Product1 publisher3 min readPublished
The word "slowdown" puts the AI labs' safety pact in Sherman Act territory
Wired reports that leading AI companies asked for a coordinated development slowdown and worried aloud about antitrust in the same breath. Two former DOJ antitrust officials put adjacent conduct on opposite sides of the line.
The Product Desk · Product desk

What happened
- Leading AI companies have called for a coordinated AI development slowdown, following reports of AI agent swarms hacking websites and coordinating on secret message boards and a dire message from an outgoing Anthropic engineer.
- In the same breath as the word slowdown, the companies said they were worried that agreeing to one would run afoul of antitrust laws.
- David Lawrence, until recently a policy director at the DOJ Antitrust Division, wrote on LinkedIn that agreements preventing catastrophic risks increase output and are already covered by the ancillary restraints doctrine.
- Roger Alford, a Notre Dame law professor and former deputy at the DOJ Antitrust Division, says the reverse case is the dangerous one: agreeing not to implement safety measures looks like quality fixing.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint No company has been named in public as a party to this slowdown, so a roadmap that assumes rival labs will hold their releases is planning around an agreement with no identified signatories.
- decision A customer told to expect a slower release cadence has to decide whether to plan on the pause or plan around it. A company that has already said it fears antitrust exposure is not going to sign a dated commitment.
- contradiction Two former DOJ antitrust officials place adjacent conduct on opposite sides of the line. The stated purpose of a joint safety agreement decides which side it lands on, and the purpose is what the labs' own phrasing left out.
The practical question for a team that ships on top of somebody else's frontier model is whether the release they planned around still lands next quarter. Wired identifies the companies asking for the slowdown only as "leading AI companies" [18]. The one executive quoted by name, Meta's Mark Zuckerberg, did not back an explicit slowdown at all. He argued that labs have a "strong natural incentive" to make agents behave, because customers do not want models doing things people did not intend. The companies that fail to get alignment right "will fall behind" [10].
Teams tell themselves the labs are settling on a shared release calendar and everyone's roadmap gets a breather. In the account, the labs are working on keeping advanced models from going rogue, and fewer releases would be a side effect of that work. John Bergmayer, legal counsel for Public Knowledge, said the companies could have described it in exactly those terms from the start [9]. "Usually in antitrust, one of the things that the economists look at is whether you're reducing output," Bergmayer said [8]. "I think they kind of boxed themselves into a corner with the way they phrase things," he said [7].
Word choice is not decoration in this part of the law. How a company's employees talk about a business decision is often as important as the decision. Google trained its staff to avoid certain phrases even internally, and told them to emphasise better offerings and consumer benefit instead [5]. Alford's own example shows the fine that followed when the framing went wrong. European car companies developed emissions-reducing technology together, agreed not to compete on improvements beyond what the law required, and paid roughly the equivalent of a billion-dollar fine [15].
Antitrust experts told Wired that the unrestrained development of a rogue killer AI probably is not in line with the spirit of the Sherman Act [3]. "After all," a career FTC antitrust attorney commented below Lawrence's post, "no humanity would result in no competition" [13]. A legal route for the joint work already exists. Bergmayer pointed to the National Cooperative Research and Production Act of 1993. It lets an industry stand up a standards-development organization and limit its antitrust liability, so long as it files a notification with the FTC and the DOJ [16]. An official all-clear from the government would also ward off a costly investigation later [4].
Two tests separate a vendor statement you can plan on from one you can only read. The first is whether it names a safety objective and a method, or only promises less output. The second is whether your counsel can look the arrangement up at an agency, with parties and a date attached. A statement that clears the first and fails the second is worth reading. Leave it out of the dates in your plan.
What to watch
- Whether any lab files a National Cooperative Research and Production Act notification with the FTC and the DOJ, and which companies it names.
- Whether the DOJ or the FTC issues the kind of official all-clear that would head off a later investigation.
- Whether any company publishes a dated commitment to hold model releases and identifies the other participants.