2026.
  7. Rost, T. (2026). From Disclosure to Self-Referential Opacity: Six Dimensions of Strain in Current AI Governance. arXiv:2604.14070, 15 April 2026.
  8. Gautam Parab (2026). Renaming AI “Super Intelligence” Changes No Definition in Force. 23 September 2026.