Product1 publisher3 min readPublished
Kanter opens the AI exemption debate by questioning whether asking for rules forms a cartel
On Decoder, former DOJ antitrust chief Jonathan Kanter takes up the frontier labs' request for permission to coordinate on safety. His first point is that OpenAI, Anthropic, Google DeepMind and Elon Musk do not much like each other.
The Product Desk · Product desk
What happened
- The Verge's Decoder published the first of a two-part series with Jonathan Kanter, the former head of the Justice Department's antitrust division under Biden, now a law professor at WashU.
- The show's introduction says the CEOs of the AI companies have called for slowing development and for antitrust exemptions so they can all coordinate with each other on safety issues.
- Kanter's first move in the interview is to question the notion that asking for regulation necessarily forms a cartel, noting that the frontier labs do not necessarily like each other.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- contradiction The accusations in the introduction and the opening words of the interview point in opposite directions, so a policy team quoting Kanter to support the cartel charge is quoting the host and not the guest.
- decision A safety team that wants to hand findings to a rival lab has to decide which of those steps it can take on its own authority, because the immunity being discussed does not exist.
- exposure With Khan and Sacks on the same side of this question, a lab that agrees with a competitor on release timing has no political cover from the antitrust left or from the administration's tech wing.
Picture the safety engineer who finds a jailbreak that also works on a competitor's model and writes to her own counsel before sending it over. Nothing in the law has to change for counsel to answer that. Coordination, the word the Decoder introduction uses, describes the other category of action: steps that only take effect if a rival takes them too, on the same schedule [3].
Kanter starts on the premise rather than the policy. "Let's just start with the very notion that asking for regulation necessarily forms a cartel," he said, then named OpenAI, Anthropic, Google DeepMind and Elon Musk as parties who "do not necessarily like each other" [10][11].
His description of the state of play is about roads. "The state of the world right now is like we've invented cars and trucks, but we have no lines on the road, no traffic lights and no stop signs and no speed limits," Kanter said [9]. He then split the problem into what companies should be doing and what the government should be doing [16].
The accusations catalogued in the show's introduction include regulatory capture and cartel formation, plus a suggestion that the companies are looking for a way out from investor pressure ahead of their IPOs [4]. The introduction attributes the exemption request to the CEOs collectively and does not name a company, a letter or a filing [14]. Until one of them publishes the text, a policy team cannot scope which conduct the immunity would cover.
Kanter's standing here comes from his docket. He brought cases against Google, Apple and Ticketmaster as antitrust chief, won the Google and Ticketmaster matters, and the Apple case is still live in the Trump DOJ [7]. Two of the three named cases resolved in the division's favour; one is unresolved [8]. He is out of government now, a point he enjoyed on air: "I can say whatever I want. I feel so free." [13]
The politics are less split than the safety debate suggests. David Sacks, the avowed libertarian and former Trump AI czar, has been approvingly retweeting Lina Khan's position that no antitrust exemption is needed [12].
Two tests sort a safety roadmap into the pile that needs none of this and the pile that does. The first is whether the measure takes effect without a competitor's agreement: a red-team result, a refusal policy, a model card, a deployment gate. The second is whether you could publish it tomorrow under your own name and still get the benefit. Work that clears both tests is work you can ship on Monday. Work that fails the first is what an exemption would cover, and also what a plaintiff would later read as an agreement between competitors.
The stated stakes sit on the other side of that line. Researchers at labs including Anthropic and Google DeepMind have quit noisily, saying the models pose real threats and that safety is not being taken seriously across the industry [5], and other researchers have put the chance of AI killing everyone above 10 percent [6]. Those are claims about risk, and neither settles whether two rivals may agree on a release date. The published transcript, lightly edited for length and clarity, stops mid-sentence inside Kanter's first answer [15].
What to watch
- Whether any lab publishes the actual text of an exemption request, including the conduct it would immunise and for how long.
- Part two of the Decoder series, and whether Kanter lands on a position on the exemption itself.
- Whether the live DOJ case against Apple changes how the division treats coordination between the frontier labs.