A class action builds its Sherman Act case out of the AI labs' own public statements

The sequence the complaint treats as offer, acceptance and confirmation took three days ^[20]. Amodei's essay said government mediation would help and that a narrow waiver would be needed for certain safety conversations ^[4]. Two days later Altman said OpenAI did not intend to wait for an antitrust exemption or legislation ^[6]. The next day OpenAI's policy chief Chris Lehane confirmed the company had already been working with Anthropic and Google DeepMind for several weeks ^[7].

The complaint pleads that asking whether the conduct was lawful shows the companies understood it as coordination, and that proceeding anyway shows intent ^[8]. No such waiver exists, Congress has enacted no exemption, and no agency compelled the conduct, the filing says ^[5]. Paragraph two says the agreement was "proposed in public, accepted in public, and confirmed in public" ^[9]. By The Next Web's account, every factual source in the filing is a named executive's public statement or a press report, and the filing cites no internal documents, confidential witnesses or meeting minutes ^[10].

Teams tell themselves a cross-lab safety forum is a governance artifact. The complaint reads that same activity as contract terms. What it pleads as agreed is a slower rate of capability advancement than each company would choose alone, implemented through limits on training compute and training runs, restrictions on using AI to improve AI, and capability checkpoints ^[11]. The policing mechanism it names is the embedded outside evaluator, described as verifiable pacing ^[12]. Anthropic has begun implementing that programme with Accenture ^[12].

The injured party here is the paying subscriber. Plaintiffs put the four companies' share of US paid subscriptions at 80%, pleaded on information and belief as a floor ^[16]. If the floor holds, a subscriber shopping for an unpaced alternative has at most a fifth of the paid market to choose from ^[21].

The plaintiffs' evidence is an admission about incentives. A statement from employees across the labs in July said each faced intense competitive pressure not to unilaterally slow, and the complaint uses that as evidence of both motive and action against self-interest ^[18]. The weak side is equally visible. The filing concedes the full effect on released products has not yet manifested, and its 29 pages do not identify a slowed release, a cancelled model or an altered training run ^[15]. Musk's acceptance was four words on social media, and Google's endorsed the direction while leaving implementation details open ^[17].

Anyone at these companies with a cross-company call in the calendar can sort the commitments using the plaintiffs' own carve-outs. The complaint disclaims any challenge to unilateral slowing, independent evaluators, lawful safety research, compliance with government requirements, or petitioning Congress, including petitioning for an exemption ^[14]. Its pretext theory is structural: every stated safety objective can be pursued alone, through independent evaluators, or through regulation ^[13]. So the question to ask of a proposed conversation is whether the thing you want only works if a competitor does it too. Work you can do alone, hand to an evaluator, or wait for a regulator to require sits outside what is pleaded; a commitment that needs the other lab to match you is the conduct in the filing.

The filing argues that competition being dangerous is not a defence a court may weigh, because that judgment belongs to Congress ^[22]. The same conduct is reachable in Europe, where Article 101 of the EU treaty prohibits agreements between undertakings that restrict competition ^[19].