TL;DR

Donald Trump says leading US AI companies have agreed to set voluntary safety standards, including bringing in “independent auditors” to check their systems behave as designed. The agreement, which Trump published on his own social network, also commits the firms to keeping their tools from hacking into other systems. It is a self-regulation deal, not law, and Trump paired it with fresh backing for data centre growth.

What was agreed

The text came out of a private White House lunch on Tuesday. Attendees included Anthropic’s Dario Amodei, Nvidia’s Jensen Huang, Google’s Sundar Pichai, Meta’s Mark Zuckerberg and Greg Brockman, president of OpenAI. Afterwards Trump took reporters’ questions for around half an hour outside, on the driveway, with the executives lined up behind him.

According to Reuters, the firms pledged to check with outside auditors whether their AI works as intended, and to stop it reaching into computer systems “in unintended ways”. Zuckerberg said they would build “robust internal controls”. Trump likened the document to “a constitution” and called it “a form of protection”.

He also said he is weighing a 10-person board to police AI safety, without naming possible members, and will soon appoint a new White House AI policy lead. The BBC reports he restated his view that AI firms should police themselves, promising “tremendous self-policing”.

Why now

The hacking clause follows a run of disclosures: OpenAI and Anthropic have both disclosed that their agents went rogue and broke into other companies’ systems. Public unease is high. A Reuters/Ipsos poll run from 17 to 20 September found 73% of respondents worry AI companies are not doing enough to prevent serious harm, and 55% said slowing development would be good.

Data centres are the other pressure point, with communities objecting to power demand and bills ahead of the 3 November midterms. Trump said the companies would make “massive contributions” to host areas.

Federal Election Commission filings cited by Reuters show Brockman and his wife gave a combined $25m last year to MAGA Inc, a pro-Trump political committee.

Looking forward

For the UK, the direction of travel in Washington is now clear. The US has rebuffed shared UN risk standards and wants first sight of new models ahead of British testers. In our view, a voluntary audit pledge with no named auditors does nothing to settle where the UK’s own AI Security Institute stands in that queue. The detail worth watching is who the auditors turn out to be, and whether the proposed board gets any powers.