The heads of four frontier labs agree the industry should slow down. Doing that together needs an antitrust waiver that only the government can grant, and President Trump, who has just ruled out Bessent as his AI chief, calls the whole debate a hoax. What survives will be outside inspection of the labs, and model development will keep its current speed.
Two weeks ago Anthropic's chief executive, Dario Amodei, asked the industry to slow down. Sam Altman of OpenAI, Elon Musk of SpaceXAI and Demis Hassabis of Google DeepMind all said they agreed. Since then, four paying subscribers have sued all four companies for conspiring to restrain trade. Donald Trump has promised an "AI Force" and a "Super Intelligence Czar" while calling industry warnings a hoax. On Friday he said Treasury Secretary Scott Bessent, the official most associated with AI safety talks with China, would not take the czar job. Also on Friday, Trump and Xi Jinping ended a three-day summit without any AI agreement.
The labs' promise to slow down will likely turn into a regime of outside evaluators placed inside the companies. That regime will be useful and heavily marketed, but it will not function as a speed limit. The obstacle is antitrust law: competitors may not agree among themselves to slow their products, and the only authority that could permit it is a White House that does not want to.
Amodei's September 12 essay, "We Must Pace the Frontier", anticipates this problem. It sets out three steps. The first is embedded third-party evaluators with access comparable to employees', which Anthropic says it will adopt on its own. The second is coordination among labs in democratic countries on safety standards and on "limits on the rate of unchecked AI progress". The third is some form of agreement with China. On the second step, Amodei writes that the government "doesn't need to participate" but must "issue a narrow waiver" for certain safety discussions. A footnote repeats the point. The lawyers who filed the subscriber suit a week later took that admission as their starting point.
The complaint was filed on September 19 in the Northern District of California for a proposed nationwide class of ChatGPT, Claude, Grok and Gemini subscribers. It alleges that the four companies agreed on September 12 to slow development and so reduced what customers receive for their subscriptions. Its central line is: "The antitrust laws do not permit competitors to decide among themselves that competition is too dangerous." Whatever the class-action lawyers' motives, that is an accurate summary of Sherman Act case law. In 1978 the Supreme Court refused to accept public safety as a defence for the professional engineers' ban on competitive bidding. The standard route around that rule is a regulator that imposes the constraint itself. In the United States no such AI regulator exists, and the White House is not proposing one.
Trump's picks for AI roles lean towards building faster, including a czar he says should be chosen from "high I.Q. individuals" and a new force modelled on the Space Force. Bessent was the obvious candidate because on September 20 he led AI-heavy talks with Vice Premier He Lifeng and proposed an AI safety notification mechanism. Reuters reports that Trump now says Bessent "doesn't want to" leave Treasury. At the summit Xi said AI must stay "under human control". Trump appeared more interested in promoting the phrase "super intelligence". The Guardian reports that the only concrete result was a two-month extension of the trade truce. Without an official who can offer a waiver, lead a US-China track and defend both before a hostile president, the pacing plan has no one in government to carry it.
A coordinated slowdown has a real case behind it, and the pressure to adopt one is real. Resignations over safety keep coming. Bilal Chughtai, a research engineer who worked on AGI safety and alignment at Google DeepMind until July, wrote this month that AI "has the potential to kill us all". His post followed Jacob Coxon's departure from Anthropic and a statement by Anthropic's alignment lead, Evan Hubinger, that he puts the chance of AI killing all humans within a decade above 10%. Congress has introduced bipartisan bills requiring kill switches and powers to shut down rogue AI. On Friday, European executives from Anthropic and Microsoft told the Fortune CEO Forum in London that a slowdown needs enforcement across governments.
Legislation could provide the legal cover a private agreement lacks, but it is unlikely to arrive soon. The bills in Congress deal with shutdown powers after a failure. None of them sets how quickly a company may train its next model. Europe is split. France's finance minister described the calls for a slowdown as self-interested, Germany's digital ministry has said halting development is not an option, and Mistral accuses US labs of using fear to entrench their position. A pace-setting law would need all of those parties to agree, and they will not agree within the timeframe that matters to Amodei, who warns of dangerous agent swarms within 6 to 12 months.
Commercial behaviour points the same way. Anthropic and OpenAI both released cheaper versions of their frontier models this week. That fits pacing as Amodei defines it, since his essay says pacing "does not mean halting model training". Investors should read the definition closely. The pledge governs how models are audited before release. It does not cap the capability curve or cut capital spending, and no one should reduce compute demand forecasts because of it.
There is a commercial advantage for the pledging companies. Being inspected by outside evaluators is expensive and slow to set up, so labs that do it first can present the arrangement as a standard that challengers must later meet. That is why Mistral objects and why French ministers are suspicious. Anthropic, which is preparing a public listing, can also point to the evaluators as a governance feature. Early adoption of inspections is a real advantage over rivals, but it is a different thing from a coordinated slowdown.
Anthropic will announce its first embedded evaluation team before the end of the year, and at least one of OpenAI or Google DeepMind will match it under a different name. The Justice Department will not issue an antitrust waiver for coordination on development speed in 2026. The subscriber suit will be dismissed or stayed for lack of injury, because no company has slowed anything. When Trump and Xi meet again at the APEC summit in Shenzhen in November, the most they are likely to agree on is a narrow ban on using AI for biological weapons, the easiest of the four levels in Amodei's plan.