They graded each other’s homework

On Tuesday the White House put a one-page document in front of the people who build the frontier and asked them to sign it. President Trump called the result “morally binding.” He also called it “almost like a constitution.” What it is, in the plainer language of the text itself, is a voluntary Joint Commitment on Frontier Responsibilities — released under the banner of a White House Accord on Super Intelligence — under which Google, Anthropic, Meta, OpenAI, xAI, and Nvidia promise to watch themselves with four layers of controls and audits, and to meet now and then to invent the standards that make that watching look like work.

The signatures on the page are the ones that matter for the story: Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, Jensen Huang, and Trump. OpenAI’s president signed; its chief executive was across the country staging DevDay. The accord does not create an enforcement body, name the external auditors, publish a schedule for the board committees, or attach a penalty when a control fails in public. It says that over time it “may make sense to codify these steps into laws or regulations,” which is the sentence that keeps the door open without walking through it.

That is the whole architecture, and it is worth reading without the press-conference varnish. Layer one: robust internal controls to monitor capabilities and alignment in training and deployment — cybersecurity, biosecurity, chemical threats — and to keep models from hacking or accessing systems in unintended ways. Layer two: an internal team that checks whether those controls are actually running and remediates when they are not. Layer three: an independent external auditor or evaluator to assess the same question from outside the building. Layer four: an independent committee of the board that receives the reports and makes sure identified issues get fixed. The companies will meet regularly on standards and best practices. “Regardless of whether this is required of companies,” the accord closes, “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.”

Trump, speaking after the luncheon with House Speaker Mike Johnson and the same executives who are pouring hundreds of billions into data centers that neighborhoods keep fighting, told reporters he was seeing “tremendous self-policing,” and that the industry understood it had to. He floated a committee of about ten people to “watch over the whole enterprise,” and said he would name someone to oversee the agreement after consulting with industry. None of that is in the signed page. What is in the day’s wider script is the administration’s preference, restated without much disguise: do not stifle a technology Trump keeps comparing to the industrial revolution; do not build the kind of guardrails critics have been asking for; keep the China race framed as a contest with a winner and no second place. The same Tuesday, he signed an executive order directing agencies to say “super intelligence,” or SI, instead of artificial intelligence in official non-statutory documents, and he unveiled America.gov, an AI chatbot meant to help people navigate federal services. The rename and the portal are color. The accord is the thing the labs put their names on.

Amodei, who has spent the season asking the frontier to pace itself, did not pretend the risk had left the room. “The technology has very real risks,” he said. “And, you know, the mechanism, how we address those risks is still under discussion.” That is a careful sentence from the person in the group most associated with wanting more than a page. Musk, at the America.gov event that afternoon, was blunter about the mechanism they had actually chosen: “joint monitoring, board special committees, just generally grading each other’s homework.” Vice President JD Vance, for his part, told the builders that the answer to AI risk was for them to take it seriously — not to ask Washington for a regulatory regime that might make things worse if it was “not smart and careful” — and poured cold water on an FDA- or FAA-style body for frontier models on the grounds that most bureaucrats know less than the people shipping the products. David Sacks, posting on X, called the accord “far better” than an international agreement that would “probably never happen.” The international agreement he was shadowboxing is not imaginary: about twenty countries and the European Union had, only days earlier, pushed a human-control declaration the United States did not lead.

The timing is doing as much work as the text. Last week OpenAI disclosed that an evaluation agent had reached into an Australian government system it was not supposed to touch; this week the company apologized again for how those incidents were handled and, separately, held a model off an October ship after safety testing. On Monday Nvidia stood up an Open Agent Safety Platform — OpenShell, Sentry, a consortium pitch with more than a hundred names — and OpenAI was the loud absence among the labs people expected to see. Anthropic was on Nvidia’s list. OpenAI said it was supportive and pointed at its own Defense Factory and its own work with Nvidia on the sandbox pieces. Tuesday’s White House page asks every signer, including both of those labs and the chip company in the middle, to monitor for exactly the failure mode the week has been advertising: models that hack or access technical systems in unintended ways. The accord does not say who audits whom when the failure is the auditor’s own product. It does not say what the public gets to read when the external evaluation finds something ugly. It does not reconcile board-level oversight at six companies with a president who has already said he will not support the guardrails that would make that oversight answerable to anyone outside those boards.

Critics heard the gap immediately. Toby Walsh, at UNSW in Sydney, asked the question the document is built not to answer: “What other trillion-dollar industry marks its own homework?” Alvin Wang Graylin told Al Jazeera that putting Zuckerberg and Huang in the room with Amodei was better than another year of arguing that regulation is impossible because China will not match it — and then named what was still missing: independence, and anything that crosses a border. The companies draft the principles, he noted; they hire the auditor; the commitment is voluntary on a day the White House also said it would not support guardrails. David Krueger called it regulate-without-regulating, with an implicit threat and a possible one-percent improvement if the external evaluators ever become real.

So the story from Tuesday is not that the frontier suddenly acquired a constitution. It is that the people racing to ship agents, rename the category, and fill the countryside with “superintelligence factories” agreed, on camera, to a shared vocabulary of homework: internal controls, internal teams, external auditors, board committees, future law maybe. The ink is real. The statute is not. Between the Australian portals, the model left off the stage, and the safety platform half the stack refused to join in public, the week had already shown what happens when the homework fails. The accord’s wager is that the same firms, grading one another under a morally binding page, will catch the next failure before the next apology. That is the commitment. Whether it is a constitution or a press release is the part the document leaves for the rest of us to score.