When the government barred Anthropic from providing access to foreign nationals, including foreign-national employees inside the United States, Anthropic pulled the models off the market. Some access has since been restored to approved entities. General public access remains another matter.²
Two cases do not prove a permanent architecture. They do, however, reveal a direction of travel. The administration is not merely asking whether some frontier models are dangerous. It is beginning to sort users into categories: government first, trusted partners next, everyone else later.
The administration calls this national security. It may even be right that national security is involved. Frontier AI is becoming genuinely dangerous. A model that is very good at code is not just a better office assistant. It can find software vulnerabilities, automate attacks, help patch critical systems, accelerate biological research and lower the skill threshold for tasks that once required teams of experts. A tool that can search software for weakness can defend the country or attack it.³
So yes, frontier AI may need brakes. The “race with China” frame can become a license for recklessness. If every safety rule is treated as surrender, the United States may release systems it does not understand because nobody wants to be accused of losing the twenty-first century.
But Trump’s answer is not a brake. It is a private roadblock.
A democratic safety regime would publish rules. It would define what level of cybersecurity, biological or autonomous-agent capability triggers review. It would apply the same standards to OpenAI, Anthropic, Google, Amazon, Meta, xAI and any other company that crosses the line. It would include independent testing, congressional oversight, judicial review, emergency procedures, deadlines and protections for legitimate cyber defenders.
A control regime works differently. Companies bring frontier models to the government first. Officials decide which partners are trusted. Access is limited not by published rules but by administrative discretion.
That is the line being crossed.
The formal Trump executive order on frontier AI says the review process is voluntary and does not create a licensing regime. That language matters. It also sounds increasingly detached from the practical effect. OpenAI previewed Sol to the government and then limited access to selected partners following discussions with federal officials. Anthropic had its models pulled after an export-control order and is now restoring access only through approved channels. If this is not formal preclearance, it is preclearance by gravity.⁴
This is where AI policy becomes part of a larger Trump pattern. The mechanism is not always censorship in the old-fashioned sense. It is conditional permission.
The administration has repeatedly shown interest in pressure points that sit between formal law and practical access: broadcast licenses, merger approvals, public broadcasting funds, university grants, library funding, federal data, procurement rules and access to government contracts. Each dispute has its own facts. But the governing instinct is similar. Do not necessarily ban the thing. Control the channel. Put the institution on notice. Make continued access depend on staying in the good graces of power.⁵
The government does not have to say, “You may not speak.” It can say, “Your license is under review.” It can say, “Your funding is gone.” Now it can say, “Your model is available only to approved users.”
That is how information control works in a modern administrative state: not always through one giant ban, but through pressure on the channels through which knowledge flows.
The administration’s stated target is China. That concern is not frivolous. China wants advanced AI for military, intelligence, cyber and industrial advantage. No responsible government would ignore that.