Trump's AI Order Is 'Voluntary.' The Leverage It Builds Is Not.

On June 2, President Donald Trump signed an executive order inviting the country’s most advanced artificial-intelligence developers to give the federal government a private look at their most capable models for up to 30 days before anyone else gets to use them. The order is careful to call this voluntary. It says so more than once, and it goes out of its way to promise that nothing in it creates “mandatory governmental licensing, preclearance, or permitting” for releasing a new model.

That promise is doing more work than it can bear. The word “voluntary” is the load-bearing beam of this order, and the structure it holds up is heavier than the framing admits. Strip away the reassurances and what the order builds is a standing arrangement in which executive-branch officials decide which AI systems are powerful enough to matter, get a head start on examining them, and help choose which companies count as “trusted partners.” That is leverage. Leverage of that kind has a way of not staying optional.

What the order actually does

Start with the text, because the details are the argument. The order asks developers of “covered frontier models” to hand the government access to those systems, under confidentiality and security protections, for as long as 30 days before a model’s wider release. It invites those same developers to “collaborate with the Federal Government to select trusted partners” who would get early access to the most capable systems. And it directs the attorney general to prioritize enforcement against people who use AI to break into computers or commit other crimes.

None of that is a licensing regime on paper. A company can decline. But notice the shape of what’s being assembled. Someone has to define “covered frontier model,” and reporting on the order indicates that work runs through a classified process involving the National Security Agency, with the Treasury Department positioned as the lead agency coordinating the effort. So the government will draw the line for which models trigger the review, do it partly in secret, and run the clearinghouse out of a department whose job has never been either AI or cybersecurity.

This is a real reversal for an administration that spent its first year treating AI rules as a brake on American competitiveness. Trump had a similar draft ready in late May and pulled it back on May 21, reportedly out of worry that it would slow the industry down. Two weeks later he signed a version anyway. The about-face is its own signal that the government now wants a seat much closer to the frontier than it had, a posture that fits the same instinct behind the administration’s push to keep advanced chips and AI leadership inside a U.S. orbit.

Why “voluntary” rarely stays voluntary

Here is the part the order’s defenders skip. A program is only as voluntary as the consequences of saying no. Once the government decides which models are “covered” and starts inspecting them, declining stops being a neutral choice and starts looking like a flag. The lab that opts in gets to call itself a trusted partner; the one that opts out gets to explain, to investors and to regulators, why it wouldn’t let the government see what it built. That asymmetry is the quiet engine that turns “you may” into “you’d better.”

The “trusted partner” designation makes the problem sharper. The order doesn’t just review models; it sets up a roster of favored companies with early access to the most powerful systems. The moment government anointment becomes a competitive advantage, the incentive shifts away from building safer models and toward winning the designation. Firms will invest in the relationship (in lobbying, in proximity, in being the kind of partner the clearinghouse likes) because that’s where the payoff moves. A safety process that rewards political access more than actual security has been turned inside out before it has reviewed a single line of code.

Industry isn’t uniformly alarmed, and that’s worth saying. Victoria Espinel, chief executive of the Business Software Alliance, praised the order for laying out “a voluntary and phased approach” and for “initiating a structured and transparent process” for industry and government to work together on AI security. That’s a fair read of the document’s stated intent. The trouble is that intent and structure diverge the instant the government holds the pen on what “covered” means and who counts as “trusted.”

The accountability gap

The deeper flaw is who can check any of this. When the definition of a “frontier model” is set through a classified process, the people expected to act on the resulting judgments (Congress, state regulators, the operators of critical infrastructure) are being asked to trust assessments they can’t independently see or verify. That’s an accountability gap with a federal seal on it.

The choice of Treasury as the hub drew an unusually blunt objection from inside the security world. “The fact that the Department of the Treasury is the lead agency acting as a clearinghouse is deeply concerning,” Nick Leiserson, senior vice president for policy at the Institute for Security and Technology, told Cybersecurity Dive, adding that “neither AI nor cybersecurity are core competencies of the Treasury Department.” When a serious analyst says the agency running your AI-safety clearinghouse doesn’t do AI or security, that is not a quibble about org charts. It’s a warning that the body holding the most sensitive information about the most powerful models may not be equipped to handle it, or to be questioned about how it does.

This is the recurring tension in how Washington has approached AI: the public keeps being told to defer to processes it cannot inspect. We’ve argued before that accountability for AI’s harms can’t be outsourced to the companies that build it, and that the build-out of the data centers and power deals behind these models deserves public scrutiny, not a rubber stamp. An oversight regime that operates in the dark fails that test from a different direction. Secrecy and capture are not opposites here. They’re the same gap viewed from two sides.

The honest counterargument

There’s a real case for the order, and it shouldn’t be waved off. The most capable AI systems genuinely do carry national-security weight, and a government that learns about a dangerous capability only after public release is a government flying blind. A 30-day window, voluntarily granted, is a far lighter touch than the licensing-and-preclearance regimes some governments have floated. If the alternative is either nothing or a heavy mandatory scheme, a phased, opt-in look is a defensible middle path, and reasonable people in the field think so.

Take that seriously, then notice what it leaves unanswered. The objection here isn’t to the government knowing what’s coming. It’s to who decides, on what authority, with what visibility, and with what pressure behind the “ask.” A defensible goal can be built on an indefensible structure, and the structure is the part that outlives this White House. Programs are written by the administration that wants them and inherited by the one that comes next. The 30 days is the easy part to like. The machinery underneath it is the part to watch.

Why it matters now

AI policy in this country has lived in a strange limbo: too important to ignore, too fast-moving for Congress to legislate, and so increasingly governed by executive orders that can be rewritten with a signature. This one moves the government from spectator to gatekeeper while insisting it has done no such thing. That gap between what the order says and what it sets up is exactly where the risk lives.

The honest version of this policy would put the definitions in daylight, give Congress and outside auditors a real way to check the government’s calls, and house the function in an agency built for it. None of that is in the order. What’s in the order is a polite request, a favored-partner list, and a classified line nobody outside the executive branch can see. Call it voluntary if you like. The leverage it creates won’t need anyone’s permission to grow.