On June 2nd, President Trump signed an executive order directing federal agencies to establish a voluntary pre-release review process for the country’s most advanced AI models. Under the order, companies are asked to give national security agencies — the NSA, representatives of the Defense Department — access to evaluate frontier models up to 30 days before public release.

The order explicitly stops short of mandating participation and bars the creation of any new licensing or permitting requirements. The White House was quick to reinforce this, posting that the government would “not be conducting oversight of all new models” and that “that level of government overreach would have chilling effects on free speech and innovation.”

How We Got Here

Two weeks before signing this order, Trump scuttled an earlier version that gave the government a 90-day review period. The shorter timeline — 30 days — was characterised by David Sacks, Trump’s former AI advisor, as a “game changer” that would allow companies to engage without significantly slowing model releases.

What changed? Internal administration battles, external pressure, and at least one catalyst: Anthropic independently brought its Claude Mythos Preview model to senior White House officials, a disclosure that exposed software vulnerabilities and raised concerns about the wisdom of releasing powerful models without prior government evaluation. The episode illustrated what security researchers had been arguing for months: frontier AI capability is outpacing any meaningful understanding of its risks.

The Republican Split on AI

The political backdrop is notable. Trump came to office promising to strip away AI guardrails. That posture has fractured the GOP. Polling from May 2026 showed strong majorities of Republican voters supporting exactly the kind of framework Trump signed into law — 71% said independent security testing should be required by law for advanced AI systems.

That’s a significant disconnect between the party’s free-market deregulatory instinct and its voters’ anxiety about AI’s societal impacts. Florida Attorney General James Uthmeier sued OpenAI over alleged risks from ChatGPT in a shooting at Florida State University. Republican Congressman Byron Donalds — Trump-endorsed to succeed Ron DeSantis — publicly broke with Trump on AI policy, saying he supported state-led regulation.

The executive order didn’t resolve this tension. It papered over it with a “voluntary” framing that gives the administration deniability while allowing it to point to action.

Whether It Stays Voluntary Is the Real Question

The most perceptive critique came from Dean W. Ball, Trump’s former AI advisor, who called the order “a mistake” and warned it could be “a first step toward a federal licensing requirement for AI models.” His concern wasn’t hypothetical — the history of voluntary frameworks in regulated industries (financial services, healthcare, nuclear) suggests that “voluntary” often becomes “expected” and eventually “mandatory” once the government establishes the infrastructure to receive and act on the information.

James Sanders at the Center for a New American Security put it more neutrally: “It’s unclear how voluntary this will stay and how voluntary it will be in practice as the AI labs try to maintain good relationships with the U.S. government.” The labs need compute, data, talent, and goodwill with policymakers. Sharing models with the NSA is now part of the landscape for maintaining those relationships.

Meanwhile, states are forging ahead regardless. The executive order was framed partly as a preemption signal — Washington wants to set the AI regulatory agenda and keep states from fragmenting the landscape. But Florida’s lawsuit against OpenAI suggests that signal isn’t landing. The states that want to regulate AI are going to regulate it.

What This Actually Means

The executive order formalises what was already happening informally — frontier AI companies sharing models with government before release, motivated by a mix of goodwill, legal exposure concerns, and the desire to stay in the administration’s good graces. Making it a structured process with a 30-day window is an institutional improvement over ad hoc disclosure.

Whether it amounts to meaningful security oversight is another question. The government has 30 days to evaluate systems that may have capabilities it doesn’t fully understand. That’s not a lot of time. And the order explicitly limits the government to evaluating national security and cybersecurity risks — not broader societal harms like labour displacement or algorithmic bias, which are what most AI safety researchers consider the more tractable problems.

The order is a start. It’s not a framework. Congress would need to act to make any of this mandatory and extend it beyond cyber threats. That legislation doesn’t exist. In the meantime, the executive order tells us something about where the debate in the Republican Party is heading — towards at least nominal engagement with AI risk — even if the mechanisms are still formless.