​​​​

U.S. President Donald Trump holds a press briefing with AI executives following a meeting on artificial intelligence, outside the West Wing of the White House on September 29, 2026 in Washington, DC. Kevin Dietsch/Getty Images

 

30, September, 2026

​​​​

No Border Clause

The White House AI accord, 29 September 2026.

 

---

 

The accord signed at the White House this week is a commitment by leading American AI companies, made alongside an American president.

 

While it has no international coordination mechanism and no cross-border provision of any kind, that does not confine it's implications. A standard without a treaty behind it travels through products, contracts and procurement instead. It reaches operators in frontier markets as a condition of access, already written, and it will be put to them as a question from a counterparty long before it is put to them as law.

 

---

​

What was signed

On 29 September, Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a Joint Commitment on Frontier Responsibilities. It is voluntary. It was described at the signing as "morally binding."

 

It has four layers. Internal controls that monitor model capability and alignment through training and deployment. Internal teams accountable for those controls. Independent external auditors who verify the monitoring works. A board-level committee that reviews the audit reports.

​

The text says these measures may make sense to write into law. Naturally, it names no standard for the auditors and no enforcement mechanism, as these will take extensive work to develop over time.

​

The same day, an executive order directed federal agencies to say "Super Intelligence" in place of "AI." It keeps the existing statutory definition pending revision, and gives the President's science adviser 60 days, to 28 November 2026, to propose legislative language on whether that definition should change. It contains no provision on international coordination either.

 

---

 

You might obsverve that this design has a precedent

Management controls. An accountable internal function. An external auditor. An audit committee of the board.

 

That is the assurance architecture of financial reporting after Sarbanes-Oxley, applied to model safety. The precedent is twenty years old, which means the behaviour of the design is not a matter of speculation.

 

The US Government Accountability Office finds that auditor attestation carries some monitoring value, and that smaller companies bore higher Section 404 compliance costs as a share of their assets than larger ones. Congress exempted the smallest issuers from the attestation requirement, and the SEC widened the exemption in 2020.

 

Three things carry over. The audit tests whether a control existed and was followed, which is a question about process. An industry forms to supply the assurance, paid by the entities it assures. And fixed compliance costs select for scale: four layers of internal governance are a line item at Google and a structural problem at a twelve-person laboratory.

​

---

​

What the audit layer cannot yet distinguish

A measurement reform can improve the fidelity of an instrument, reading the real thing more accurately. Or it can improve the favourability of the output, producing better readings without better measurement. From outside the two look the same.

 

The accord does not separate them. A policy analyst quoted at the signing put the structure in one sentence: "The companies drafted the principles, they hire the auditor, and the commitment is voluntary."

 

That describes a design, not a motive. A reader at distance cannot verify which case they are looking at, and nothing in the text lets them find out.

 

---

 

Where the binding action is

Not at the signing. The Vice President said the intended instruments are the existing authorities of the FTC and the Department of Justice. The Speaker of the House said in mid-September that Congress should not panic into regulation. The bipartisan Great American AI Act has been a discussion draft since June, opposed by more than two hundred state legislators.

 

The live contest remains the one opened by last December's executive order: federal preemption against state law, argued in court. Colorado's rewritten AI Act takes effect on 1 January 2027. California's SB 53 is in force and in dispute.

 

The document to watch is the 28 November drafting exercise. A definition that stays broad reaches most deployed systems. A definition narrowed to frontier capability leaves most of them outside any federal rule. The order adopts the first today and commissions the second as a possibility.

 

---

 

How a standard without a border crosses over to other markets

It crosses through the people who must answer to it.

 

A bank, an insurer or a ministry that deploys a frontier system has to satisfy its own board, its own auditors and its own regulator that the dependency is controlled. So it asks the supplier for evidence. The supplier asks its suppliers. The provisions travel down the chain in terms of service, model documentation and procurement requirements, and no statute needs to name the company at the end of it.

 

Sarbanes-Oxley spread this way. Businesses that were never public issuers found themselves asked by customers to evidence their own controls, because the customer's auditor required it.

 

The European Union's deforestation regulation has the same shape. It falls due on 30 December 2026 for large and medium operators, after repeated postponement. It is written for operators in Europe, and it reaches producers elsewhere as a requirement to supply data about their own supply chains, which operators must collect and keep for five years.

 

No provision of the accord requires any of this. The mechanism is older than the accord and does not depend on anyone's intention.

 

The consequence falls on the operator in a frontier market, and it is narrow. Counterparties with obligations will ask for evidence: controls documented, data lineage traceable, a record of who did what and when. An entity that can produce it is contracted, integrated and financed. An entity that cannot is refused by nobody in particular. It is simply not legible to the party deciding.

 

​​