Trump Signs Voluntary AI Review Order After Pushback
Trump signed a narrowed AI executive order giving the government 30 days of voluntary pre-release access to frontier models, after industry lobbying gutted the original 90-day mandatory proposal.
6 min read
The executive order Donald Trump was supposed to sign on May 21 finally arrived Tuesday - but the version he signed barely resembles the one that was pulled hours before the ceremony two weeks ago.
Trump signed an order establishing a voluntary framework for government access to frontier AI models before they go public. Companies that produce "covered frontier models" can choose to submit their systems for government testing up to 30 days before release. Participation is optional. Licensing is explicitly off the table.
"provide the Federal Government with access to covered frontier models, subject to appropriate confidentiality, cybersecurity, insider-risk, and intellectual-property protection" for up to 30 days before release.
That's the core of what remains. The order also bars the government from using this framework as a back door to mandatory oversight: "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."
How the Original Order Got Narrowed
The story of this EO is a story about how quickly Washington's AI posture shifts when the right people make calls.
Trump delayed the signing in late May, saying publicly that he disliked "certain aspects" of the order. Behind that vague statement was a direct push from David Sacks - the venture capitalist who served as Trump's AI czar from January until his 130-day term as a special government employee ran out in March. Sacks, now on the President's Council of Advisors on Science & Technology, had reportedly argued for a review window closer to two weeks. He got 30 days, and he secured two other wins that mattered more: voluntary participation and a restriction to "advanced models only."
David Sacks, Trump's former AI and crypto czar, directly lobbied to narrow the order from 90 days mandatory to 30 days voluntary.
Source: commons.wikimedia.org
The original draft required a 90-day mandatory review period - long enough that a lab releasing a model every few months would spend a quarter of its time in a government queue. The two versions side by side:
What Changed Between May 21 and June 2
| Original Draft | Final EO |
|---|---|
| 90-day review period | 30-day review period |
| Mandatory framework | Voluntary participation |
| Applied broadly | Advanced models only |
| No licensing exclusion | Licensing explicitly banned |
Who Gets What
AI Companies
The final order is close to what frontier AI labs would have written themselves. Voluntary participation means OpenAI, Anthropic, Google DeepMind, Meta, and xAI can each decide whether to engage. There's no penalty for staying out.
The "covered frontier model" definition will matter enormously here. The order directs agencies to develop benchmarks for assessing "advanced cyber capabilities" - that test determines whether a model falls under the review framework at all. A lab that argues its model doesn't meet the threshold faces no review. Expect legal teams to argue exactly that.
Companies that do participate also get strong IP protections. The government must handle submitted models "subject to appropriate confidentiality, cybersecurity, insider-risk, and intellectual-property protection." In practice, this means reviewers can't share what they see during the 30-day window.
The Public and National Security Rationale
The security motivation behind this order is real, even if the implementation is light. White House officials pointed to advanced models like Anthropic's Claude Mythos as the explicit reason for urgency - a system capable of identifying major software vulnerabilities at scale. Project Glasswing showed what that looks like in practice: Mythos found thousands of zero-day vulnerabilities over several weeks, giving the national security community a concrete picture of what adversaries with similar systems could do.
The order's response to that threat is to ask companies nicely to share their work in advance. The power of that ask depends on whether labs see a benefit in cooperating. Most of the major ones already do: five frontier labs - OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI - signed formal pre-deployment evaluation agreements with CAISI by May. This executive order largely codifies a relationship that was already forming.
The China Dimension
The 90-day proposal didn't fail because the White House changed its security assessment. It failed because the people advising Trump argued that mandatory, lengthy pre-release review would hand China a structural speed advantage. The US is in a sprint to stay ahead on frontier AI, and every week a model sits in government review is a week a Chinese lab doesn't have to wait.
That argument won. It also sets a ceiling on how far any future administration can push mandatory AI oversight without triggering the same pushback.
The Pieces That Weren't Cut
The order establishes a new AI cybersecurity clearinghouse for agencies to share vulnerability information - one of the provisions that survived industry review.
Source: pexels.com
Three provisions survived the narrowing intact.
The order directs the DOJ to focus on AI-related crimes, specifically hacking and unauthorized access involving AI systems. That's a signal the administration is thinking about adversarial AI use by criminal actors, not only foreign governments.
A new "AI cybersecurity clearinghouse" will be established for agencies to share vulnerability information about AI systems. Its usefulness depends on what labs are willing to disclose and how quickly the clearinghouse can process and distribute findings - neither of which is specified.
Finally, the order directs multiple agencies - the Committee on National Security Systems, the Department of Homeland Security, the Office of Management and Budget, and the National Cyber Director - to develop standardized benchmarks for assessing AI cyber capabilities. Those benchmarks will determine the "covered frontier model" threshold. The benchmark design is where the real fight happens next.
What Happens Next
The immediate question is which labs will voluntarily submit for review. The five already operating under CAISI agreements are the most likely early participants. A lab that opts out will face pointed questions about why.
The benchmark definition is the bigger issue. Agencies have no set deadline for publishing those standards. Until they do, "covered frontier model" is an empty category, and the 30-day window applies to no one.
Congress hasn't moved on AI legislation, and there's no indication that changes this year. This executive order is the only binding federal action on AI oversight in the US - and it's designed to be as light as the industry would allow.
Sources:
