The Executive Order That Asked Instead of Demanded

Trump signed a scaled-back AI executive order on Tuesday, replacing the 90-day advance review window in last month's abandoned draft with a 30-day voluntary submission process — and industry officials had pushed to cut it to 14 days, according to three White House officials familiar with the matter.

The order, signed privately, asks AI companies to give the government a look at their most powerful models before public release. An earlier version had called for up to 90 days of advance notice; the signed order shrinks that window significantly, reflecting the outcome of an intense lobbying effort that included a last-minute call from former Trump AI czar David Sacks that helped derail the May 21 signing ceremony, as POLITICO reported at the time.

Trump participated in a small, high-level meeting at the White House on Monday to discuss the final version, according to two officials and another person familiar with the matter. The signed order was reviewed by OpenAI, Anthropic, and Google before it went out.

The administration is asking. It is not demanding.

Here is the line from the draft order that survived into the final version: "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models."

That distinction is not a technicality. It is the story.

The order, in effect, appears to reflect a calculation that the federal government lacks the leverage to compel disclosure from companies now central to national security. OpenAI, Anthropic, Google, and Meta are not regulated entities in the traditional sense. They are strategic assets. And the order acknowledges, in plain English, that the most it can do is request access and hope the companies cooperate.

This is new. The United States government has managed transformative technologies before — but never from this position. The CLIPPER chip debate of the 1990s is the closest analogy: the NSA wanted a backdoor in consumer encryption, the tech industry resisted, and the government ultimately backed down because it lacked the leverage to compel compliance. The result was a voluntary key-escrow system that nobody used, followed by an encryption landscape the government never controlled. The AI executive order is the same dynamic at a larger scale and higher stakes, except the companies being asked to volunteer are also the ones with the most sophisticated understanding of what the ask means.

That asymmetry is the second-order effect nobody is talking about yet. The companies that opt in early — reviewing their models, sharing findings, writing the Glasswing framework's standards — are not just complying. They are writing the rulebook. Voluntary participation in a government review process becomes the de facto standard for what adequate security looks like. Companies that sit out, or lack the resources to engage, inherit a framework designed without them. The order asks for cooperation. What it may produce is a compliance floor set by the best-resourced players, with everybody else measured against it.

The proximate trigger for the review mechanism is Anthropic's Mythos Preview model, released in April as part of Project Glasswing. Mythos found thousands of high-severity vulnerabilities across every major operating system and browser. VP JD Vance said the administration was prioritizing data protection in the wake of Mythos. The Pentagon pressed for an advance-look mechanism. The order was the response.

But the version that reached the Oval Office had already been substantially weakened by industry pressure. The 90-day window was cut to 30. Industry officials pushed for 14. What arrived at the signing ceremony was a request, not a requirement — and the companies knew it.

Former Trump AI czar David Sacks had warned, before the May 21 postponement, that voluntary reviews might one day become mandatory. That warning reads differently after Tuesday's signing: the voluntary framing was not an endpoint but a negotiating position. The companies showed up, reviewed the draft, and negotiated the window down before it became official.

Project Glasswing, the consortium Anthropic announced in April, is the other half of this story. Twelve founding partners including Microsoft, Google, Amazon, Cisco, and JPMorganChase. Anthropic committing up to $100 million in usage credits and $4 million to open-source security organizations. It is a reasonable response to the security concerns that drove the executive order. It is also, not incidentally, a pre-emptive compliance framework — one that was apparently sufficient to satisfy the government's concerns enough to accept a voluntary 30-day window instead of the 90-day mechanism first proposed.

Whether the voluntary framework holds is the open question. What Tuesday's signing confirms is that the asking-not-demanding dynamic is now official policy. The labs know what they have. The government is learning what it does not. And for now, at least, 30 days — or possibly 14 — is the best the state could get.