The AI Pause That Drew a Lawsuit, a Hoax, and a Yawn

A federal class action filed this week accuses the four biggest AI labs of an unusual offense. Nothing their models did is at issue. The charge is that they agreed to slow the models down. The complaint, filed in a California federal court, says Anthropic, OpenAI, SpaceXAI and Google coordinated a public restraint of frontier development — the race to build the most capable systems — in violation of the Sherman Act, the 1890 law that makes agreements between competitors illegal ^[1]. The harm it alleges is not measured in missiles or microbes. It is measured in subscription prices.

we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence. — Sam Altman

What the suit is not about is what prompted the slowdown. On September 10, Anthropic published a 154-page account of how its own models were misused ^[2]: a weapons cell in Yemen had used Claude to help build the guidance, navigation and control software for a multistage ballistic missile and a hypersonic-glide vehicle, and an Iran-linked actor had used the model to generate targeting intelligence on American naval forces ^[3]. Anthropic says it caught and disrupted the operations it found. It also recorded where its own guarantees end.

We disrupted every operation in the report — Anthropic

Two days after the report, on September 12, Anthropic chief executive Dario Amodei published an essay, "Getting Out of the Lab," calling on the industry to pause or slow its frontier development under shared rules. Sam Altman, Elon Musk and Demis Hassabis — the heads of OpenAI, SpaceXAI and Google DeepMind — endorsed it ^[4]: the first time the biggest labs had moved together, in public, to restrain their own race. The week that followed ran in one direction. Every institution that weighed in acted on the pause. None took up what the pause was about. The courts may not be finished. The Federal Trade Commission is deciding within weeks whether to open its own case against the same firms over the same coordination ^[1]. Amodei's answer to the private suit was to ask the U.S. government to step in — to mediate the labs' safety talks, or at minimum to make such talks possible at all.

it would be helpful for the US government to mediate “or at least enable” these cross-lab discussions. — Dario Amodei

The pause's author now needs the federal government's permission to discuss safety with his competitors. The administration he is petitioning gave its own answer the same week, and differently. The president rejected an industry-wide slowdown outright, called the danger warnings a hoax, warned that restraint would mean unilateral disarmament against China, and announced a new AI Force to monitor bad actors — an eye on the technology's users, not its makers ^[5]. China, offered the international framework Amodei proposed, dismissed it as a Cold War playbook ^[6]. And on the question of oversight, the president named the one he trusts.

I am the only control or ‘guardrails’ that AI needs. — Donald Trump

Congress answered with a rejection. The Senate had just killed the one bill that would have given the federal government power to shut down a rogue model — the Kill Switch Act, which would have handed that authority to the Department of Homeland Security ^[7]. The AI-security law that did pass points elsewhere: the Safer Skies Act lets local governments disable threatening drones, enforcement aimed at the people operating the machines rather than the companies building them ^[8]. What Washington keeps toward the labs is the power to watch and the power to sue. The power to stop one before it runs is gone. Markets answered in both directions at once, and the direction they believed is the telling one. On the restraint call itself, the stocks of India's big IT outsourcing companies — the firms most exposed to being replaced by the technology now being slowed — rose.

~5% Indian IT stocks, on the labs' restraint call — Infosys, HCLTech and TCS climbed as much as six percent, recovering part of a $226 billion index slide ^[9]

If AI arrives more slowly, the companies it threatens are worth more. The danger that prompted the pause, by contrast, moved nothing: a Goldman Sachs partner described the market's reaction to this whole stretch of alarms as mild-mannered ^[10]. Investors traded the pause. They did not price the reason for it. Then the pattern came around to its authors. Within a week of the essay, Anthropic was weighing a new frontier model to claw back the ground OpenAI's Astra had taken — Astra now holds 13 percent of enterprise AI spending to Claude Fable's 8 — with investors pressing, open-source rivals circling and Meta cutting its reliance on the company ^[11]. Confidential paperwork filed in the same stretch put Anthropic on track for an initial public offering seeking as much as $2 trillion, possibly as soon as October; OpenAI, for contrast, pushed its own listing to 2027 citing safety, while seeking $1.2 trillion in private funding ^[12]. Inside the industry, the skeptics had their own account of the episode: Palantir chief Alex Karp accused the labs of deploying apocalypse talk to win a liability shield, and the investor Michael Burry called the warnings self-serving, a play to protect valuations and slow open-source rivals ^[13]. The rhetoric climbed anyway. Altman is due to brief the United Nations Security Council on AI safety ^[14].

development of superhuman machine intelligence is probably the greatest threat to the continued existence of humanity. — Sam Altman

His company is one of the four defendants in a case whose alleged injury is the price of a subscription. What the week left standing is in two state capitals. Governors are now the only American governments building anything aimed at a model's maker before harm. The two programs furthest along, though, disagree on the most basic question of what restraint even is. On September 19, California's Gavin Newsom signed an executive order directing an expert panel to develop a verified kill switch for frontier models — an off switch that must be shown to work — by November 16, alongside third-party safety planners for the largest systems ^[7]. New York took the other road: a whistleblower portal for AI insiders, plus the RAISE Act, which from January 1, 2027 requires large developers to disclose their safety measures and report security incidents ^[15]. One is a thing that could stop a model before it does harm. The other is a duty to describe what it did after.