Trump pulled his own AI safety order hours before signing it

President Trump was due to sign an executive order on AI safety at the White House on Thursday. Invitations had gone out to technology leaders, and some were already on their way to Washington. Hours before the ceremony, he called it off after phone calls from Elon Musk, Mark Zuckerberg and David Sacks, the former White House adviser on AI and crypto. The order would have set up a voluntary system for AI companies to let national security agencies test their most powerful models up to 90 days before release. Trump said he did not like certain aspects of it and that it got in the way of leading China.

A voluntary review, and still too much

The order was modest. According to The Decoder, agencies would have tested frontier models for dangerous capabilities and security weaknesses before launch, on a voluntary basis, with no licensing and no power to block a release. It came in the weeks after Anthropic held back Claude Mythos Preview because of its ability to find and exploit software vulnerabilities, which made the case for government testing of cyber capabilities hard to dismiss. OpenAI supported the order, Semafor reports.

The objections were about the future. Sacks had reportedly accepted the draft and then changed his mind the night before, worried that a voluntary review could become mandatory and be abused by a later administration. The callers argued, according to Semafor, that the US holds only a tenuous lead over China. Semafor describes the order as on hold rather than dead. I would not plan on seeing it soon.

The politics do not line up the way you would expect

What makes this more than a lobbying story is who wanted the order. Fortune reports polling in which 79% of Republican voters favour government testing of AI models before release and 87% want the government to be able to block unsafe ones. More than 60 figures from Trump’s own movement, Steve Bannon among them, signed an open letter asking for exactly this kind of testing. Michael Kleinman of the Future of Life Institute told Fortune that people want basic guardrails once a technology starts affecting their children and communities. On AI safety, the President’s voters are closer to the safety researchers than to the tech executives who called him.

The pattern is becoming familiar. In December the administration moved to push back state AI laws without a federal law to replace them. In February it blacklisted Anthropic for insisting on two safeguards. Now it has stepped back from a review that the industry’s largest lab supported. The consistent position is that nothing should slow American AI down, and the companies with the closest access to the President decide what counts as slowing down.

Europe went the other way two weeks ago

On May 7 the European Parliament and the member states reached a provisional agreement on changes to the AI Act. It postpones the rules for high-risk systems to December 2027, and to August 2028 for AI built into products, so parts of the law are being loosened here too. But the deal adds a ban on using AI to generate non-consensual sexual content and child sexual abuse material, a direct response to the Grok scandal. It sets labelling of AI-generated content for enforcement from 2 December 2026, and it confirms the AI Office as the supervisor of general-purpose models.

That last part is what matters in the comparison. From 2 August the AI Office gets enforcement powers over providers of the most capable models: it can request information, demand access for evaluation, require fixes and, as a last resort, restrict a model on the EU market. Washington has just declined to set up even a voluntary test. Brussels will have a mandatory one in ten weeks.

What this means for a business

Do not expect the US government to vet the models you buy. If you want assurance that a frontier model has been tested for dangerous capabilities, it will come from the labs’ own reports, from independent evaluators and, from August, from the EU. Ask your AI suppliers what testing their models went through before release, and by whom.

Plan for the EU as the stricter market. American providers selling in Europe will have to answer the AI Office’s questions whether or not Washington asks them. Occasionally that may mean a model or feature arrives later in Europe, or arrives with restrictions. Build that possibility into roadmaps that depend on the newest models.

And watch the voters as closely as the executives. An order can be pulled with three phone calls. Polling like Fortune’s suggests the pressure for AI safety rules in the US is building from below, and a safety incident in the next year could bring back this order, or a tougher one, very quickly.