AINEWS
Jonathan Kanter, who led the DOJ's antitrust division through the Google search case, says the law already covers catastrophic risk coordination. The harder question is who defines 'safety.'
Jonathan Kanter ran the Justice Department's antitrust division from 2021 to 2024. He opened the Google search case and brought landmark cases against Apple and Live Nation–Ticketmaster. On a recent episode of The Verge's Decoder podcast, he was asked whether AI companies need an antitrust exemption so they can coordinate on safety. His answer, in effect, was no, with a condition attached.
Kanter is now a Distinguished Teaching Professor at Washington University in St. Louis Law and a professor of technology policy at Carnegie Mellon. He was confirmed as Assistant Attorney General for the Antitrust Division in 2021 on a bipartisan 68–29 Senate vote and ran an 800-person division with a budget above $300 million. The cases he led are the recent reference points for how U.S. enforcers treat coordinated conduct by dominant firms.
The Decoder episode opened with a striking number: the chance that AI kills everyone is, according to "other researchers" cited in the segment, greater than 10 percent. That figure is scene-setting, not a load-bearing claim. The structural question Kanter was answering is narrower and more useful than the headline suggests: whether two or more frontier AI labs can lawfully sit in a room, share a third-party evaluation of a model's dangerous capabilities, and agree on a coordinated response, without becoming the kind of cartel U.S. antitrust law exists to break.
Under current law, Kanter's answer is probably yes, with a familiar set of conditions. Information exchanges that are reasonably tailored to a legitimate safety purpose, open to non-firm-affiliated researchers, and not extending to pricing, output, or hiring, have long lived inside antitrust law. The harder question, in his framing, is upstream: who gets to define what counts as "safety" before coordination starts. If the definition is set by the firms themselves, and the coordination is broad, the visible result looks indistinguishable from a cartel with a safety press release stapled to it. If the definition is set externally, by a regulator, a nonprofit evaluator, or a pre-agreed protocol, the same coordination is plausibly lawful.
Kanter's framework is being tested, in real time, by a pair of unusual political allies. Libertarian investor and former Trump White House AI czar David Sacks has approvingly retweeted Lina Khan saying there is no need for an antitrust exemption for AI safety coordination. Sacks is not known for defending progressive trustbusters. Khan is not known for siding with the AI industry. The fact that they land on the same answer, that exemption is not needed but coordination is fine if the safety definition is external, is one of the few clean signals in the current debate that the antitrust line is real, not a partisan beat.
The critique Kanter has to answer is not theoretical. The Verge's segment notes that researchers at Anthropic and Google DeepMind have publicly quit over what they describe as inadequate attention to model risk. Skeptics read the exemption request as three things. Regulatory capture: a small number of large labs defining "safety" in a way that excludes smaller competitors. Attempted cartel formation: shared restraint dressed up as a safety pact. IPO timing: firms asking for coordination cover ahead of public listings, when investor pressure on deployment timelines is highest.
Kanter's prior DOJ remarks and his April 2025 essay "Competition and freedom" for WashU Source are consistent with the Decoder position. He treats competition policy as a load-bearing tool for innovation, not a friction to be waived when the technology is consequential. The China dimension, raised in the interview, sharpens the point: in a world where the U.S. and China are racing on frontier capability, the legal test for what is and is not coordination has to be drawn before any safety pact is signed, not after.
The reader-facing version of Kanter's framework comes down to four checks. Is the safety definition set by a third party, or by the firms themselves? Is the information shared narrowly tailored to catastrophic-risk evaluation, or does it extend to pricing, hiring, or model release schedules? Is participation open to independent researchers, or limited to incumbents? And is there a credible exit, a way for a lab to decline to coordinate without being locked out of the evaluation that the rest of the industry uses?
If the answer to all four is the safe one, the pact is plausibly lawful. If the answer to any is no, the same pact reads, to an enforcer, as a coordination agreement with a safety label. The visible shape of the current exemption request, in Kanter's telling, is producing exactly that ambiguity. The fix is not a new law. It is a clearer, externally anchored definition of what counts as "safety" before the firms sit down.