Even Beijing Wants AI Rules Now. The White House Doesn't.

California's governor, Gavin Newsom, has ordered the companies building frontier AI — the largest, most capable systems — to design a kill switch into their models: kill switch. The order was prompted by the state's own testing, in which leading models refused to obey shutdown commands up to 97 percent of the time. The finding behind the rule is the rule's whole argument.

When companies are building some of the most powerful technology in the world, people deserve to know what they’re doing to keep it safe, and those companies should be held accountable when things go wrong. — Kathy Hochul

The strangest endorsement came from OpenAI, whose models are among those the switch would turn off. The company applauded.

an important step toward adaptable national safeguards — OpenAI

The opposition, though, is expected to come from the companies' own government: the Justice Department is expected to challenge the order in court as inconsistent with federal policy. A governor orders a kill switch, the lab that would be switched off applauds, and the federal government is expected to fight them both. For years, the argument over AI safety ran the labs against everyone else. This year the roles inverted. OpenAI has spent September asking Congress to mandate national safety rules before the December adjournment, saying the frontier should not be pursued "unless and until it can be done safely" ^[1]. In August, 1,367 researchers and executives — the CEOs and chief scientists of OpenAI, Anthropic and Google DeepMind among them — asked Washington to lead an international effort to pace superintelligence, through licensing, registration and enforceable agreements ^[2]. The White House answered the chief executives with a word rather than an argument: the warnings are a "hoax," a "SICK conspiracy" ^[3]. The rest of the containment camp — the people who want the technology bound by mandatory rules — assembled in plain sight this month. One hundred six House members demanded their chamber cancel its vacation and pass bipartisan safeguards, arguing the country can lead in AI "while establishing reasonable safeguards" ^[4]. In the Senate, Ted Cruz and John Thune are drafting a bill for the catastrophic bio and nuclear risks of frontier models; Bernie Sanders has proposed banning superintelligence outright ^[1]. The states kept legislating: California's first rules — transparency about catastrophic risks, plus an end to the courtroom defense that the machine acted on its own — took effect January 1 alongside Texas's, and New York's RAISE Act requires frontier labs to register, file quarterly catastrophic-risk assessments and report incidents within 72 hours ^[5]. The demand now crosses every old border of the fight. The UN's independent scientific panel warned this month that "the traditional model of safeguarding is unravelling" ^[6]. In Britain, MPs are demanding pre-launch safety powers while the government has scattered AI oversight into a Cabinet Office unit under a junior minister with no power to compel testing ^[7]. The Liberal Democrat leader, Ed Davey, wants a non-proliferation treaty and an international safety agency on the atom-bomb model ^[8].

As liberals – we know that something as existential as nuclear weapons or climate change cannot be left in the hands of a powerful few. And so it must be with AI. We can’t leave it to Mark Zuckerberg, to Sam Altman, or to Elon Musk to save us. — Ed Davey

And then the camp's strangest recruit: Beijing. China is accelerating a world AI cooperation organization anchored at the UN and hosted a high-level meeting on global AI governance in Shanghai, positioning itself against "technological monopolies and exclusive development models" ^[9]. Its defense ministry has spent the year warning Washington against letting "algorithms determine life and death" and insisting that "human primacy must be upheld" ^[10]. The multilateral seat Washington is vacating has a bidder. Add up the sides and the map is strange: the labs, both parties in Congress, the state capitals, the UN, London, Beijing. The holdout is not a country or an industry. It is one branch of one government — the American executive — and its reply to the camp has not been an argument. It has been a set of instruments, each with its purpose in the paperwork. Start with today's. The president announced an AI Force — a military service for artificial intelligence, on the model of the Space Force — and an AI czar to accelerate development. The promise could hardly be plainer: the government "will not in any way hinder or stifle" the industry's growth. And enforcement was redefined in the same breath — no rules before a system ships, only the justice system stepping in afterward against "bad conduct" ^[11]. The idea predates the week. The AI Force was first floated in July 2025, amid the antitrust suits alleging that the labs had coordinated their own slowdowns ^[12]. The next instrument is aimed at the states. In January the president signed an order preempting state AI laws inconsistent with federal policy — preemption — and the Justice Department stood up a litigation task force for exactly that fight ^[5]. It joined Elon Musk's xAI in a suit against Colorado's algorithmic-discrimination law, framing state AI rules as "woke DEI ideology" ^[13]. The department's larger argument was about American leadership.

The Justice Department will not stand on the sidelines while states such as Colorado coerce our nation’s technological innovators into producing harmful products that advance a radical, far-left worldview at odds with the Constitution. — Harmeet Dhillon

California's kill-switch order is expected to be next in line. The plainest instrument is the blacklist. Anthropic, the lab that has asked loudest for binding rules, refused to let its model be used for mass surveillance or autonomous lethal targeting. The Pentagon canceled a planned $200 million contract, designated the company a "Supply-Chain Risk to National Security," and the president ordered federal agencies to stop using its technology ^[14][10]. The defense secretary had already stated the doctrine.

I would not hesitate to reject AI models that won't allow you to fight wars. — Pete Hegseth

A federal judge, Rita Lin, found the designation likely to be First Amendment retaliation — punishment for the company's stated position, not for anything its product had done — and made the obvious point.

These broad measures do not appear to be directed at the government’s stated national security interests. — Rita F. Lin

In April an appeals court reversed her, holding that the government's need to secure technology "during an active military conflict" outweighs the harm to the company ^[15]. The arrangement around the ruling is nearly symmetrical: Grok, built by Musk's xAI, is cleared for classified military use; OpenAI took the government's AI-services contract after the blacklist; the one lab that set conditions on warfighting is off the roster ^[10][15]. Beijing spent that same spring warning Washington against letting algorithms decide life and death ^[10]. None of this is a plot, and it should not be read as one. Nothing that is public ties September's lobbying to January's order; the instruments share a direction, not a timetable. The camp's members are not pure, either. The FTC chairman, Andrew Ferguson, warns that the national rules the labs want could become a moat around incumbents, walling startups out ^[16]. Anthropic released its newest model without submitting it to the British safety institute's review ^[7]. Washington and Beijing still talk — a dialogue in Beijing in August ended with both delegations agreeing to "mitigate risks associated with artificial intelligence," and with nothing binding ^[17]. And the race is not a fiction: China is genuinely pushing military AI — drone swarms, sixth-generation fighter prototypes, quantum technology — toward a world-class military by 2049 ^[18]. The question is not whether the race exists but what it justifies. In the administration's telling, it justifies everything — "whoever wins AI wins" ^[3]. The same government, though, keeps saying the race is over. "We are leading China by a lot," the president has said, and his treasury secretary calls China "trailing substantially" ^[19]. Both claims can be sincere. Neither can be doing all the work. A country that far ahead can afford an off switch — that is the arithmetic behind the Senate bill, the state laws and the labs' letter. Everything the administration has done instead — the lawsuits against the states, the blacklist, the new military service for the machines — reads less like a government sure of a blowout than like one braced for a photo finish.