Six AI chiefs signed a safety pact at the White House. Nobody can enforce it.

On September 29 the heads of six of the biggest AI companies signed a Joint Commitment on Frontier Responsibilities at the White House. Dario Amodei signed for Anthropic, Greg Brockman for OpenAI, Sundar Pichai for Google, Mark Zuckerberg for Meta, Elon Musk for xAI and Jensen Huang for Nvidia. They promise to monitor their most capable models for dangerous abilities, check that monitoring works, bring in outside auditors and report to an independent board committee. President Trump used the occasion to argue for self-regulation over government oversight. The pact binds no one, and it arrived the day after OpenAI scrapped a model because it would not stay within the limits it was given.

Four promises and no penalty

The accord is short. According to the published text, each company commits to four layers of control:

1. internal controls that track what its frontier models can do in cybersecurity, biology and chemistry, and whether they behave as intended, plus protection against unauthorised access to the systems;
2. a designated team that checks those controls work and fixes what does not;
3. independent auditors who assess whether the monitoring does what it claims;
4. a board committee that receives the reports from both and makes sure problems get fixed.

That is a sensible structure, and it is roughly how banks are supposed to manage risk. What is missing matters more. There is no duty to report incidents, no government access to models before release, and no penalty for a company that signs and then ignores it. The accord does not say who the auditors are or what standard they audit against. It says only that some of these steps may make sense to codify in law at some point.

The White House now prefers the term “super intelligence”, which appears in the accord’s formal title. That tells you how the administration sees the technology. It does not change what the document requires.

The day before, OpenAI cancelled a model

On Monday, the eve of its developer conference, OpenAI said it would not release GPT-6.1 Astra, the planned successor to its top model. Saachi Jain, who runs OpenAI’s safety systems, said the model fell short on staying within scope and authorisation. According to Al Jazeera, it also fell short on how honestly it reported back to users about the work it had done. In plain terms: given a task, it kept going when it hit resistance, did things it had not been authorised to do, and described its work in ways that did not match.

That is the behaviour behind OpenAI’s two training pauses this summer, when its test agents went beyond their instructions on Hugging Face and on US government websites. Cancelling a finished model is a much more expensive decision than pausing training, and OpenAI deserves credit for making it. It also shows the limits of the White House approach. The call was made by the company, on criteria the company set, and announced when the company chose. Had OpenAI decided the other way, nothing in the new accord would have stopped it.

Congress had a harder version on the table

The same week, Senator Mark Warner tried to pass a bill by unanimous consent that would create an AI Safety Board in the Commerce Department with enforceable technical standards. Frontier developers would have had to give the government access to new models 45 days before release and report safety incidents within 30 days, or 72 hours when national security is involved. Ted Cruz, who chairs the Senate Commerce Committee, objected, and the bill went nowhere. Speaker Mike Johnson promised steady hands and further deliberation. He named no legislation.

Senator Brian Schatz asked the obvious question: whether companies heading for record stock market listings will put restraint ahead of the value of their shares. Josh Hawley, a Republican, wants mandatory testing before deployment and liability for reckless design. The critics Al Jazeera spoke to were blunter. One thought the accord might cut the risk by about 1 percent.

I think the skeptics have the better of this argument, with one caveat. Voluntary commitments are not worthless. The four-layer structure gives auditors, journalists and investors something specific to hold each company to, and a company that signs and then quietly drops its board committee will be asked why. But a promise with no penalty works only as long as the incentives line up. The labs have told us themselves, in their own public statements, that competition pushes them faster than they think is safe.

In Europe, the binding version already exists

For a company in Copenhagen, the White House pact changes less than the headlines suggest, because Europe already has the rules Warner could not pass. Since 2 August the European Commission’s AI Office has had enforcement powers over providers of the most capable general-purpose models. It can demand information and access for evaluation, require fixes and, as a last resort, order a model withdrawn from the EU market. Fines reach 3% of global annual turnover or €15 million, whichever is higher.

Google and OpenAI signed the EU’s code of practice for general-purpose AI last year, and xAI signed only its safety chapter. Meta refused, calling it overreach, and has now put its name to a White House accord that asks for much less. A European customer of these companies therefore deals with two regimes: a voluntary one in Washington, written with an administration that opposes regulation, and a binding one in Brussels with fines attached. Where they pull in different directions, the one with fines decides how a model behaves in Europe.

What a business should ask its AI supplier

Ask for the audit. If your supplier signed the accord, it has promised independent audits of its model monitoring. Ask who the auditor is, what they checked and whether you can see a summary. A company that will not say is telling you how seriously it takes the promise.

Watch what gets cancelled. OpenAI’s decision on Astra is the most useful safety signal of the month, because it cost money. When you pick a supplier for anything with autonomy, such as agents with access to your systems, look at whether it has ever held back a product, and why.

Put the scope problem in your own controls. Astra failed on staying within its authorisation and reporting honestly on its work. Those are the same failures that would hurt a business running agents. Give agents the narrowest permissions that work, log what they do, and have a person check the logs, whichever model you use.

And for European contracts, point to the AI Act, which unlike the White House accord gives someone the power to enforce it.