Grading Each Other's Homework

On 29 September 2026, at the launch of a government chatbot called America.gov, Elon Musk described the document he had just signed. “We agreed to a number of things that include joint monitoring, board special committees, just generally grading each other’s homework,” he said, according to Nextgov.

It is the most accurate summary of the deal anyone has offered, and it came from the man who signed it.

The document

The “White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities” runs two pages. It was signed by the President, Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Musk of SpaceXAI, and Jensen Huang of NVIDIA. Every American frontier lab put a name on it, including the one the Pentagon had blacklisted six months earlier and the one whose owner sat at the President’s left at lunch, per the seating chart the President posted.

It commits each company to four layers, per the copy Nextgov reported: “robust internal controls to monitor the capabilities and alignment of its models during training and deployment around areas like cybersecurity,” an internal team to oversee those controls, partnerships with independent external auditors, and an independent board committee to read the reports. The last paragraph begins: “Over time, it may make sense to codify these steps into laws or regulations.” The next sentence: “Regardless of whether this is required of companies, we believe that implementing these controls and audits is critical to ensuring a safe future for everyone, and each of our companies are committed to doing this.”

Asked whether any of it was binding, the President said it was “morally binding.” He also said: “I think I’m seeing tremendous self-policing. And they understand that they have to self-police.” He promised about ten people to “watch over the whole enterprise” and an overseer named “in coming days.” No overseer had been named by 3 October.

The same afternoon he signed Executive Order 14434, under which the executive branch “will not acknowledge the usage of ‘Artificial Intelligence’ and ‘AI’ in any applicable setting.” The technology was renamed the same day it was handed back to the people building it.

The plan it was copied from

Seventeen days earlier, on 12 September, Amodei published “We Must Pace the Frontier.” Its first step was embedded evaluators: a third-party team “such as METR” with “ongoing, employee-like access,” whose job is “to verify adherence to safety practices and commitments, report incidents, and help assess the alignment” of models and training pipelines. Anthropic committed to it unilaterally and listed the furniture: “Desks in our offices, access badges, and company laptops.” The reviewers would have the right to publish “without editorial control by Anthropic,” and the company wrote that it “can’t redact findings just because they are unfavorable.”

Sam Altman answered the same day: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.” Musk posted: “Dario is right.”

Lay the accord beside the essay. Internal controls on training and deployment. External evaluators. A board that reads what they find. The accord is the essay’s first step, signed by everyone.

What did not make the trip is the next paragraph of the essay: “The most effective method of pacing is via regulation that targets all US frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily.” Nor did the framework Anthropic had published in July, which asked that governments “be able to verify companies’ safety claims, enforce safe practices through civil penalties, and ultimately have a way to slow or block the deployment of AI models that pose a serious risk of catastrophic harm,” and called itself “the strongest policy proposal from any frontier lab or policymaker to date.”

The substance went in. The enforcement stayed out. “Morally binding” is what a commitment is called when the clause that covers “those who are unwilling” has been removed.

Each party, in its own words

The case for the accord as written was made by the Vice President, and it is not a stupid one. “The solution to some of the AI risks is for you guys to take the risk seriously, not to come to the government for a regulatory regime that may make things worse if it’s not smart and careful,” JD Vance told the executives. He rejected an FDA- or FAA-style body because “most bureaucrats just know way less about this than the people who are actually building these products,” and said the FTC and the Justice Department already have the authority to reach products that harm consumers.

Zuckerberg called it “a start and an accord that the whole industry can come to.” Pichai called it “historic and consequential.” Amodei, who had asked for a statute and received a pledge, said “the mechanism, how we address those risks is still under discussion,” and, on the way out, “If we do this right, if we work with the president and everyone here we can win safely.”

The critics split along a line that does not match any party. Bradley Tusk of Tusk Ventures told CNBC the CEOs asking to be regulated were not converts to state power: “They want something that creates an equalizer for each other.” Alex Pascal of the Berkman Klein Center wanted “robust legal liability, regulation and fundamentally changing the race dynamics.” Senator Mark Warner, introducing a bill for “rigorous testing and evaluation” before deployment, put it as a question: “The president’s response? To rename it and tell the companies developing it to regulate themselves.”

One reading says the incumbents wanted a rule that binds their rivals. The other says the incumbents were let off. Both readings point at the same two pages.

The sticks stayed in the drawer, not out of the house

Voluntary did not mean unenforced. It meant enforced by other instruments. Two weeks before the lunch the President had posted that his administration already had “tremendous CRIMINAL and REGULATORY power” over these companies. The day after the signing, the Federal Trade Commission opened a broad probe into the safety of AI systems, “including Anthropic and OpenAI,” on allegations of unfair or deceptive acts. It is an investigation. No complaint, no finding. No report has named SpaceXAI, Google or Meta as a target.

So the arrangement has a shape. The labs supply the controls. The White House supplies the signature page. Enforcement arrives, when it arrives, through procurement, the FTC and Truth Social, aimed by the same office that wrote “morally binding.”

The grader with the lowest mark

The day before the lunch, a chart circulating on X ranked Musk’s Grok 4.6 and 4.7 three tiers below the best models from Anthropic and OpenAI. He called it “accurate,” with the caveat “for now.” Days earlier he had said SpaceXAI would reach “pole position in about 6 months.”

The lab promising to pass the leaders signed the leaders’ plan, endorsed its author in four words, and named the arrangement accurately. The lab that wrote the plan signed it too, without the part it had said mattered most. Every signatory got something. Nobody got a referee.

Six companies agreed to grade each other’s homework, and the only teacher in the room called it morally binding.

Click.

The receipts (free, on this site): US CAISI · Dario Amodei · the universal capture mechanism