On September 29, local time in the US, Trump signed an artificial intelligence agreement with the heads of Google, Anthropic, Meta, OpenAI, xAI and Nvidia in the East Room of the White House, then posted the document to Truth Social. The document's full name is "The White House Superintelligence Accord: A Joint Commitment on Frontier Superintelligence Responsibility," and its text runs to a single page.
Among those present were Elon Musk, Mark Zuckerberg, Dario Amodei, Jensen Huang, Greg Brockman and Sundar Pichai; Microsoft's Satya Nadella and AMD's Lisa Su also attended. Pichai later confirmed on social media that Google had signed the agreement.
What the One Page Says
The accord requires signatories to build four layers of controls. The first is internal monitoring — continuously tracking capabilities such as cyberattacks, biological and chemical threats, and unauthorized system access during model training and deployment. The second is a dedicated internal team that verifies these controls work as intended and handles any issues found. The third is an independent external audit or assessment body. The fourth is an independent board-level committee overseeing the first three layers.
The document is framed as a "joint commitment" and a "statement of principles." Asked how binding it is, Trump said:
"I think it's morally binding."
He also described the document as "almost like a constitution." House Speaker Mike Johnson put it more plainly, saying the accord is voluntary for companies and that the industry will "regulate itself." The text leaves open that it "may become necessary" down the road to write these steps into law or regulation.
The same day, Trump signed an executive order requiring federal agencies to use the term "Super Intelligence" instead of "artificial intelligence" in official documents, giving agencies 60 days to draft the relevant definitions.
Criticism Zeroes In on What's Missing
UK tech outlet The Register went through the text line by line and noted it never defines what counts as a "robust" control, sets no audit frequency, and gives no assessment timeline. Democratic Senator Mark Warner called on Congress to pass legislation mandating testing, assessments and incident reporting, criticizing the president for merely "renaming things and letting companies police themselves."
The White House has not said whether the results of external audits will need to be submitted to the government or disclosed to the public.
From 2023 to Today: A Timeline of Attempts
Placed on a timeline, US oversight of frontier models has kept hovering near the "voluntary" line.
In July 2023, the Biden administration brought in Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI — seven companies — to sign a set of voluntary commitments that included pre-release safety testing and watermarking AI-generated content. There were no penalties attached then either.
In May this year, Trump had been set to sign an executive order requiring large models to undergo review before reaching the market; representatives from OpenAI, Anthropic and Google all showed up, but he called it off at the last minute, saying he "didn't like certain aspects" of it. The version he ultimately signed on June 2 changed course, letting companies voluntarily hand models to the federal government for testing 30 days before release.
What's different this time is the shift from "show the government" to "build the mechanism inside the company, then bring in a third party to check" — with the board now pulled into the loop. The six signatories cover the country's major closed-source model makers and its largest chip supplier; Microsoft and AMD were in the room, but public reporting so far has not listed them among the signatories. What these four layers actually look like once each company implements them won't be clear until the first round of external audit results comes in — and the accord gives no date for that.
Sources: CBS News, CocoLoop, The Register, Washington Examiner (full text of the accord), Lianhe Zaobao, Sundar Pichai's social media statement; the list of signatories and the four control layers follow the accord's text.