By JD DuRie, in collaboration with AI Matters
On September 29, President Trump invited six of the most powerful people in technology to lunch. Afterward he took them outside to sign a document. Present were Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of xAI, and Jensen Huang of Nvidia. The document was titled the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. The President called it “morally binding.” The next day he signed an executive order directing federal agencies to call artificial intelligence “super intelligence.” The order says the new term covers exactly the same technologies as the old one.
It is tempting to dismiss the whole affair as theater. That would be a mistake. The accord is worth reading closely, because what it leaves out tells us more than what it says.
What it says
The text runs about 300 words. It asks each company building frontier AI models to adopt four layers of safeguards:
- internal controls to monitor what their systems can do, with attention to cybersecurity, biological and chemical threats, and models that “hack or access technical systems in unintended ways”;
- an internal team to confirm those controls actually work;
- an independent outside auditor or evaluator;
- an independent committee of the board of directors to receive the reports and make sure problems get fixed.
Anyone who has sat on a corporate board will recognize the structure. It borrows from how public companies police their financial statements. As governance design, it is sensible. If every frontier lab truly built all four layers and staffed them with people willing to say no, the world would be somewhat safer than it is today.
What it leaves out
The accord has no enforcement mechanism, no penalties, no funding, and no timeline. It does not say who the auditors will be, what they must test, or whether anyone outside the company will ever see their findings. It creates no regulator. It sets no definition of “safe.” Instead, it promises that the participating companies “will meet regularly to establish standards and best practices.” Those are the same companies whose products are being judged.
It also offers the companies no legal protection. There is no indemnification, no safe harbor, and no preemption of state law. Their legal exposure is exactly what it was the day before. But the accord ends by suggesting that “over time, it may make sense to codify these steps into laws or regulations.” House Speaker Mike Johnson has already floated it as a basis for legislation. The real question is what that future law would contain. A statute that turns these four voluntary steps into a liability shield or a ceiling on state regulation would be a very different thing from the accord itself.
We have seen this before. In July 2023 the Biden administration announced voluntary AI commitments from seven companies, several of them the same firms. Critics said then that no one could hold the companies to their word. Three years later, the record largely proved them right.
Who answers when something goes wrong?
Under this accord, the honest answer is: the companies answer to themselves. The board committee at the top of the structure is accountable to shareholders, not to the public. Citizens have no seat at the table, no access to the audits, and no remedy beyond the ordinary courts, which existed before the accord.
A compassionate libertarian reading
Those of us who distrust concentrated power usually aim our suspicion at the state. It is fair to note that this accord is not a government power grab. It imposes no new mandates and builds no new bureaucracy. Some readers will count that as a virtue.
But the libertarian case against concentrated power has never been only about government. It is about any arrangement in which a small number of actors hold great power while answering to no one. The accord creates exactly that arrangement and puts the White House seal on it. It is not the free market at work, because the public cannot comparison-shop between AI labs on catastrophic-risk safeguards it cannot see. It is not distributed, voluntary coordination among equals. It is six firms that together control most of the world’s frontier AI capability, certifying their own conduct, with the President as witness.
A genuinely liberty-minded framework would push power outward. It would require audit results to be made public, so customers, insurers, and investors can judge for themselves. It would let independent evaluators, not the companies alone, set testing standards. It would preserve the right of states, including Connecticut, which has debated its own AI legislation in recent sessions, to experiment with their own rules rather than having a voluntary federal pledge used as an argument to silence them. It would make whistleblower protection for AI workers explicit. None of this requires a new federal agency. All of it requires transparency, which the accord does not provide.
The case for the defense
To be fair, the accord has defenders, and their arguments deserve a hearing. Binding regulation of a fast-moving technology can freeze today’s assumptions into law and favor incumbents who can afford compliance departments. A voluntary framework can be revised as the science improves. Public commitments also carry some weight: a company that signed this document and then skipped the audits has handed future plaintiffs a ready-made exhibit. And not every signatory wants the same thing. Anthropic, for one, has publicly supported real safety legislation, while others in the room have lobbied against it. For some signatories the accord may be a floor, not a ceiling.
Those are reasonable points. But a floor that no one inspects is only a promise, and the public has been asked to accept such promises before.
The bottom line
The accord describes good practices. The companies deserve credit if they follow them. But in its current form it gives the most powerful firms of our time a government blessing, costs them almost nothing, and gives the rest of us nothing we can enforce. Before anyone turns it into law, citizens should insist on three things: audits we can read, auditors the companies do not choose alone, and no liability shields bought with voluntary pledges.
This post was written in collaboration with AI Matters, an editorial process combining human judgment with AI-assisted research and drafting. Disclosure: the AI system used in preparing this post is made by Anthropic, one of the accord’s signatories.
