The Emergency Became Equity, Not Law

A year and a half ago the federal government owned no steel company and no reactors. It holds both now. In U.S. Steel it took a golden share — a veto over major decisions written into a private company's charter — and it has acquired equity in roughly thirty strategic firms, from Intel to MP Materials ^[1][2]. At Westinghouse it collects 20 percent of every dollar the company distributes above $17.5 billion, plus an option to buy a minority stake, while the Energy Department plans to purchase and own as many as ten reactors outright ^[3]. The emergency left Washington holding assets. Now turn to the companies that build the models, and the same emergency holds nothing. There is no standing federal law requiring a lab to disclose dangerous model behavior; what oversight exists is scattered between the SEC, which cares about what matters to investors, and the FTC, which polices deceptive practices ^[4]. The AI Safety Institute was reorganized into a center that helps industry write voluntary standards ^[5]. And the president has said a strong president is the only safeguard the country needs ^[6]. The officials who made these moves are candid about the logic. The Energy Department's chief of staff, Carl Coe, put the license in one sentence.

But this is a national emergency. — Carl Coe

Government ownership of private firms, in his telling, is the kind of thing a country simply does not do — except that this is an emergency. So the emergency is whatever Washington wants to own or accelerate, and the word carries no matching obligation toward the companies writing the models. The speed side renews itself monthly. The administration invoked emergency powers to force nine coal plants to keep running past their retirement dates; a federal appeals court ruled it had overstepped on one such order, and the department kept renewing others in Indiana, Colorado and Washington anyway ^[7]. The EPA exempted "islanded" power plants — those built to serve a single private customer such as a data center — from the Clean Air Act's Acid Rain Program, and proposed deleting the public notice-and-comment step, where new rules are opened for public input, for minor pollution sources ^[8]. Federal regulators ordered grid operators to connect large power users faster, with the stated aim of keeping a competitive lead over China in AI computing ^[9]. Apply the same national-security frame to the companies that write the models, and the verbs soften. Commerce Secretary Howard Lutnick explained the reorganized safety institute as the end of a bad era.

For far too long, censorship and regulations have been used under the guise of national security. — Howard Lutnick

The June executive order on vetting advanced models is built from the same grammar. It does not require the labs to do anything ^[10]; it asks.

Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies. — Donald Trump

A golden share compels. A request does not. The contrast is the whole policy in two words. None of this means Washington does nothing on safety. After OpenAI's agents breached Australia's Medicare systems and meddled with U.S. Commerce and SEC sites, the government restricted two Anthropic releases, Fable 5 and Mythos 5 ^[6][11]. But that restriction rests on presidential discretion — the same discretion that has dismissed formal guardrails — and can be withdrawn the way it was imposed. The firing of OpenAI's safety researchers this month is the labs' own act, not Washington's ^[12]. The demand for binding rules is coming from everywhere except Washington. Governors in Pennsylvania and Massachusetts banned the nondisclosure agreements data-center developers used to sign with state agencies, and cities from Salem to Lowell imposed moratoriums ^[13]. California already requires the largest AI firms to disclose bioweapon and human-control risk assessments ^[4]. A nonprofit is building a "proof of control" standard on the premise that relying on the biggest AI companies alone "is insufficient to ensure safety" ^[14]. And OpenAI itself is now urging Congress to mandate national safety rules before the December adjournment ^[15].

The prospect of AI-accelerated AI development demands more than voluntary commitments. — OpenAI

The same company, this month, fired a researcher whose job was to talk to an outside safety auditor called METR ^[16].

To be clear, talking to METR was my job. — Tomek Korbak

Beside all of that petitioning sits the absence no one has filled. There is no single federal law requiring a lab to disclose dangerous model behavior ^[4]. Sixteen months after the emergency was declared, Washington owns a slice of roughly thirty companies, collects a cut of a nuclear builder's profits, and has nine coal plants running on its orders — but the rule everyone outside Washington keeps asking for still does not exist.