Four named plaintiffs. One paid subscription each, to ChatGPT, Claude, Grok or Gemini. On Friday they filed a proposed class action in the Northern District of California claiming Anthropic, OpenAI, SpaceXAI and Google made an illegal deal to slow the pace of their own development. The complaint dates the agreement to September 12, the day Dario Amodei's essay went up and Altman, Musk and Hassabis each answered it in public, in agreement, within hours.
The instinct is to file this under nuisance. It is not. It is the first time the slowdown's form, rather than its motive, has been put in front of a judge, and the form is the problem.
Section 1 of the Sherman Act does not care why competitors agree to produce less. An agreement among rivals to restrict output is the textbook per se violation; the court does not weigh whether the restraint was wise. The labs' lawyers know this, which is why the plaintiffs' best exhibit is not the September essay. It is the July statement, signed by senior staff across the labs, that acknowledged the "intense competitive pressure not to unilaterally slow" and asked the government to back a global effort to slow automated AI development. That sentence is a confession and a plea in one breath: we cannot do this alone, and we cannot do it together either, so please make us.
The plaintiffs don't object to the companies individually deciding to slow their own progress in favor of safety. They instead argue in the complaint that antitrust laws forbid them from taking the "shortcut" of agreeing to "substitute collective restraint for individual accountability."AP Wire
That framing is sharper than it looks. Nobody is being sued for pacing their own model. Anthropic can ship Claude whenever it likes. What the complaint attacks is the handshake, and the handshake is the whole design: the labs said out loud, per Reuters' reconstruction of the ten days, that no one of them could pause while the others kept going. A slowdown that only works if everyone signs is, in the statute's vocabulary, a cartel with a safety rationale. And safety rationales have a bad record in this courtroom. In 1978 the Supreme Court told the National Society of Professional Engineers that a ban on competitive bidding was illegal even if the engineers sincerely believed bidding produced unsafe bridges. The Court's point was not that the engineers were wrong about bridges. It was that the antitrust laws had already decided competition, not the competitors, gets to set the pace.
A slowdown that only works if everyone signs is, in the statute's vocabulary, a cartel with a safety rationale.
The honest objection is that this suit is a long shot. Four subscribers, a proposed class, a theory of harm that requires a jury to believe they were injured by getting a slower Claude. Public statements that happen to agree are not the same as an agreement; courts have been clear that parallel conduct alone is not conspiracy, and every defendant here will say it decided to slow on its own and simply said so on the same afternoon. The plaintiffs may never get past a motion to dismiss.
Grant all of it. The suit still does its work before it loses, because the risk it names is now priced. Every joint statement from here on goes through counsel. Every coordination call becomes a discovery target. The labs' general counsels will do what general counsels do: move the coordination to the only room where it is legal, which is one with a government at the head of the table. That is the room the July signatories asked for. The complaint does not stop the slowdown. It removes the option of doing it on a handshake, and leaves two: competition sets the pace, or a regulator does.
The labs spent a week arguing about whether they meant it. Four subscribers just made the sincerity irrelevant. A voluntary slowdown was never a stable state. It was the intermission before someone decided whose law applies.