Trump's AI Policy Is "Minimal Regulation" in Name Only

On June 2, Trump signed an executive order that does two opposite things. It prohibits mandatory government licensing of AI models. And it establishes classified NSA benchmarking, a Treasury cybersecurity clearinghouse, DOJ prosecution directives, and up to 30 days of pre-deployment vetting for frontier models before public release. White House AI advisor David Sacks confirmed the review window was cut from 90 days to 30 so labs could comply without delaying commercial releases ^[1]. The "minimal regulation" claim holds if you measure it the way a constitutional lawyer would. No federal AI statute has passed Congress. State legislatures are filling the vacuum themselves, with Colorado narrowing its 2024 anti-discrimination law into a transparency regime and Connecticut passing a comprehensive AI accountability act with whistleblower protections for frontier-model developers ^[2][3]. By the metric that matters in law, the formal regulatory regime is minimal. What operates alongside that minimal formal regime is something else entirely. The same week as the executive order, the administration signed a presidential memorandum directing the Pentagon to accelerate military AI integration ^[4]. Defense Secretary Hegseth's AI-first strategy awarded procurement deals to seven firms, including Google, Microsoft, Amazon, Nvidia, and OpenAI. Anthropic, which had refused to allow Claude in autonomous weapons and mass surveillance, was designated a supply-chain risk, had its $200 million contract terminated, and was barred from working with other defense contractors ^[5]. Export controls give the government a kill switch over commercial availability. Anthropic's models were restricted globally for 18 days and restored only after the lab implemented a government-approved safety classifier, agreed to report malicious activity, and provided pre-release access ^[6]. When OpenAI launched GPT-5.6, federal officials limited the rollout to roughly 20 trusted partners and approved additional customers case by case, drawing criticism from Representative Lori Trahan that the government was deciding access without legal oversight ^[7]. The administration is also negotiating equity stakes in the companies it regulates, discussing a 5% stake in OpenAI valued at roughly $42.6 billion, which Trump described as making the American public a partner in the companies, while excluding Anthropic from those talks ^[8]. The government has built its own AI capability, too. A Palantir-Nvidia government-dedicated stack gives federal agencies a frontier system independent of the labs, making the government both regulator and competitor. Palantir CEO Alex Karp framed the project as the alternative to outsourcing American AI to Silicon Valley ^[9]. The architecture extends internationally through Pax Silica, a 35-nation supply-chain coalition launched with a $50 million AI-powered cargo verification platform ^[10]. The word "voluntary" is what holds both claims together. The framework is legally non-binding. That makes the "minimal regulation" description accurate. But export controls, supply-chain-risk designations, procurement decisions, and equity negotiations operate under separate national-security authority. Compliance becomes effectively mandatory, and none of it is called regulation. Meta, the only major lab that has not signed the voluntary review framework, is being pressured to submit ^[11].

While we are working through the details, we hope to sign the agreement soon — Meta

OpenAI, while complying with the voluntary framework, has called it insufficient, arguing instead for a federal regime with mandatory pre-release evaluations ^[12].

If artificial general intelligence is going to benefit all of humanity, the world needs more than voluntary commitments, individual company policies, and isolated regulatory interventions. — OpenAI

The labs had asked for an IAEA-style multilateral body and a coordinated global slowdown. What they got was national-security gating ^[13]. The architecture's deepest weakness is built into its instruments. They were designed for physical chips, but AI capabilities travel through software APIs that export controls cannot gate. China's Zhipu AI trained a competitive model, GLM-5.2, entirely on 100,000 Huawei processors that U.S. export controls were meant to deny ^[14]. Anthropic, after accusing Alibaba of using 25,000 fraudulent accounts to generate nearly 29 million Claude exchanges, made the case to senators directly ^[15].

Distillation attacks turn hundreds of billions of dollars in American investment and [research and development] into a massive subsidy for our geopolitical competitors — Anthropic

A bipartisan congressional draft, the Great American AI Act, appeared two days after the June 2 order, proposing mandatory audits for large firms and a three-year preemption of state AI laws ^[12]. It has not passed. Until something like it does, AI governance in the United States runs through national-security discretion with no published standard, rewarding cooperators and punishing refusers, all of it technically voluntary.