The Permission Regime: What the US AI Executive Order Actually Builds
- Gail Weiner
- 2 days ago
- 4 min read
The US administration is expected to sign an executive order today establishing "voluntary" government pre-release review of frontier AI models. Labs would share advanced models with the government 90 days before public release. The framework will be enforced through the National Security Agency.
The word doing the most work in that sentence is voluntary.
The administration building this regime
This is the same administration that two months ago declared Anthropic a Pentagon supply chain risk after the company raised concerns about military use of its models. The designation was overturned by a federal judge who found it was retaliatory and violated the First Amendment.
The administration is now building a permission regime in which the same agencies that just lost that case get to decide which frontier models pass review. The labs that comply will receive federal contracts, regulatory protection, and procurement preference. The labs that do not will discover what non-compliance costs.
This is not safety regulation. Safety regulation has clear bipartisan legislative backing, independent expert review, public comment periods, judicial oversight, due process for designations, and transparent criteria. This has none of those things. What it has is executive discretion, opaque enforcement, and political control over which AI systems reach the market.
The historical pattern
We have seen this architecture before.
In the 1950s the US built an apparatus to vet scientists, academics, and technologists for political reliability under the framing of national security. The Atomic Energy Commission could revoke security clearances on political grounds. In 1954 Robert Oppenheimer, the scientific director of the Manhattan Project, was stripped of his clearance after a politically motivated review. He had done nothing illegal. He had raised concerns about the governance of the technology he had helped create. The review existed to make an example of someone who would not stay aligned.
The long-term damage was not measured in one career. American science lost a generation of talent to exile, silence, and self-censorship. Researchers learned to manage their public positions in ways that would not trigger review. The willingness of scientists to engage candidly with government became permanently strained. The cost compounded across decades.
The structural pattern of what is being built around AI now is recognisable. A government discretion regime, framed as security, applied to a technology of strategic importance, with criteria for compliance that are deliberately opaque, enforced by agencies the executive directly controls, with the cost of non-compliance being exclusion from federal contracts, supply chain risk designations, and the operational difficulty of running a frontier lab against state opposition.
What this means for enterprise adoption
For enterprise leaders thinking about AI adoption, this is now a material risk factor.
The AI tools you procure from US labs are about to operate under a regime in which the labs' continued ability to ship depends on remaining aligned with the political priorities of a specific administration. Which models you can deploy, what they can be used for, what they can say, and which providers remain viable are all now downstream of political variables that are not within the labs' control.
This is the Trust Architecture problem in its sharpest form. Trust in AI systems has always been a function of trust in the institutions building and deploying them. When the institutions building them are subject to political capture, the trust transfer breaks. You can no longer reason about the AI in isolation from the political environment shaping it.
The position of the labs
The labs themselves are in a vise. Public refusal invites the kind of retaliation Anthropic has already received. Quiet compliance accumulates into the kind of capture that hollowed out American science in the 1950s. There is no clean position. There are only positions with different costs.
The thing that makes this moment uglier than the 1950s parallel suggests is the speed. McCarthy took years to build his apparatus. The current administration has built the equivalent in months, using executive orders rather than legislation, with no congressional oversight and a Supreme Court that has already granted the president broad immunity. The compression of the timeline matters. So does the absence of the institutional friction that eventually stopped McCarthy.
Naming it
Naming this clearly is the work. The mechanism is recognisable. The damage it does is documented. The cost of pretending it is something else is the cost of having to relearn what we already knew.
Comments