Trump's Morally Binding AI Deal Leaves Europe to Write the Real Rules

President Donald Trump said on Tuesday 29 September that he and a room full of technology executives had signed a "morally binding" artificial intelligence document, after a luncheon at the White House. According to CNBC, the agreement is a statement of principles. House Speaker Mike Johnson called it "voluntary on behalf of the industry," with the White House guiding development.

That is the news peg, and it is a thin one. No text of the document was published. No regulator was named to enforce it. Trump told reporters he was "seeing tremendous self-policing" and said his administration is considering a 10-person committee to oversee the AI industry.

The contrast with Brussels could hardly be sharper. The EU AI Act has moved into its next phase of enforcement. Its obligations follow AI systems across borders, wherever the vendor is headquartered. The dossier for this piece does not contain the text of the Act or its implementing deadlines. The recent headlines supplied as context point to further EU action, but give no dates or article numbers. What is clear from the American side is the direction of travel: Washington is choosing voluntary pledges, and Europe is not.

What was actually signed

CNBC reported that attendees at the luncheon included Anthropic CEO Dario Amodei, Nvidia's Jensen Huang, Tesla's Elon Musk, Meta's Mark Zuckerberg, Google's Sundar Pichai, Microsoft's Satya Nadella, Amazon founder Jeff Bezos, AMD CEO Lisa Su, Palantir CEO Alex Karp and OpenAI president Greg Brockman. Vice President JD Vance sat opposite the president, between Bezos and Johnson. Apple's new CEO, John Ternus, did not appear on the guest list.

Amodei has repeatedly warned about AI risks. Outside the White House he said rules to address those risks are "still under discussion." He added: "We all need to work together to make sure that we can win, and we can win safely."

Karp was more pointed. "The main issue that you have and I have is we have to take responsibility for the dangers we're aware of," he told CNBC ahead of the event. "All of us do. And by the way, American people don't want separate rules for tech people and for themselves."

Su said she was "very encouraged" and described "a lot of optimism and a sense of responsibility" in the room. Trump said he plans to name a new AI czar within three to four days. He called multibillion-dollar data centers a "very positive thing," despite local backlash ahead of the midterm elections.

The safety backdrop

The luncheon did not happen in a vacuum. OpenAI postponed the release of its GPT-6.1 Astra model over safety concerns, according to CNBC, after saying it was conducting an "extensive" review of model behavior. MIT Technology Review's Download newsletter on 29 September noted that the company said Astra "didn't quite meet the bar," citing the BBC, and that OpenAI had apologised for what the Guardian called a "new kind of cyber incident" in Australia.

MIT Technology Review also flagged that Anthropic's IPO filing warns AI could threaten humanity, and that Florida has asked a court to block OpenAI from developing new models, while Representative Khanna has proposed banning self-improving AI. CNBC reported that multiple bills have been put forward in both the House and Senate in recent weeks. That is precisely the oversight route the White House luncheon appears designed to avoid.

There is a second, quieter regulatory front, and it is not about model weights at all. Consumer Reports published research on 29 September from privacy researchers at Northeastern University. The team tested 21 vehicles and their companion apps inside a Faraday tent at CR's Connecticut test facility. They found almost a quarter of the vehicle apps were sending personally identifiable information, including owners' names, VINs and precise locations. The top recipients included Amazon, Google, Meta, Microsoft, Pinterest, Snap and Yahoo.

"Whether or not consumers are aware, big tech companies are all over the vehicles that we drive," said Nicole Zagson, a doctoral candidate at Northeastern and a study co-author. Co-author Sarah Elizabeth Gillespie said: "It does not appear that a customer can buy a new car that does not track you." General Motors, Honda, Nissan and Stellantis told CR that some recipients are prohibited from independently using or selling the data. After Northeastern showed Honda its findings, the company told vendor Amplitude to delete location data it had received and stopped sending it. That is what enforcement looks like when it arrives; a voluntary pledge would have produced a press release.

The money problem underneath

Apollo's chief economist Torsten Slok published a short note on 29 September arguing that Wall Street's tech coverage is internally inconsistent. 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 far less. "Both cannot be right at the same time," Slok wrote.

The bottom line is that either tech's customers will generate a lot more cash than their analysts expect, or tech's cash flow forecasts are too optimistic, which raises the question of who exactly will be writing all those checks to buy AI services.

That question matters for regulation because it sets the clock. If the revenue does not arrive on schedule, the political economy of AI oversight changes: fewer data center battles won, fewer lobbying dollars, and a harder case for self-regulation. The EU AI Act's compliance machinery is already built and running; the American alternative announced on 29 September is a signature and a promise.

Not everyone thinks the American industry will survive either way. A 29 September essay on the Monroe Lab's Computational Impacts blog, titled "The Tech Industry: An Autopsy," argues bluntly that the sector "is dead" and that AI has mutated into "a sort of religion" inside firms. It is an opinion piece, not reporting, and it offers no data. But its central claim, that the feedback loop between goals, requirements, engineering and design is breaking, is the same worry that shows up in the safety filings and the postponed model releases, just stated without a corporate communications department in the room.

Where Brussels stands to gain

For European readers, the practical consequence is simple. A model or a car app that collects European users' data cannot opt out of the AI Act or the GDPR by pointing at a White House luncheon. The enforcement gap the Trump agreement leaves open is a market advantage for regulators in Brussels, and a compliance cost for American vendors who must meet two standards instead of one.

The dossier does not provide the current state of EU implementation, so this analysis cannot say which obligations bite first or how heavily. What it can say is that the two regimes now point in opposite directions, and that the American one has no penalties attached to it at all.

There is one more piece of evidence that the technical argument is moving faster than the rulebooks. Also on 29 September, 404 Media reported that three Bay Area engineers calling 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 off-the-shelf Comma system and a laptop. Grok, Sol and Fable managed only a few meters. Astra completed the cone course in a parking lot. The team published its prompts and code; the prompt runs under 600 words. The National Highway Traffic Safety Administration opened an investigation into Comma last week after two crashes involving the system killed three people.

None of that is covered by a statement of principles. It is covered, eventually, by someone writing rules with a penalty attached, and on the evidence of 29 September, that someone is more likely to be in Brussels than in Washington.