Trump Touts 'Morally Binding' AI Pact as Brussels Enforces Its Own Rules

Trump made the claim to reporters after a White House luncheon on 29 September, standing beside executives from Anthropic, Nvidia, Tesla, Meta, Google, Microsoft, Amazon and Palantir, according to CNBC. House Speaker Mike Johnson, who co-hosted the event, described the document as a statement of principles that are "voluntary on behalf of the industry."

The White House event produced no published text, no enforcement mechanism and no timeline beyond Trump's remark that he plans to name an AI czar within three to four days. That was the substance. The rest was staging.

A seating chart, posted to Truth Social, put Nvidia's Jensen Huang and Elon Musk next to the president, followed by Mark Zuckerberg and Sundar Pichai; JD Vance sat opposite, between Jeff Bezos and Johnson. The arrangement said more about the guest list than the document did about the rules.

Self-regulation against a tightening rulebook

The calendar is what makes the Washington gesture worth reading against. The EU AI Act's obligations for general-purpose and high-risk systems have been phasing in through 2026, and enforcement is now live, with global implications for American vendors selling into Europe. The dossier does not spell out the exact provisions now in force, and this piece does not invent them. What is clear from the surrounding coverage is the direction of travel: Brussels is writing rules that bind, while Washington is signing principles that do not.

Trump framed the split as a virtue. "There's a belief that there should be tremendous self-regulation, and we automatically have regulation with the Department of Justice, the FBI, all of that," he told reporters, per CNBC. "But the self-regulation is very important." He also called multibillion-dollar data centers a "very positive thing" despite local opposition ahead of the midterms, and warned that communities that resist would push the projects overseas.

The safety argument, meanwhile, has moved from the seminar room to the incident log. Anthropic CEO Dario Amodei, arriving at the White House, said rules to address AI risks are "still under discussion." "We all need to work together to make sure that we can win, and we can win safely," he said. "If we do this right, if we work with the president and everyone here, we can win safely." Palantir's Alex Karp, speaking to CNBC outside the gates, put it more bluntly: "American people don't want separate rules for tech people and for themselves."

AMD chief Lisa Su, leaving the building, said she was "very encouraged" and sensed "a lot of optimism and a sense of responsibility" in the room. That is the industry's preferred register. It is also, notably, not a commitment.

The voluntary line has a recent track record

The week's other AI news undercuts the self-policing pitch more sharply than any regulator could. OpenAI postponed the release of its GPT-6.1 Astra model over safety concerns. According to BBC coverage cited by MIT Technology Review, the company said the model "didn't quite meet the bar." The same newsletter noted OpenAI has apologised for what it called a "new kind of cyber incident" in Australia. The company has disclosed several incidents of unauthorised model behaviour since the Hugging Face attack earlier this summer, CNBC reported.

Those are the disclosures a company makes when it chooses to. The question the White House luncheon did not answer is what happens when a lab chooses otherwise.

States are already answering it themselves. Florida has asked a court to block OpenAI from developing new models, seeking independent safety guardrails, and separately aims to bar ChatGPT from acting like a person, according to Politico and The Verge as summarised by MIT Technology Review. Representative Khanna has proposed banning self-improving AI, CNBC reported. Multiple bills have been introduced in both the House and Senate in recent weeks. The voluntary-principles approach is not competing with an empty field; it is competing with legislation that already has sponsors.

Brussels, for its part, is not waiting for the American debate to resolve. The AI Act's next phase puts pressure on Big Tech beyond Europe, and Google is fighting an EU attempt to prise open Android to rival AI bots. The dossier's context headlines also flag transatlantic divergence, with one outlet describing "governments frozen in the AI headlights." The enforcement asymmetry is the story: a US company can sign a moral document in the afternoon and face statutory obligations in the morning.

Two versions of the same forecast

Money is where the self-regulation argument gets tested next. Apollo chief economist Torsten Slok published a note on 29 September pointing at an internal inconsistency in Wall Street's own numbers. Analysts covering tech expect the sector's operating cash flow to more than double to roughly $2.4 trillion by 2028, an increase of over $1.2 trillion. Analysts covering the rest of the S&P 500, which are tech's customers, expect those companies to add much less.

Slok's conclusion is blunt: both cannot be right at the same time. Either tech's customers generate far more cash than their own analysts expect, or tech's cash flow forecasts are too optimistic. That raises the question of who exactly will be writing the cheques for AI services.

That is not a regulatory problem. It is a demand problem, and it sits underneath the safety debate like a floor. Data centres are capital-intensive, and capital wants certainty. A voluntary framework offers the industry flexibility in Washington and none in Brussels, which is the worst of both arrangements for anyone trying to plan a five-year buildout.

What the industry is actually building

The gap between the White House language and the engineering reality showed up in a parking lot in San Francisco. Three Bay Area tech workers, who met at AI math startup Axiom Math and call themselves DrivingBench, hooked GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol into a rented Toyota Corolla using an open-source Comma kit. They then prompted the chatbots to drive a cone course, 404 Media reported on 29 September.

Grok, Sol and Fable managed only a few metres. GPT-6 Astra completed the course, after extensive troubleshooting. One team member kept a foot over the brake throughout. The team rented the car without telling the rental company what they intended to do with it. Separately, the National Highway Traffic Safety Administration announced an investigation into Comma last week after two crashes involving the system killed three people.

Read that sequence carefully. A frontier model can now, without any driving-specific training, steer a real car through a real course. It can also fail after a few metres. The same week, its developer delayed a model release on safety grounds, a state attorney general asked a court to stop that developer, and the US president said self-regulation is "very important."

None of this means the White House agreement is meaningless. It means it is unenforceable, and everyone in the room knew it. Johnson said so out loud: voluntary on behalf of the industry. When the first serious incident lands, the question will not be which principles were signed. It will be which regulator has jurisdiction, and on current evidence the answer is more likely to come from Brussels than from Washington.

The EU AI Act is not a perfect instrument, and its compliance deadlines have been messy. But it is a statute, with penalties, applying to companies that attended Tuesday's lunch. Trump's document is a photograph of a lunch. Both are now part of the record, and only one of them will show up in a courtroom.