Trump signs one-page AI safety pledge with Google, OpenAI, Meta, xAI and Nvidia

The document was signed by Google, OpenAI, Anthropic, Meta, xAI and Nvidia, heise online reported on 1 October. It fits on a single A4 page, signatures aside. The German outlet noted that the seating plan Trump posted on Truth Social drew more likes and comments than the declaration itself.

What the paper actually commits the companies to is thin. Each firm is declared responsible for making its own technology safe and for building products in a way that lets users trust them. To that end, the signatories say they should run internal processes that prevent safety problems or at least detect them quickly. An internal team should monitor those measures. A separate committee drawn from the board of directors should receive the safety teams' reports and oversee their work.

Controls should monitor the capabilities and alignment of AI systems in cybersecurity, biosecurity and chemical threats. Models should not hack or otherwise gain unauthorised access to technical systems. The companies should also meet regularly, as heise notes they already do in a smaller circle, to develop standards and best practices.

"Zusammengenommen würden diese Maßnahmen dazu führen, dass die Unternehmen, ihre Kunden und die Öffentlichkeit wieder auf die Sicherheit der KI-Systeme vertrauen könnten," the paper concludes, according to heise.

Heise's own reading is sceptical. The outlet points out that most of these measures, the directors' committee aside, have already been announced by the AI companies, and that the first two have long been in place while organisations keep getting hacked. It also flags the grammar: the text says "we believe that [...] should" throughout, never "we will". Only in the second-to-last sentence does the paper suggest that enshrining these measures in regulations or laws might prove useful at some point. Maybe.

The timing invites the scepticism. Heise cites ongoing investigations by the US trade regulator and lawsuits including one brought by Florida against OpenAI as evidence that trust in these products is not currently justified. MIT Technology Review's 29 September edition of The Download lists more of the same week: OpenAI scrapped a new model, GPT-6.1 Astra, over safety concerns, according to the BBC; the company apologised for what the Guardian called a "new kind of cyber incident" in Australia; Florida asked a court to block OpenAI from developing new models, per Politico, and to stop ChatGPT acting like a person, per The Verge; and Anthropic's IPO filing warns that AI could threaten humanity, the Financial Times reported.

Antitrust is where the binding rules sit

Nothing in the declaration changes competition law, and that is where the harder obligations on these companies are being written. Google's position in advertising technology was the subject of a US antitrust ruling that let the company keep both sides of its ad business. Recent headlines tracked by this desk record a judge finding Google broke antitrust law and then allowing it to retain its entire ad-tech operation, a decision that also ended rigged auctions. The same coverage notes Bank of America calling the outcome a "major boon" and raising its target price for the ad platform Magnite to $38.

That contrast matters for reading the White House lunch. A voluntary pledge signed by the companies themselves, monitored by committees the companies appoint, is a different instrument from a court judgment that imposes remedies. Heise's observation that the paper only suggests regulation "might" be useful later is the point: the signatories keep the pen.

Elsewhere in the dossier, the week's most recent items show how wide the gap is between declared principles and enforcement. STAT reported on 1 October that the Department of Homeland Security appealed a court ruling blocking limits on how long international students and postdoctoral researchers can stay in the US. DHS wants a four-year cap, arguing the longstanding policy of allowing researchers to stay for the length of their training promoted fraud. A coalition of academic groups sued, saying the policy violates federal law on new regulations and would jeopardise US scientific preeminence. A federal judge issued a preliminary injunction while the case proceeds, and the administration now hopes the First Circuit will overturn it. The appeal was filed on Wednesday.

Also on 1 October, STAT's Morning Rounds carried a KFF Health News analysis on immigrants losing Medicaid under last year's tax bill, and a KFF finding that 14 per cent of new applicants and 7 per cent of people renewing benefits in Nebraska were denied or lost coverage for failing work requirements.

Tech.eu reported the same day that Belgian tech companies raised €1.09 billion in the first half of 2026, with Kpler alone accounting for around 79 per cent of the total on an €859.2 million round. Analytics led by value, followed by software at around €97.6 million and transportation at €35.5 million.

What the pledge leaves out

The declaration's control list is specific about cyber, bio and chemical threats and about alignment. It says nothing about competition, market structure or the remedies a court can impose, which is precisely where Google's case ended up. It says nothing about liability when an AI agent acts on its own, a question MIT Technology Review raised after Meta's Muse agent shared a user's address while negotiating a Marketplace sale, as the Guardian reported.

MIT Technology Review made a related point about claims on 29 September. Anthropic said its new molecular biology lab had made its first discovery, with AI agents flagging a previously uncatalogued pattern around an enzyme, reminiscent of what led to CRISPR. Biologists pushed back. Some questioned whether finding a pattern amounts to a discovery at all. One said his team had already found the same pattern, raising questions about whether Anthropic's system had learned from his conversations with Claude. The newsletter's conclusion applies neatly to the one-pager from the lunch: what is novel for an AI system may be routine, unsurprising or not that consequential to a specialist, and grand claims can make real progress harder to recognise.

Heise's account of the seating is the last detail worth keeping. Musk sat to Trump's left, Nvidia to his right. Anthropic chief executive Dario Amodei was six places away behind Meta, Altimeter and Broadcom, while OpenAI's Greg Brockman took a similarly distant seat on the other side. Google got the place next to Musk.