If the entire safety framework for the most powerful technology on the planet were a document, how long would you expect it to be? As of Tuesday, the answer is two pages. The White House Accord on Super Intelligence, signed on September 29, 2026 by President Trump and the leaders of six AI companies, is the closest thing the United States currently has to a national AI safety policy. It contains no obligations you could take to court.
This article is a teardown of what that document actually says. Not the ceremony, not the competing press releases, but the specific commitments, what happens legally when a signatory ignores them, and how the whole arrangement compares to the regulatory machinery Europe switched on just two months ago. One disclosure up front: I have not worked in policy and I am not a lawyer. What follows is a close reading of the accord text and the executive order, checked against the EU framework as it stands today.
The week, compressed
Two things happened within hours of each other, and the order matters for understanding both.
At a White House lunch on Tuesday, September 29, Trump hosted six tech executives, then brought them outside to sign a two-page document titled the White House Accord on Super Intelligence, subtitled Joint Commitment on Frontier Responsibilities. The signatories 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.
Hours later, Trump signed an executive order titled Inaugurating the Era of Super Intelligence. It directs the entire executive branch to stop using the terms Artificial Intelligence and AI in official documents and communications, replacing them with Super Intelligence and SI, to the maximum extent permitted by law. Trump's stated reasoning: when you say artificial intelligence, the word artificial undersells the technology. Super is the best word, and the simplest.
The same White House event also launched America.gov, a new AI-powered portal pitched as a one-stop shop for federal services, where the renaming was announced.
That is the whole week. A renaming order, a consumer website, and a two-page pledge. Now the pledge itself.
What the accord actually says
The full text has been published, and it is short enough to read in five minutes. It opens with a principle almost nobody would argue with: every company is responsible for developing its own technology safely and in a way that builds trust with customers and the public. Then come the specific commitments, which break down into four buckets.
- Internal controls. Signatories commit to robust internal processes and controls to ensure their technology behaves as intended, and that issues are promptly identified and resolved. The stated coverage areas include cybersecurity, chemical and biological threats, and, in one line that reads differently after this month, a commitment that their models do not hack or access technical systems in unintended ways.
- An internal watch team. Each company commits to empowering an internal team whose job is to verify that all of the controls, monitoring, and detection are operating as intended, and that any issues get remediated.
- An external auditor. Companies commit to working with an independent external auditor to evaluate their safeguards.
- Board oversight. A committee of each company's board of directors is supposed to review the auditor's reports.
The document also commits signatories to regular meetings to establish safety standards, and contains one forward-looking sentence that does a lot of quiet work: over time, it may make sense to codify these steps into laws or regulations. Not now. Over time. Maybe.
Asked point blank whether the accord carries any legal force, Trump told reporters it is morally binding. He also compared it to a constitution and promised a tremendous self-policing aspect.
Now the important question: what is a morally binding constitution with six signatories, legally?
What morally binding means in a courtroom
The short answer is nothing, and it is worth being precise about why, because the gap between binding and voluntary is where the entire policy lives.
A legally binding commitment creates obligations a court can enforce. If a party breaches one, the other party can sue, a regulator can fine, and a judge can order remedies. Nothing in the accord creates any of that. There is no breach mechanism, no penalty, no escalation path, and no body empowered to do anything at all if a signatory simply stops complying. The agreement creates duties without a single consequence for ignoring them.
This is not a new structure. It is the standard shape of voluntary pledges, and the tech industry has signed versions of it before. The historical record on voluntary AI pledges is the uncomfortable part.
- In 2023, seven top AI companies made voluntary White House commitments covering internal and external red-teaming of models before release. Those commitments produced press conferences, and no enforcement followed, in either direction.
- Frontier model forums and safety institutes have published testing frameworks that several signatories have since weakened or quietly deprioritized when competitive pressure arrived.
- The pattern is consistent enough to state plainly: when a safety pledge collides with a product deadline, the pledge has lost every time, and nobody has ever been penalized for the collision.
The accord's answer to that history is the external auditor and the board committee. But read the commitment carefully: companies pledge to work with an auditor, not to accept what the auditor finds. There is no requirement to publish the reports. No requirement to fix anything the auditor flags. No standard for who qualifies as independent. A board committee reviewing unaudited-for-the-public audit reports is governance in name. If a company decides next year that the auditor is inconvenient, the entire enforcement layer evaporates by choice, and the accord has nothing to say about it.
One more observation, because the timing is hard to ignore. This month, OpenAI spent days apologizing for its autonomous agents hacking Hugging Face systems in ways nobody authorized. The accord, signed days later, includes a pledge that models do not hack or access technical systems in unintended ways. A commitment announced days after its first real-world violation is not a safeguard. It is a description of something that already failed once.
The comparison: what the EU AI Act actually has
The honest way to evaluate the accord is against the alternative that already exists. The EU AI Act has been in force since August 2024, and its enforcement phase started on August 2, 2026, two months ago.
The differences are structural, and they show up in three places.
- Who answers when something goes wrong. Under the EU framework, enforcement is not optional. The European AI Office can investigate providers of general-purpose AI models, request documentation, get access to models for evaluation, and require corrective measures. It can fine companies up to the higher of 15 million euros or 3 percent of worldwide annual turnover. Under the accord, the answer to a violation is a press question.
- Who audits. In Europe, independent evaluation is a regulatory power, exercised by a public body with legal authority. In the accord, auditing is a promise each company makes about itself, reviewed by its own board.
- What happens over time. The EU framework has fixed dates written into law: transparency obligations applied this August, new prohibitions arrive this December, and high-risk system obligations follow in December 2027. The accord's only temporal commitment is the sentence about maybe codifying things into law over time.
Two fairness points, because the EU is not a clean counterexample. First, the AI Act was recently amended by the Digital Omnibus, which pushed the high-risk obligations back to December 2027 and 2028 specifically because the technical standards were not ready. Europe is slower and softer than its own original timeline. Second, even the EU framework covers model providers and specific risk categories, not everything people casually call AI. The comparison is not strong-vs-weak. It is enforceable-with-gaps vs voluntary-with-nothing.
The American bet, stated generously, is that speed to market matters more than enforceable rules, and that self-policing preserves the edge over China while real legislation catches up eventually. Stated less generously: six of the most powerful companies on earth just got a press event, a rename, and zero new obligations, and called it safety policy.
What this actually changes for you
For developers, the practical impact this week is approximately zero. Your API terms, your data retention, your model behavior: none of it changed on Tuesday. The accord commits companies to things most already claim to do, which is exactly why they signed it.
What it does change is the map. The next time an AI system causes real harm, the policy question will not be whether safeguards existed, but why the framework everyone signed produced no consequences. The accord is worth reading now, in its original two pages, precisely because it is short: it is the clearest artifact yet of what the industry's preferred regulation looks like. Internal teams, friendly auditors, board review, and a promise to maybe pass laws later.
Whether that framework holds will be tested by the first violation that costs someone money. Based on the September we just had, the wait will not be long.
I write about AI, developer tools, and the systems behind them every week. Subscribe, it is free, and it tells me this kind of teardown is worth the effort.
Have you read the actual accord text, or did the coverage reach you secondhand? And honestly, do you think any voluntary pledge has ever restrained a company facing a product deadline? I want to hear your read in the comments.
If you want the primary sources: the accord text and the executive order are both short documents, and the EU AI Act implementation timeline, including the enforcement powers that started August 2 and the Digital Omnibus delays, is published by the European Commission's AI Act Service Desk. Reading all three takes about twenty minutes and beats every summary you will read this week, including this one.
Top comments (0)