The domestic danger is greater. If a model is too dangerous for public use but available to the state and its chosen partners, the danger has not been eliminated. It has been concentrated.
Frontier AI is not just another product. It is a tool for searching, persuading, monitoring, coding and automating. It can help defend a power grid. It can also help surveil dissent. It can find software flaws. It can also find institutional pressure points.
The more powerful these models become, the more important the rule of law becomes. If the government is going to decide who may use frontier AI, the standard cannot be presidential trust. It cannot be private conversations with favored executives. It cannot be a list that officials can change at any time. It cannot be a secret finding that one company is dangerous while comparable systems remain available elsewhere.
America needs AI governance. It does not need AI patronage.
Congress should create a real frontier-AI framework: published capability thresholds, independent evaluations, emergency suspension authority with short deadlines, judicial review, congressional reporting and explicit rules for government use. It should protect legitimate cybersecurity defenders, coordinate with allies before cutting them off, and make clear when remote model access is an export and when it is not.
There are two bad futures here. In one, Silicon Valley releases increasingly dangerous tools because every restraint is treated as surrender to China. In the other, the government keeps the most powerful tools close, grants access to its friends, delays access to the public and calls the result safety.
The first future risks serious harm by speed. The second risks authoritarianism by asymmetry.
Trump may be right that frontier AI needs brakes. That does not mean he should control the brake pedal. A president who treats the press, universities, libraries, civil servants and scientific data as things to dominate should not decide in secret who gets access to the most powerful information tools ever built.
Export control may be the legal label. The practical effect is domestic access control. And the political danger is plain: a government that restricts public access to intelligence tools while preserving privileged access for itself is not disarming the technology. It is arming the state.
Part II on Wedneday this week: If American AI becomes permissioned infrastructure, Canada and Europe have a chance to build the democratic alternative — starting with power, compute and sovereignty.