Pacing becomes a price-fixing claim

When the heads of the largest AI companies called for slowing the frontier down, the objection from Washington and Beijing was strategic. Both capitals said no within 48 hours. The objection that arrived on Friday is narrower and, for the companies, more awkward: that agreeing to go slower together is illegal.

Four people who pay for ChatGPT, Claude, Grok and Gemini subscriptions filed a proposed class action in the Northern District of California against Anthropic, OpenAI, SpaceXAI and Google DeepMind, alleging that their public calls to decelerate amount to an unlawful agreement among competitors. The proposed class is everyone in the country who pays for those products.

The evidence pleaded is, unusually, public. Dario Amodei's essay on 12 September argued for industry-wide cooperation on slowing advances in favour of stronger safety work, and the other chief executives said much the same thing within the day. A July statement signed by senior employees acknowledged the "intense competitive pressure not to unilaterally slow" development, which the complaint reads as evidence that unilateral restraint was not viable and coordination was therefore necessary.

That is the shape of a classic Section 1 case. Competitors may each decide independently to sell a worse or slower product; what they may not do is agree with each other to do it. The complaint's theory is that subscribers paid for a service whose improvement was deliberately restrained by agreement.

The defence writes itself and is not weak. Calling publicly for regulation is petitioning, which is protected; saying in an essay that an industry should be paced is speech, not a contract; and none of the companies has actually slowed anything that anyone can point to. Proving an agreement from parallel public statements is the hardest thing in antitrust.

The plaintiffs' lawyer is not arguing it quietly. Nick Rowley, the lead attorney, said AI "will quickly spin out of human control and could kill us all if we allow AI safety and protocol ... to be controlled by private self-serving agreements between the world's most powerful 'for profit' technology companies", which is a safety argument pressed into service as a consumer-harm claim.

That tension runs through the case. The complaint's stated concern is that the companies are too powerful to be trusted with safety decisions; its legal theory is that consumers were denied faster AI progress. A jury would be asked to award damages for the slowing that the same filing says was self-serving rather than genuine.

None of the four companies had commented when the suit was reported. They face a familiar problem: any statement defending the pacing call as principled restraint is a statement about conduct the complaint is trying to characterise as an agreement.

The commentary saw this coming. The Guardian argued last week that letting AI firms coordinate to pace the frontier was a revival of an old corporate manoeuvre for getting around competition law, and the point cuts both ways — if the coordination is real, it is a cartel; if it is not real, the safety commitment is rhetoric.

What makes this more than a nuisance filing is the timing. Washington has already refused the antitrust waiver that would have made coordinated pacing lawful. Without that waiver, every joint safety initiative the industry announces now carries litigation risk, and the labs must choose between coordinating visibly and coordinating not at all.