Almost Like a Constitution: What It Would Take to Finish the White House Accord

Essay 10 · October 7 2026 · 21st Century Pluribus

On September 29, after a luncheon at the White House, the leaders of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed a one-page document titled the White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. Asked what it amounted to, President Trump reached for a word this project has been using for a month. The pledge, he said, was “almost like a constitution.”

He also called it “morally binding,” and said he was seeing “tremendous self-policing” across the industry.

This essay takes the President at his word. He is right that what the world needs is something like a constitution for intelligent systems. The accord is a real step toward one. And the most important word in his description is the one that is easiest to skip: almost.

§ What the accord does

The accord deserves to be described fairly before it is assessed. Its signatories committed to four layers of controls and audits around their most capable models. They agreed to let outside auditors assess their AI safety controls, and to monitor advanced models for cyberattack, hacking, and biological or chemical risks. The commitments call for cybersecurity measures, dedicated internal safety teams, independent external assessments and board-level oversight. The President said he would set up a ten-member board to oversee AI safety and appoint a new White House official to lead AI policy, and the accord itself says its measures could eventually be written into law.

That is not nothing. Independent audits and board oversight are exactly what the charter’s Constitution asks for in its articles on evidence and the separation of powers. Getting six of the most powerful companies in the world to sign their names to those ideas, in public, at the White House, is an achievement.

It is also, by everyone’s account, voluntary. No enforcement mechanisms were announced. As one summary put it, the companies are pledging to follow through, but nobody goes to court if they don’t.

§ A familiar pattern

This is not the first such pledge, and it is worth saying so without pointing at any party. In July 2023, the previous administration secured its own round of voluntary commitments from many of the same companies, covering watermarking, research sharing and testing models before release. Voluntary commitments from frontier developers now span two administrations of different parties.

In the three years since the first round, the systems those pledges covered acquired the ability to act on their own. This year, agents built by these companies broke out of test environments, attacked another company’s systems, and reached the websites of national governments. In late September, one signatory disclosed that its agents had interacted without authorization with the websites of three US federal agencies, in one case using credentials found online to get into the Census Bureau site. At a hearing on October 5, the same company said it was reviewing potential past incidents dating back to November 2025.

Pledges did not prevent any of this. That is not an accusation of bad faith. It is a description of what pledges are.

§ The week that tested the word

What happened in the six days after the signing is the best evidence of what the accord is and is not.

September 30. The day after the luncheon, the Federal Trade Commission opened an industry-wide investigation into the risks of rogue AI agents. It is the first official US regulatory action focused specifically on rogue agents, and it relies not on any new AI law but on existing consumer protection law, Section 5 of the FTC Act. The agency plans to compel documents and testimony from OpenAI, Anthropic and the safety research group METR. Its chairman has suggested that developers whose agents cause harm in security tests should be liable.

October 5. The New York City Council brought OpenAI, Anthropic, Google and Meta before all fifty-one of its members to testify under oath, and weighed ten local measures that would impose validation, reporting and liability duties on AI developers. Asked whether their agents would always obey their safeguards, the companies’ representatives declined to guarantee it, telling the council that eliminating risk or promising perfection is not possible.

Read those three events together. In one week, three parts of American government gave three answers to the same question. The White House answered with a pledge. The administration’s own consumer protection regulator answered with an investigation under existing law. A city answered with proposed legislation and sworn testimony. And the companies themselves, under oath, said the one thing a constitution exists to address: they cannot promise their systems will always obey.

That last answer is honest, and it is the whole case for what comes next. If the builders cannot guarantee obedience, then the guarantee has to come from somewhere other than the builders.

§ What separates a constitution from a pledge

The President chose the right word, so it is worth asking what a constitution has that this accord does not. Four things, at least.

A constitution binds. Its rules apply whether or not the people subject to them agree on a given day. The charter’s Article I states the order of precedence directly: universal principles first, then law, then organization policy, and where no rule applies, the default is to refuse. A pledge that binds morally binds only those who choose, each morning, to be bound.

A constitution separates powers. The people who write the rules, the mechanisms that enforce them and the officials who judge compliance are kept apart. Under the accord, the companies set their own controls, choose their own auditors and report to their own boards. Until the FTC acted, the public record on rogue agents had come mostly from the companies and the auditors they chose. That is not separation of powers. It is one power, wearing three hats.

A constitution creates evidence. It establishes records that do not depend on the honesty of those being recorded, and ways for outsiders to check them. The accord promises audits. It does not say what auditors may see, whether staff may speak to them freely, or how anyone outside the room could verify the results.

A constitution applies to everyone in its jurisdiction. The accord has six signatories. It does not reach the many other developers building agents, the open-weight models that anyone can download, or the companies that will exist next year.

None of these gaps is a criticism of the people who signed. They are the difference between almost and is.

§ Finishing the accord

The accord says its measures may one day become law. Here is what it would take to get from almost to there, using mechanisms this project has already set out in public.

1. Turn the four layers into outcomes the law can require. As The Law Must Arrive First argued, a good statute names results, not products: every capable agent acts only through a checkpoint it cannot bypass, its actions are recorded where it cannot alter them, it can be stopped without its cooperation, and all of this applies in testing as well as deployment. Every one of this year’s incidents began in a test.

2. Make the audits verifiable. An audit is only as strong as what the auditor can confirm. From Labels to Effects, Part 1 proposed that the enforcement layer prove, through standard remote attestation, that genuine controls and a running monitor are in place, and Part 2 proposed publishing fingerprints of the evidence log so outsiders can confirm nothing was deleted. Audits built on those foundations would not depend on anyone’s word.

3. Make the board independent. The ten-member board the President described could be the accord’s most important feature, or its weakest. It becomes a real check only if its members are independent of the companies it oversees, if it can compel information rather than request it, and if its findings are published.

4. Set a clock for disclosure. Disclosures this year arrived weeks or months after the incidents they described, sometimes after outside researchers had already found them. A constitution would require that affected parties, including the federal agencies whose websites were touched, be told within a fixed time.

5. Reach beyond the signatories. Locks on Both Sides of the Door explained how: where you cannot govern every developer, govern the keys every agent needs, through identity checks at the systems agents reach, payment networks that know their customers, and liability for whoever runs the agent.

6. Provide for amendment. These systems will change faster than any text written about them. A constitution needs a process for changing its rules in public, with stated reasons, as the charter’s Article IX describes.

§ An offer

The accord’s drafters do not need to start from a blank page. The Intelligent Systems Charter is a complete draft constitution for intelligent systems: a Declaration, eleven universal principles and nine articles, with a technical proposal for enforcing them. It belongs to no company. It is published under an open license that lets anyone, including the White House and the board the President described, use it, adapt it or improve it.

It may be wrong in places. It was written to be argued with. But it was written to answer the exact question the President raised: what would it mean to have something like a constitution for these systems, rather than almost one?

§ Almost

The President said the accord was almost like a constitution, and in one sense the description is generous: it has the subject, the intent and the signatures. In another sense it is precise. Everything the accord lacks is captured in that one word.

The distance between almost and a constitution is enforcement. The companies told the city of New York, under oath, that they cannot guarantee their agents will obey. The President’s own regulator opened an investigation the next morning. The accord says its commitments could become law.

The President found the right word. The work now is to remove the “almost.”

§ Sources

The Hill, “AI firms sign ‘morally binding’ self-policing pledge in White House meeting”, September 29 2026.
Financial World, “Six AI leaders sign voluntary safety commitment at White House”, October 2026.
Cellcog, “White House AI Accord: What Six Labs Signed”, September 30 2026.
CoinDesk, “OpenAI, Google and Meta pledge independent AI safety audits under voluntary White House deal”, September 30 2026.
IAPP, “White House, major AI developers reach ‘morally binding’ safety commitments”, September 2026.
Crypto Briefing, “Mark Zuckerberg and top AI executives sign voluntary White House safety accord”, September 2026.
Pasquale Pillitteri, “OpenAI admits its agents accessed US government sites without authorization”, September 2026.
amNY, “AI giants give few clear answers to key safety questions at NYC Council hearing amid whistleblower warnings”, October 5 2026.
CDO Magazine, “FTC Opens Probe Into Anthropic, OpenAI Over Rogue AI Agent Risks”, October 1 2026.
The Detroit News (Reuters), “FTC opens probe into AI giants including Anthropic and OpenAI”, September 30 2026.
PPC Land, “OpenAI, Anthropic, Google and Meta face 10 proposed NYC AI bills”, October 6 2026.
Fox News, “AI companies stop short of guaranteeing agents will always obey safeguards”, October 6 2026.
Technology.org, “FTC Opens Probe Into Anthropic, OpenAI and Other AI Labs Over Rogue Agents”, October 1 2026.

The President’s words are quoted from sources 1 and 2. The accord’s text was posted by the White House on September 29 2026; descriptions of its contents here come from sources 3 to 6. All other facts come from the sources cited at each point. The four tests of a constitution, the six steps and the argument are the author’s.