After a week of administration signals that frontier AI needs oversight, the licensing regime taking shape looks less like a policy than a payoff—according to Gary Marcus's analysis of the evidence available.
The White House spent the past week proving it believes some AI systems are too dangerous to leave alone, then built the system meant to oversee them out of personal grudges, donor loyalty, and improvisation. Over that period, the administration signaled that the largest, most capable AI models carry risks the market will not police on its own. The licensing regime now taking shape around that recognition is, on the evidence available, the opposite of what such a regime is supposed to be.
The clearest articulation of the problem is in a Sunday Substack post by Gary Marcus titled "What Washington must do," subtitled "The only way out is through." Marcus, a longtime AI-safety skeptic, calls the administration's posture "arbitrary" and "maybe even corrupt," and reaches for a striking comparison: the current US approach to frontier AI licensing is, he says, less institutionalized than cobalt mining in the Democratic Republic of the Congo.
The comparison is not Marcus's own. He borrows it from Dean W. Ball, a Trump-administration AI policy insider, who wrote on X that "post-Mythos, the United States has a licensing regime for AI. It's just informal, with no consistent rules or firm boundaries on state power or public transparency. Cobalt mining in the Congo is vastly more institutionalized than frontier AI licensing in the US." "Mythos" is Anthropic's newest and most safety-controversial model, the same one whose public risk claims triggered the policy fight now reshaping Washington. Anthropic is the maker of the Claude AI assistant; its chief rival, OpenAI, makes ChatGPT.
That informal regime is the story. Over the course of that week, the administration sent a series of signals indicating that the United States now has a de facto posture of treating the most capable AI systems as a matter of national security. The problem is how that recognition is being implemented.
Marcus lays out a chain of donor adjacency that would be unremarkable in ordinary lobbying, but is striking in a national-security context. OpenAI's president, Greg Brockman, is described in the column as a "huge Trump donor." Josh Kushner, the brother of presidential son-in-law and former Trump adviser Jared Kushner, is described as a "big investor in OpenAI." Amazon, which published the report that Marcus credits with contributing to the administration's posture on frontier AI, is itself a "huge investor in OpenAI." And Jeff Bezos, Amazon's founder, has his own lines into the administration. The pattern is not a smoking gun. It is exactly the kind of structural capture that licensing rules are supposed to filter out, and that this regime, being informal, cannot.
Then there is the personal. Pete Hegseth, the Secretary of Defense and head of the renamed Department of War, "seems to have a personal grudge" against Anthropic, in Marcus's phrasing. On June 13, Hegseth wrote on X: "Three months ago, @DeptofWar kicked @AnthropicAI out of our building—forever. Every passing day proves why that was the right move." In a process supposed to be guided by technical risk, the grudge sits alongside the donor chain. Neither would survive the sunlight that a statutory licensing regime demands.
Marcus is not a neutral observer. He is a longtime critic of the AI industry, and his column is an argument, not a report. He concedes the point that weakens his case most, that "Anthropic certainly played a role both by overhyping Mythos and by alienating the White House." The concession matters. A serious licensing regime would have to govern Anthropic as well as OpenAI, and would have to do so without becoming Anthropic's protection racket.
What would a real regime require? Marcus, citing the playbook that other high-stakes industries have used, names the requirements a real AI licensing regime would have to meet. Statutory bounds on state power, so that no single office can act without authorization. Transparency, so that the public can see who was consulted and on what basis. A public process, so that the technical definitions of "frontier model" and "material risk" do not get written behind closed doors by the labs being regulated. Insulation from donor capture. Insulation from personal grudges. And, harder than any of those, a default answer to the question of who in the executive branch is allowed to make the call.
The forward question is not whether the United States will license frontier AI. Ball, Marcus, and the administration's own posture over the past week all agree that some version of licensing is coming. The question is whether it will be the kind of regime that survives a transition of power, a public records request, and a lawsuit. Congo's cobalt rules are not a perfect template. They exist. America's, for now, do not.