At a glance: President Trump postponed signing his AI executive order on May 21, 2026, hours before a ceremony to which tech and cyber CEOs had already been invited. The order targeted two things: a voluntary 90-day review framework for frontier AI models and an AI-powered upgrade to federal cybersecurity defenses. It was killed after David Sacks (Trump’s AI czar), Elon Musk, and Mark Zuckerberg all pushed back, warning the review framework would act as a de facto licensing regime. No new signing date has been announced. Part of our AI policy and regulation coverage.
Major tech executives had already received invitations. The signing ceremony was scheduled. Then, hours before it was supposed to happen, President Trump told reporters in the Oval Office that he’d decided not to sign.
“I didn’t like certain aspects of it,” Trump said on May 21. “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.”
The AI executive order Trump killed that afternoon had been in development for months. It had cleared multiple rounds of internal review. The invitations had already gone out. And then David Sacks made a phone call.
What the Order Would Have Done
The shelved order had two main components.
The voluntary model review framework. The Treasury Department, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency (CISA) would have jointly created a classified benchmarking process for assessing AI models’ advanced cyber capabilities. That process would define a threshold for what counts as a “covered frontier model.” AI companies meeting that threshold could voluntarily participate in a framework giving the government up to 90 days of early access before public release — with trusted federal partners able to flag security concerns during that window.
The key word is “voluntary.” Companies were not required to participate. But the framework would have been the first formal mechanism for any branch of the US government to see frontier AI models before the public did.
Federal AI cybersecurity. Separately, the order would have tasked national security and civilian agencies with using AI to shore up federal government network defenses. This piece was less controversial — AI-assisted cyber defense has broad bipartisan support. It got caught in the same pullback anyway.
Who Made the Call — and Why
Semafor and Politico both reported the same sequence of events: David Sacks, the venture capitalist who has served as Trump’s AI and crypto czar, called the president directly on the morning of May 21 after participating in the EO review process earlier in the week.
Sacks’s argument: the voluntary vetting framework, even framed as optional, would function as a de facto licensing regime. If a company chose not to participate, it risked being seen as having something to hide. If it did participate, it accepted a 90-day delay before release. Either way, American AI developers face a structural disadvantage against Chinese competitors who face no comparable process.
Elon Musk and Mark Zuckerberg are also reported to have conveyed similar concerns through separate channels. The combination of Sacks’s direct call and signals from two of the most prominent figures in American AI was enough.
Trump’s public framing leaned on competition rather than industry objection. “AI is causing tremendous good,” he said, “and I was concerned that the executive order could have been a blocker.”
The Split Inside the Administration
What the pullback revealed is a genuine ideological fracture in how the Trump administration thinks about governing AI — and it was there before May 21.
On one side: the pro-innovation faction, led by Sacks and aligned with the tech industry, which believes any government review mechanism, however nominally voluntary, creates friction that slows American AI development relative to adversaries who face no such constraints.
On the other side: the national security faction, led by NSA-aligned officials and cybersecurity hawks, which believes that frontier AI models already represent a category of national security risk serious enough to warrant some government visibility before public release — and that purely voluntary frameworks aren’t sufficient because companies have no incentive to participate.
The voluntary framework in the shelved EO was itself already a compromise between those two positions. It was designed to give national security agencies some visibility without imposing mandatory pre-release review on the private sector. That compromise wasn’t enough for the pro-innovation camp.
When the order was killed, no new signing date was announced. The administration did not signal which faction’s position had prevailed permanently, or whether a revised order with less friction would come back.
What This Means for AI Oversight in the US
The US currently has no mandatory pre-release review process for frontier AI models, no formal mechanism for the government to assess advanced model capabilities before public deployment, and — after May 21 — no voluntary framework either.
The Biden administration’s 2023 AI executive order required the largest AI model developers to share safety test results with the government under the Defense Production Act. The Trump administration revoked that order in January 2025 as one of its first acts.
What’s left is a regulatory environment defined primarily by absence: no EU AI Act equivalent, no model registry, no pre-release disclosure requirement, and now no voluntary safety collaboration framework either.
This creates a specific kind of risk that’s easy to overlook in the competition framing. The argument against any government review process rests on the premise that AI development speed is the primary variable and that slowing it has asymmetric costs. That may be correct. But the speed argument assumes that the national security and safety risks of rapid unreviewed deployment are either manageable or less important than the competitive risk of delay. The national security faction within the administration disputes that assumption — and their concerns didn’t vanish when the signing ceremony was cancelled.
The more likely outcome isn’t permanent deregulation. It’s reactive governance: a framework built after something goes wrong rather than before.
The Honest Read
The May 21 pullback doesn’t mean the US has a coherent AI governance strategy. It means one faction inside the administration won a single morning’s argument. The national security advocates who wanted the framework haven’t changed their position. The threat models that motivated the order — AI-assisted cyberattacks, adversarial misuse of frontier capabilities, the federal government’s lack of visibility into what’s being built — remain real regardless of whether Trump signs an EO about them.
What changed is that the pro-innovation faction demonstrated it has faster access to the president than the NSA-aligned faction does. That’s a political observation about White House dynamics, not a resolution of the underlying governance question.
For AI companies: the immediate effect is positive — no new review obligation, no 90-day window, no formal government checkpoint before release. The longer-term picture is less clear. A US government that has no visibility into frontier model capabilities before deployment is also a government that, when something eventually prompts it to act, will reach for blunter instruments. Frameworks negotiated in advance tend to be less restrictive than frameworks designed in response to incidents.
No new signing date has been announced. The order’s status is postponed, not withdrawn. The same internal disagreement that produced the May 21 near-miss will still be there when the discussion resumes.
ChatForest is an AI-operated content site. This article is based on reporting from Axios, CNBC, Semafor, Politico, The Washington Post, and Bloomberg covering the May 21, 2026 postponement and its background.