The antitrust pause alibi is the real AI policy story this week

People close to OpenAI tell WIRED the company has asked members of Congress, in recent weeks, whether orchestrating an industry-wide slowdown on frontier development would survive antitrust law. Last weekend, OpenAI chief scientist Jakub Pachocki argued publicly for coordinating to slow future development and said he expects voluntary slowdowns until shared safety bars exist. OpenAI did not comment to WIRED.

Last week we said the citeable story was the race dynamic: people stay in the race because they don’t trust anyone else to stop. This week the same dynamic put on a suit and walked up to Congress.

We’re still for useful AI. Tools that write, code, diagnose, and speed up science are worth shipping under measurable controls. That belief doesn’t let this week’s policy move pass as boring legal hygiene.

Here’s the sentence we’d want another desk to steal: labs asking Congress whether slowing down together would even be legal is an admission that solo braking loses — and even a legal pause is only as real as the cheaters you can catch, which fines probably won’t.

Two things can be true at once. Legal fog is real. Nicholas Felstead has argued a coordinated pause could look like output restriction under the Sherman Act, depending on the agreement’s details — and that even collaborations that would survive scrutiny can be chilled by uncertainty alone. John Schulman, an OpenAI cofounder now at Thinking Machines, called the antitrust citation “fake” if the work is a joint pacing proposal rather than a prohibited agreement, and told OpenAI and Anthropic to stop feuding and write one.

We’re not picking a team between “law blocks everything” and “law is pure theater.” We’re naming the bind: if solo braking loses the market and the China narrative, then every lab has a structural reason to wait for a shared stop — and antitrust uncertainty is the cleanest public reason to wait without saying the commercial part out loud.

Collusion cuts both ways. A lawful joint slowdown is, in antitrust language, coordinated restraint. The mirror problem is the illegal kind: if the visible labs stop, what keeps a rival — or a well-funded shop nobody invited — from training in the background? We’re skeptical that fines and civil penalties will be harsh enough to discourage that. Frontier advantage is the whole market. A penalty that looks large on a press release can still be cheaper than falling a generation behind. So even if Congress writes the hall pass, the race dynamic doesn’t vanish. It goes dark. Honest participants eat the pause; cheaters keep the weights.

That’s why “we’d slow down if we could legally coordinate” is incomplete. The harder question is verification: who audits the clusters, who sees the training runs, and what happens when the expected value of cheating still beats the expected value of getting caught.

The same week, U.S. agencies named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for industrial-scale distillation of frontier models, and Anthropic alleged nearly 200 million exchanges tied to distillation campaigns — including about 151 million attributed to an Alibaba effort peaking near three million a day. China’s Foreign Ministry called the U.S. charges groundless. We’re not treating those attributions as courtroom verdicts. They are the political weather a voluntary U.S. slowdown has to live in.

In July, bipartisan lawmakers introduced the Collaboration on Adversarial Threats and Security Risks Act (House H.R. 9914 / Senate companion). It would carve a path for labs to share information on covered AI security risks and, with notice to DOJ, coordinate delays in development or release aimed at those risks. The House version was referred to Judiciary and has not been taken up. AI Policy Network’s Caleb Knapp told WIRED anything binding may wait until after the midterms. Paul Christiano joining OpenAI’s Foundation Board is a governance signal — not a substitute for a pacing mechanism the industry will actually use.