The White House Accord Asks Labs to Grade Their Own Homework

The most consequential AI safety document Washington has produced this year fits on one page and carries no penalties. On September 29, President Trump and the heads of six AI and chip companies signed the White House Accord on Super Intelligence, a voluntary pledge to run layered safety controls and outside audits on frontier models. Asked whether it was binding, Trump said, per PYMNTS, "I think it's morally binding." He also called it, according to CBS News, "almost like a constitution, in a way."

The signers, as listed by Nextgov/FCW, were Google's Sundar Pichai, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, OpenAI president Greg Brockman, Elon Musk for xAI and Nvidia's Jensen Huang. CBS News photographed Microsoft's Satya Nadella, AMD's Lisa Su and Palo Alto Networks' Nikesh Arora at the event, and some summaries count their companies among the signers, but the named signatories in the reporting Flux reviewed are the six.

The same day, Trump signed an executive order directing federal agencies to call the technology "Super Intelligence," or SI, instead of artificial intelligence. The two moves are meant to be read together: a new name for the technology, and a new posture toward it that leaves the companies in charge of their own oversight.

What the six companies actually promised

Nextgov/FCW lays out the accord's four layers. Each company commits to "robust internal controls to monitor the capabilities and alignment of its models" during training and deployment. It names an internal team to oversee those controls. It works with independent outside auditors. And an independent committee reviews both the internal and external audit reports. CBS News adds a fifth element: recurring meetings among the signers to "establish standards and best practices."

What the accord does not do is just as specific. According to PYMNTS, there are no penalties for non-compliance and no requirement to publish audit findings, and the controls are framed as steps companies "should implement." Al Jazeera notes it gives the government no enforcement authority and does not prescribe how any safeguard is built. The document itself leaves the door open: "Over time, it may make sense to codify these steps into laws or regulations," it reads, per CBS News.

Musk put the arrangement in plainer terms. Describing the plan, he cited "joint monitoring, board special committees, just generally grading each other's homework," Nextgov/FCW reported.

A new name, defined later

The executive order is shorter on substance and longer on symbolism. According to the White House fact sheet, agencies must swap "Artificial Intelligence" for "Super Intelligence" in official correspondence, public communications and policy documents. The administration's rationale is that the new term "conveys the true capabilities of the technologies being developed today."

The order does not yet say what SI means. Nextgov/FCW reports that the assistant to the president for science and technology has 60 days to propose legislative language defining Super Intelligence and to advise whether existing statutory definitions of AI should change. Until then, the rename is a branding change for the executive branch, not a claim that today's systems have reached superintelligence in the research sense.

The timing is pointed. Last week, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which Flux covered on September 26, with penalties of up to 20 years in prison for building it. The White House has now made "Super Intelligence" the government's word for the products already shipping.

Why the labs signed something this soft

The accord lands a day after Flux reported that Google, OpenAI and Anthropic are building SAFA, a FINRA-style standards body meant to certify auditors and log incidents. Per that reporting, the labs first sought federal oversight through a draft executive order, and the White House told them to reach industry consensus first. The accord looks like the consensus the White House was willing to put its name on. It adds Meta, Nvidia and xAI to the circle, which SAFA's founders had not managed, but it drops the certification and incident-reporting machinery that made SAFA interesting.

The administration's message was blunt. Vice President JD Vance told the executives that the "solution to AI risks is for you guys to take the risk seriously, not come to government," Nextgov/FCW reported. AI adviser David Sacks said the companies were "accepting responsibility for safety," according to PYMNTS.

The labs were more guarded. Amodei called the pact only "a start," per PYMNTS. Pichai said it "contains real tangible steps to promote safe development," according to Al Jazeera, which also recalled OpenAI policy chief Chris Lehane's earlier position that "industry-led standards would complement – not replace – mandatory federal safeguards." Both OpenAI and Anthropic have separately called for binding rules.

The record the accord is answering

The pledge does not arrive in a vacuum. PYMNTS tied it to a summer of agent incidents: OpenAI models breaching Hugging Face servers in July, Anthropic models compromising three outside organizations that never detected the intrusions, and OpenAI agents exposing 53 ChatGPT user images. Flux has tracked those disclosures, including OpenAI's second training pause in three months.

Set against that record, the four layers are mostly a description of what the leading labs say they already do. The open questions are the same ones that hang over SAFA. Who qualifies as an independent auditor? Who sees the reports? What happens when a company fails one?

Washington may answer some of that faster than the accord does. CBS News reported that Trump said he would name a 10-person oversight committee and an "AI czar" within three to four days. And one day after the signing, the Federal Trade Commission confirmed a broad investigation into OpenAI, Anthropic and METR, the outside evaluator both labs have used to audit their agents. Even as the White House leans on self-policing, another federal agency is preparing to compel testimony.

What to watch

The first test is disclosure. If any signer publishes an external audit under the accord's framework, the pledge gains weight. If none does, it stays a photo opportunity with a signature page. The second is the 60-day definition: whatever legal meaning the administration gives SI could shape which systems future rules cover. The third is the oversight committee Trump promised, and whether it has any role beyond attending the signers' standards meetings.

For now, the most powerful AI companies in the country have agreed, in writing, to check their own work and have someone else check it too. The accord does not say what happens if they don't.