On September 29, the leaders of six major AI and technology companies sat down to lunch at the White House and signed a one-page document with the president. Google, Anthropic, Meta, Nvidia, OpenAI, and X. The document is titled the "White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities." Trump posted it on Truth Social. He called it "almost like a Constitution" and "morally binding."
It is not a law. It is not an executive order. It creates no federal legal requirement and no penalty for ignoring it. "Morally binding" is the president's phrase, not a legal category. That matters, because everything else about this story depends on understanding what the paper actually is: a voluntary promise, signed in public, with no enforcement behind it.
The promise itself has four parts. Each company commits to robust internal controls that monitor what its models can do during training and deployment, watching for cybersecurity, biosecurity, and chemical threats, and making sure models do not hack or access systems in unintended ways. An internal team makes sure those controls actually work. An outside auditor independently checks the work. And a committee of the company's board of directors reviews the reports and makes sure problems get fixed.
Credit where it is due: the substance is sound. This is the monitoring and detection approach. Watch the systems. Keep records. Verify. The argument this blog exists to make: the answer to rogue agents is tools, not bans. When the biggest labs in the world sign a document saying the fix is internal controls and independent audits, they are agreeing with the builders. I will take the win.
And to be clear, the controls themselves are not the enemy. Monitoring a model for unintended cyber access, testing before deployment, having someone check your work, these are good engineering practices. The fight is not over whether to do them. It is over who defines the threshold, what compliance costs, and what happens when the checklist becomes law.
The companies also agreed to "meet regularly to establish standards and best practices." The accord does not say how transparent those meetings will be, who else gets a seat, or how those standards would eventually become law. And then comes the sentence that made me put down my coffee: "Over time, it may make sense to codify these steps into laws or regulations."
This is not the only room where this is happening. The Information reported on September 24 that Google, OpenAI, and Anthropic have been finalizing plans for a private Standards Authority for Frontier AI, modeled on Wall Street's FINRA, with a possible launch in late 2026 or early 2027. OpenAI confirmed the talks. According to the reporting, the original idea was a public-private partnership with federal oversight, and the plan shifted to industry self-regulation after a draft White House executive order stalled. Critics are already warning that if the largest vendors set the accreditation and audit rules, compliance costs could squeeze out smaller and open-source developers. Cohere's CEO reportedly called it "a cartel by any other name." That is his characterization, not mine, but it tells you the industry itself is not united on this.
Now my theory, labeled as such: this is how a moat gets built without anyone voting on it. The concern is not that anyone will explicitly ban open weights. It is that if these requirements become law without size-based exemptions or a low-cost compliance path, the price of external audits and corporate oversight becomes a wall. For Google, those are rounding errors. For a two-person shop training a model in a garage, or an open-source collective with no board of directors at all, they are the end of the road.
I want to be fair, because the other side deserves its say. A voluntary accord is not a federal licensing scheme. Nobody's model was seized. Nobody needs permission to train. The same day, the president said he will never stifle an industry he says will be bigger than the Industrial Revolution, and a week earlier he told the UN he rejects any global scheme to control AI. If the choice is between this piece of paper and a federal agency that licenses who may build, I will take the paper. And the accord itself says the companies are committing "regardless of whether this is required," which undercuts the pure cynicism read. Some of this may be genuine.
But here are the questions nobody at the lunch answered, because nobody at the lunch was in a position to ask them.
Who audits the auditors? Who selects them? Who pays them? What qualifies them, and who decides? The accord does not say. That does not make the audits fake, but it is worth saying out loud.
What happens to the builder with no board? The accord assumes a company with directors, compliance teams, and auditors on retainer. Most open-source AI work has none of those things. When the voluntary standard becomes the legal standard, what is the on-ramp for the project with three contributors and no legal department?
And what becomes of the committee? The president also floated a possible oversight committee of "maybe 10 people" to "watch over the whole enterprise." Ten people, chosen by whom, accountable to whom? That is not a detail. That is the whole question.
Scan the signatures one more time. Every signature belongs to a major corporate player. The open-source community, the home labs, the independent researchers, and the small AI companies are not represented among the signatories. I am not saying the people in the room are villains. I am saying the room was incomplete.
I am not against regulations. I am against regulations only the big guys can afford. A voluntary promise among giants is fine as far as it goes. But when the promise comes with a plan to become law, written through a process the rest of us have no visibility into, the small builder should be paying attention.
The accord is morally binding. The question is: on whom?
