MacCarthy: Extend US AI Safety Review to Open Weight Models

Mark MacCarthy has an argument in TechPolicy.Press that the Trump administration's new voluntary AI safety review program has a hole in it: it apparently does not cover open weight models. That means highly capable Chinese open models that US startups and enterprises are already running, plus American open models from Meta and start-ups like Reflection AI, can reach the US market without any independent prior review, while closed frontier systems still go through the process.

MacCarthy, an adjunct professor at Georgetown teaching technology policy and ethics, argues the exemption traces to a lobbying push led by Nvidia and the newly-formed Little Tech Association, which defended open models as engines of competition. He says the industry overshot: whatever business decision a developer makes about releasing weights, the underlying risks are the same, so the safety obligation should be too.

The enforcement question is where the piece gets sharper. He acknowledges the standard objection, that once weights are published they are shared globally in minutes and a US ban only stops legitimate American companies and researchers from accessing them, while bad actors abroad would simply download them. His answer is to condition US market access on compliance with safety protocols for every model, open or closed, domestic or foreign, and to pair that with US-China cooperation on shared AI risks, pointing to discussions scheduled for September as a starting point.

The essay is an argument, not a policy document, so several things it does not spell out are load-bearing: who inside the administration draws the open-versus-closed line, what the review criteria would actually be, and how a market-access condition would be enforced against models whose weights are already in the wild. Two of our tracked experts have shared the piece this week, so expect the framing to keep surfacing as those September talks approach.

If regulators pick up MacCarthy's framing, the current safe harbor for open weight developers narrows quickly, and Meta, Reflection AI, and the Chinese labs whose models US firms depend on land inside the same review process as the closed frontier houses, with all the downstream compliance cost that implies.