Trump Killed His Own AI Safety Order at the Last Minute

On May 21, the photo op was arranged. The AI and tech CEOs were ready. The executive order on AI and cybersecurity was apparently minutes from being signed. Then Trump pulled it.

He told reporters in the Oval Office that he had "postponed" the order because he didn't like "certain aspects" of it. He added: "I think it gets in the way of, you know, we're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead."

Fine. That's a legible position. But here is the part that keeps nagging at me: according to CNN and other outlets with knowledge of the draft, the order was built around a voluntary framework. AI companies would share advanced models with the government before release. No mandatory disclosure. No hard deadlines. The enforcement mechanism for noncompliance was essentially nothing. One version of the draft called for a review period of up to 90 days; the companies were reportedly pushing for something closer to 14.

So Trump killed a toothless safety measure because it was too much regulation. The accelerationists won the day, and what they defeated was a framework that, by design, could not have stopped them from doing anything.

Axios reported that Trump's AI adviser David Sacks also opposed the order, and that one White House source described the whole thing as unnecessary, "just something doomers wanted." That framing tells you a lot. Safety-minded review of frontier models isn't a doomer position. It's what Anthropic itself has been requesting. The irony here is that it was Anthropic's model, Mythos, that reportedly set off the alarm bells inside the administration in the first place. NPR noted that Anthropic's announcement that Mythos was "too powerful to release" due to cybersecurity concerns had prompted the White House to seriously consider the executive order at all. The administration got spooked by a lab telling them a model was dangerous. Then, when given an opportunity to institute even a soft review process, it blinked.

I want to be precise about what was on the table. The draft order, per CNN, would have set up a voluntary "clearinghouse" involving the Treasury Department, other agencies, and AI companies to identify security vulnerabilities in unreleased models. It also called for more hiring at the US Tech Force. Nothing in the leaked text mandated anything binding. Companies that didn't want to participate didn't have to participate.

The timing is strange in another way. Trump's first day in office included rolling back Biden's AI executive order, which required leading labs to actually share safety test results with the government. Now the replacement, a much weaker ask, has been pulled before it could even land. There's no federal AI safety framework of any kind on the horizon, and Congress hasn't passed anything either.

What this really surfaces is a structural problem that isn't going away. The administration wants to stay ahead of China in AI. Anthropic, an American lab, told the government it had built something it considered too dangerous to release. That's a genuine conflict in the accelerationist worldview, and it has no clean resolution. If you want American labs to win the race, you probably want some process for knowing what those labs are building. You can't do both "all speed, no guardrails" and "we trust our labs are being responsible." The killed order was an attempt to square that circle, badly. Now there's just the circle.

Axios noted that the White House's Office of the National Cyber Director is still working on separate AI security initiatives. So the story isn't finished. But as of today, the most powerful government in the world has less formal AI oversight than it did before Donald Trump took office. That's not spin. That's just what happened.