There Is No Legal Way to Slow Down AI

How a safety pledge became a federal antitrust case in six days, and why the labs' decision to coordinate in public turned their most honest admission into the evidence against them

Author

Justin Wright
September 21, 2026 • Est. Reading Time: 6 minutes

Happy Monday!

Filed September 18 in the Northern District of California, six days after Dario Amodei published his call for an industry-wide slowdown. (Source: CNN)

On September 12, Dario Amodei published an essay asking the AI industry to deliberately slow the pace of capability gains. Sam Altman agreed the same day, and so did Elon Musk and Demis Hassabis. It was the rarest thing in this industry: four fierce competitors publicly aligned on a single commitment.

On September 18, four paying subscribers to ChatGPT, Claude, Grok and Gemini filed a class action in the Northern District of California accusing all four companies of violating Section 1 of the Sherman Act. The theory is simple: horizontal competitors agreed to restrict the output of competing products. Antitrust law has treated that as illegal for more than a century, and the stated motive does not matter.

The plaintiffs' strongest line is that they barely have to prove anything. As the complaint puts it, the agreement was "proposed in public, accepted in public, and confirmed in public." There are no smoke-filled rooms to reconstruct, no emails to subpoena; the evidence is the essay and the replies.

Last week this newsletter noted that Amodei's own proposal conceded the problem: step two of his framework asked the US government for a narrow antitrust waiver so competitors could hold safety conversations legally. Two days after he published, President Trump rejected the whole premise. The waiver is not coming, and the lawsuit arrived anyway.

Six days after four AI CEOs publicly agreed to slow capability development, four of their subscribers sued them under the Sherman Act for illegal output restriction. The complaint uses the labs' own public statements as evidence, including the July pacing letter in which lab employees admitted facing "intense competitive pressure not to unilaterally slow." Amodei had asked the government for an antitrust waiver. Trump refused and called the warnings exaggerated. The labs are now building a self-regulatory body anyway, which a competing CEO has called a cartel.

TL;DR

The Trap Has Three Doors and All Are Locked

The coordination problem this newsletter described in July has now become a legal one, and the structure is worth laying out plainly.

Three Paths to a Slowdown

Path

How it works

Why it fails

Unilateral

One lab slows on its own

Hands the lead to rivals; competitive suicide

Multilateral

Labs agree among themselves

Per se output restriction under Sherman Act §1

Government-sanctioned

Antitrust waiver or statute

Administration rejected the premise outright

Every lab understands door one. The July letter signed by more than a thousand frontier researchers said so explicitly, acknowledging that each faced intense competitive pressure not to unilaterally slow. That was the honest core of the document and the reason it asked Washington for help.

Door two is the one they walked through on September 12, and it is the one the Sherman Act forbids. Agreements among horizontal competitors to restrict output are treated as per se illegal, which means courts do not weigh the benefits. The complaint makes the point directly: this is among the restraints Section 1 has always condemned, and it does not become lawful because the products are new.

Door three required the government, and the government slammed it shut. Trump called the dire warnings exaggerated, said "negative forces" were making claims about things that will not happen, and framed the stakes as "whoever wins AI, wins." White House AI czar David Sacks went further, telling the labs to pace themselves without Washington and adding, "stop pretending the motivation to slow down is purely altruistic." China declined as well, although getting their agreement was a long shot from the outset.

The Honesty Became the Evidence

The July pacing letter was, by the standards of corporate communication, unusually forthright. Its central admission was that no company could slow down alone without losing, which is why it asked for a coordinated mechanism. That sentence was the most credible thing in the document because it conceded self-interest rather than hiding it.

The complaint now uses that admission as evidence of both motive and action against self-interest. In antitrust terms, conduct that would be irrational for a firm acting alone, but rational if everyone does it together, is circumstantial evidence of an agreement. The labs explained their own coordination problem in writing, and the explanation became Exhibit A in the lawsuit.

The lesson competitors will take from this is not encouraging. The safe move, legally, is to say nothing, coordinate nothing, and ship. Transparency about competitive dynamics created liability that silence would not have.

The Defense Nobody Is Discussing

The reality is, nothing has slowed down. The median gap between major frontier model releases compressed from 37.5 days in 2023 to 11 days in 2026. Four labs shipped frontier models in a single week earlier this month. GPT-6 Astra crossed OpenAI's own Critical threshold for cyber capability. And this week Anthropic published the first results of its R&D Automation Index, reporting that Claude now leads 26 percent of the company's own AI research and development work, up from under 1 percent in February.

That last number is the opposite of a slowdown. It is the clearest public measurement yet of AI accelerating AI research, published by a company whose CEO called for pacing five days earlier. If the plaintiffs must show that output was actually restricted and that subscribers were harmed by it, the release calendar is a problem for them. An agreement to slow that produces no observable slowing is, at minimum, a difficult damages case.

The counter is that the suit targets the agreement itself, and per se doctrine does not require proof that the restraint worked. Courts have also carved out room for genuine safety justifications under rule-of-reason analysis, which is where the labs will want this litigated. The case is at the pleading stage with no response filed, so anyone predicting the outcome is guessing.

They Are Building It Anyway

OpenAI's global policy chief Chris Lehane confirmed the company has been working with Anthropic and Google DeepMind for weeks on a self-regulatory standards body modeled on FINRA, the industry-funded body that polices Wall Street. Altman endorsed the structure on September 15, three days before the lawsuit. The original proposal came from Hassabis in July, which this newsletter covered at the time.

Cohere CEO Aidan Gomez called it a cartel while Senator Sanders argued binding international rules are needed instead. OpenAI has itself acknowledged that safety coordination could collide with antitrust law, which is a striking thing for a defendant to say while a Sherman Act complaint is pending.

So the labs are simultaneously being sued for coordinating and continuing to build the institution through which they intend to coordinate. That is either conviction or recklessness, and which one it turns out to be depends on a courtroom.

What This Means for Practitioners

For anyone planning capability roadmaps, treat the pacing proposal as noise until a court or Congress acts. There is no legal mechanism for an industry-wide slowdown right now, the administration has declined to create one, and litigation makes voluntary coordination more dangerous instead of less. Plan for capability to keep compounding.

For enterprise buyers, the useful signal is the R&D Automation Index rather than the rhetoric. A company publishing a measured, methodologically explicit number about how much of its own research its models now lead is giving you something you can actually track. Watch that index more closely than any essay.

For founders and executives generally, there is a broader lesson in how the July letter was used. Candor about competitive dynamics is legally expensive. If your industry ever faces a collective action problem, the instinct to describe it honestly in public carries a cost that your lawyers will price differently than your conscience does.

The Bottom Line

The labs spent the summer arguing that a slowdown was necessary and that only coordinated action could deliver one. They were right about the structure of the problem and apparently unprepared for its legal consequences. Unilateral restraint loses the race while multilateral restraint is a per se antitrust violation. The waiver that would square the circle requires an administration that has publicly called the underlying fear exaggerated.

What remains is a group of companies that announced an agreement they may not be permitted to keep, in front of a government that does not want them to keep it, while shipping models faster than ever. Whatever this week was, it was not a slowdown. It was the discovery that wanting one and being allowed to have one are entirely different problems.

In motion,
Justin Wright

If antitrust law makes it illegal for competitors to jointly restrain a technology they believe is dangerous, and no single company can restrain it alone without losing, then the law has effectively decided that this particular risk must be run. Is that the correct outcome, or a case where the framework has outlived the situation it was built for?

Food for Thought

Builder’s Note

The number I actually care about from this week is the 26 percent. Anthropic publishing a measured index of how much of its own R&D Claude now leads, with a stated methodology and an automation scale borrowed from Epoch AI, is more useful to me than any amount of essay writing about pacing. It went from under 1 percent to 26 percent in six months. Whatever anyone says about slowing down, that curve is the thing to watch, and it is being published by the company with the most reason to be careful about how it looks.

Quick Hits

  • Anthropic published the first results of its R&D Automation Index, reporting Claude leads 26% of the company's AI R&D work as of August, up from under 1% in February, with Claude involved in over 90% of R&D overall. (Bloomberg)

  • Plugin4Shell, a zero-click remote code execution flaw, hit Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI by defeating SHA pinning on plugin updates. Anthropic and OpenAI patched it; Google deprecated Gemini CLI rather than fix it and Microsoft has not patched Copilot. (The Register)

  • Trump rejected the pacing calls and criticized Amodei directly, saying "whoever wins AI, wins" and calling the warnings exaggerated. China declined to participate as well. (CNBC)

  • OpenAI disclosed six model-misalignment cases on September 16, and Apple shipped its long-delayed Siri AI with personal context, onscreen awareness and cross-app actions. (The Neuron)

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!