Can Antitrust Law Criminalise AI Safety Coordination?

A Lawsuit That Turns Safety Into a Liability

The AI cartel lawsuit filed on September 18, 2026, accuses Anthropic, OpenAI, SpaceXAI, and Google of coordinating to deliberately slow AI development — violating U.S. antitrust law and degrading the competitive value of paid subscriptions that millions of users rely on daily.

The core accusation is simple, even if its implications are not. The plaintiffs allege that these four firms coordinated to deliberately slow AI development, violating U.S. antitrust law and degrading the competitive value of the paid subscriptions that millions of users rely on daily. ChatGPT, Claude, Grok, and Gemini subscribers, the suit argues, are paying premium prices for a product whose progress has been artificially throttled.

What makes this case structurally unique is the weapon the plaintiffs have turned against the defendants: the defendants' own moral language. Safety, caution, responsibility — these were the watchwords these companies used to build public trust and political goodwill. If coordination in the name of safety constitutes illegal output restriction, then every public statement about existential risk becomes potential evidence. The lawsuit seeks to represent a nationwide class of paid subscribers — a constituency broad enough to carry real financial weight.

The Sherman Antitrust Act does not include an exemption for good intentions. And that is the contradiction the Northern District of California will now have to resolve: whether the most powerful companies in commercial history can agree, however informally, to slow the engine of civilisation — and call it ethics.

Six Days From Essay to Alleged AI Antitrust Violation: The September 12 Timeline

The oldest question in competition law is deceptively simple: when does coordination become conspiracy? Six days before the lawsuit was filed, that question acquired a precise timestamp.

On September 12, 2026, Anthropic CEO Dario Amodei published an essay proposing industry-wide coordination to decelerate AI development on safety grounds. The same day, Sam Altman of OpenAI, Elon Musk of SpaceXAI, and Demis Hassabis of Google publicly signalled their agreement. Plaintiffs argue this simultaneity is not coincidence — it is the evidentiary core of their cartel allegation.

Amodei did not stop at the essay. He formally requested a "narrow antitrust waiver" from the U.S. government, seeking legal cover for safety-based industry conversations among direct competitors. That request matters legally: it implies Amodei understood that without such a waiver, the coordination could violate antitrust law. If the architects of "safety" sought an exemption, they arguably acknowledged the legal exposure themselves.

The plaintiffs' theory, filed under the Sherman Antitrust Act, holds that the defendants violated competition law by agreeing to decelerate AI advancement. Under U.S. antitrust doctrine, even output restrictions framed as public goods can constitute illegal collusion if competitors act in concert. The legal threshold is critical here: courts distinguish between parallel conduct — where firms independently reach similar conclusions — and explicit agreement, which triggers per se liability.

The plaintiffs' challenge is that September 12 produced no leaked memo, no signed document. What it produced was four powerful executives, in the same sector, expressing the same position, within hours. Whether that constitutes a meeting of minds or a coincidence of conscience is precisely the question the Northern District of California must now answer.

The Subscriber as Plaintiff: Paying for Progress That Was Withheld

Consider the quiet absurdity here: a consumer pays a monthly fee for access to the frontier of human cognition, and the very companies collecting that fee may have agreed, privately, to keep that frontier where it was. The plaintiffs' legal theory is novel but structurally tight. Artificial slowing of AI progress constitutes a quality reduction for paid users, meaning that subscribers to ChatGPT, Claude, Gemini, and Grok received less than the competitive market would have otherwise delivered.

This is not a claim about deceptive advertising. It is a claim about withheld value. If four dominant providers coordinate to suppress the pace of model improvement, the subscriber's monthly fee no longer buys a share of a competitive race. It buys access to a managed plateau.

Lead attorney Nick Rowley sharpens the argument with a warning that cuts in the opposite direction: that private, self-serving safety agreements carry their own catastrophic risk, potentially allowing AI to spin out of human control precisely because oversight becomes a closed-door arrangement among incumbents rather than a transparent regulatory process. The legal strategy thus holds two tensions simultaneously: coordination harmed consumers, and unaccountable coordination could harm everyone.

Sam Altman's public position complicates any defence that coordination required formal legal cover. He stated plainly that OpenAI does not believe it needs to wait for an antitrust exemption to begin safety coordination. If no exemption was considered necessary, the question for the court becomes what legal basis the defendants would invoke. The subscriber, paying for progress, is left asking the same question as the judge: what, precisely, was the justification for slowing down?

The Sherman Antitrust Act does not include an exemption for good intentions.

Washington's Divided House: The AI Czar Versus the Antitrust Purist

The Capitol does not speak with one voice on artificial intelligence. Inside the same legislative week that produced the California lawsuit, two sharply opposing instincts collided in Washington, each staking a claim on the future of the technology, each incapable of fully accommodating the other.

Donald Trump's response was immediate and categorical. He labelled efforts to limit AI development a "conspiracy," dismissing the safety rationale with the bluntness of a man more interested in geopolitical arithmetic than existential risk. His counter-proposal: an AI task force, a designated AI czar, a coordinated national instrument designed not to slow the technology but to ensure that when the next inflection point arrives, it arrives under an American flag before a Chinese one. The framing was nakedly strategic. Safety, in this construction, is not a public good; it is a competitive liability.

Then came Josh Hawley. The Republican senator, no ally of Silicon Valley, delivered the line that most precisely defined the structural fault line: "There is no world in which I would agree to give the most powerful companies in the history of the world an exemption from antitrust laws." Short. Definitive. It closed the door Amodei had tried to open with his narrow waiver request. Here was the contradiction made visible — a White House accelerating into deregulation on one side, and a senator raising the oldest market-governance shield on the other. Two Republicans, one courthouse filing, and a policy vacuum that neither antitrust doctrine nor executive order has yet been designed to fill.

The Architecture of Concentration: Chips, Capital, and Regulatory Jurisdiction

Antitrust law has always struggled with markets that consolidate before regulators can map them. The AI hardware stack presents a more acute version of this problem: by June 2024, the DOJ and FTC had formally divided oversight of Nvidia, Microsoft, and OpenAI between themselves — a jurisdictional partition that acknowledged, implicitly, that no single agency could surveil the full vertical chain of AI power. That division was an admission of structural reality, not a remedy for it.

Compare this to the European response. French competition authorities raided Nvidia over its dominance in the AI chip market, a move that signals cross-border antitrust convergence rather than a coordinated transatlantic strategy. The EU proceeds through raids and investigations; the US proceeds through jurisdictional carve-outs. Both approaches reveal the same underlying anxiety: that a single chokepoint in the chip supply chain can determine who builds the future.

The capital layer compounds the concern. Microsoft's $13 billion investment in OpenAI created a secondary concentration vector well before the September 2026 lawsuit was filed. Then, in 2025, Nvidia and OpenAI signed a memorandum outlining $100 billion in planned infrastructure investment — a figure that dwarfs most sovereign technology budgets. If the alleged behavioural coordination described in the lawsuit represents the soft infrastructure of collusion, these capital flows represent its hard architecture. The legal question is no longer whether concentration exists. It is whether any regulatory jurisdiction, divided or otherwise, is architecturally equipped to contain it.

What the Courtroom Cannot Decide: Extinction Warnings, Clinical Proof, and the European Contrast

Jacob Coxon did not leave quietly. His resignation from Anthropic, paired with a documented warning that AI could drive human extinction by 2030, places the existential stakes in terms no antitrust filing can adequately contain. A court can adjudicate market harm. It cannot adjudicate civilisation-scale risk. These are not the same problem, and treating them as interchangeable is precisely what makes this lawsuit both urgent and inadequate.

The data on AI's public benefit is not speculative. Peer-reviewed research published on PubMed confirms that AI already outperforms non-specialists in diagnostic accuracy for pulmonary function tests — a measurable, clinical gain with direct patient outcomes. If deceleration delays the diffusion of such capabilities by even one year, the consumer harm is not merely a degraded subscription. It is counted in misdiagnoses. Yet Semantic Scholar research flags automation bias as a persistent systemic risk in human-AI collaboration, a liability that neither faster nor slower deployment eliminates. The evidence does not resolve cleanly into a pro-acceleration verdict.

The European dimension cuts deeper still. The EU AI Act imposes mandatory safety protocols on high-risk AI systems, effectively requiring the kind of coordinated compliance behaviour that U.S. antitrust law now scrutinises as cartel conduct. This cross-border regulatory paradox is not theoretical: if safety coordination is legally compelled in Brussels and legally prosecuted in San Francisco, one jurisdiction will ultimately set the global norm. The AI cartel lawsuit, then, is not merely a dispute over subscriber value or market competition — it is a proxy battle over who holds the authority to set the pace of the most consequential technology in human history. The question no judge in the Northern District of California can answer is the one that matters most: who decides the pace of civilisation, and by what authority do they hold that role?