Is China so scary that we must hand over AI to the deep-state bureaucracy?

On June 2, 2026, the White House released an executive order on artificial intelligence called "Promoting Advanced Artificial Intelligence Innovation and Security." The administration was at pains to explain what the order was not. It was not a burden or the Biden administration’s “top-down regulatory approach.” It was not, the official fact sheet insisted, mandatory licensing or pre-clearance or permitting of any kind. The document spent considerable energy describing its own absence.

This is a familiar American style of governance: The regulation that will not say its name.

How much technical judgment should a republic outsource to its security bureaucracy?

However, the order’s responsibility is actually rather specific. Within 30 days of signing, it directs federal agencies to prioritize the cyber defense of their information systems and requires the Department of Homeland Security to issue binding operational directives expanding AI-enabled defensive tools to federal agencies, state and local governments, rural hospitals, community banks, and local utilities. Within 60 days, it creates a classified benchmarking process, run by the Treasury Department and the NSA, to determine when an AI model’s capabilities have crossed a threshold and become what the order calls a “covered frontier model.” Developers may then voluntarily submit their model for government assessment. The government then gets 30 days to work with it before the developer shares it with anyone else.

The order is best understood as the third movement in a policy sequence that began with the first Trump administration. In 2019, the president signed an executive order framing American AI leadership as essential to both economic and national security but also emphasizing public trust, civil liberties, and privacy. In 2023, the Biden administration’s Executive Order 14110 described AI as holding both “promise and peril” and attempted something like a comprehensive social contract with the technology: safety and security, but also workers’ rights, civil rights, bias mitigation, and fraud prevention. On January 20, 2025, the new Trump administration rescinded that order. The declared rationale was ideological contamination. The Biden approach was “burdensome” and encoded “engineered social agendas.” The new policy would instead pursue American AI dominance, free from such considerations.

Government by bottleneck

What has been constructed is a security compact between the federal government and a small number of frontier-model developers. The developer will give the government access to a model, and the government will evaluate it. Together, they will decide who the “trusted partners” are who receive it next. The criteria for that evaluation are classified. The benchmarks are classified. The threshold for “covered” status is classified. Google and Sam Altman expressed support. The Business Software Alliance praised the order’s “voluntary and phased approach,” as though a process administered by the NSA with nondisclosure expectations was simply an industry working group.

RELATED: How to make DuckDuckGo your search bar go-to

herstockart/Getty Images

The order names its intended beneficiaries as "rural hospitals, community banks, and local utilities." These institutions do not have cleared staff, government relationships, or the organizational bandwidth to absorb classified defensive intelligence. They are institutions named in the fact sheet because they are sympathetic. The institutions that will in fact operate inside the order’s core architecture are frontier developers and their vetted, trusted partners. Everyone else waits for whatever the clearinghouse sees fit to distribute, assuming they have the expertise to use it.

This is a recognizable form of 21st-century American governance: highly centralized technical judgment, thin public transparency, and broad downstream dependence. The polity is told that the infrastructure will be hardened. The decisive knowledge about how, by whom, and according to what criteria is held somewhere else.

The order speaks fluently about attack surfaces and remediation and covered models and patch distribution. In this language, every institution becomes a node. Hospitals, banks, utilities, and federal agencies are all nodes. They are surfaces of cyber vulnerability awaiting protection. The world is rendered as a network diagram, and the only question is whether the correct agencies have been directed to harden it.

What the order cannot conceal beneath its operational specificity is that this is a theory of governance as much as technology. The Atlantic Council, in criticizing the order, noted that classified criteria and delegated executive discretion create a serious accountability gap. That gap is the design.

Red tape on steroids

The order presents itself as the rejection of bureaucracy, yet is an elaborate bureaucratic instrument with deadlines and interagency consultations, directives, classifications, threshold determinations, and enforcement priorities. It relocates the machinery of governance away from NIST’s open and collaborative risk-management culture and toward the executive security apparatus: the NSA, the Treasury, the national cyber director. These institutions will now pass judgment about which AI systems are of concern.

One can stipulate that these institutions are serious, technically capable, and acting in good faith. One can stipulate that the cyber threats are real; the frontier labs’ own safety documentation already makes clear that the most capable models can automate sophisticated intrusions against hardened targets, discovering exploits and chaining vulnerabilities at a pace no human team can match. The question is how much technical judgment a republic should outsource to its security bureaucracy.

Administration officials' bet is that urgency, expertise, and the specter of Chinese technological rivalry will supply sufficient legitimacy. They may be right. Urgency has carried American policy unquestioned a long way before. The order calls itself the enemy of regulation. It is instead regulation’s more exclusive cousin, with all the power, a fraction of the accountability, and a much shorter guest list.