AI Labs' Secret Safety Pact Faces Antitrust Test
OpenAI, Anthropic and Google's Private Safety Talks Put Antitrust Law to the Test
Update, Sept. 17: The Justice Department says frontier AI labs have not requested a meeting on antitrust concerns around safety coordination, Associate Attorney General Stanley Woodward told a Fordham University panel Thursday, per Reuters — meaning the coordination debate is still happening entirely outside any regulatory framework.
The most consequential AI news of the week didn't arrive in a research paper or a product launch. It arrived in a briefing room, almost as an aside: the three companies racing hardest to build superintelligence have already been sitting down together, for weeks, deciding how to handle the risks of what they're building.
Chris Lehane, OpenAI's global head of policy, told reporters on Tuesday that OpenAI has been working with rivals Anthropic and Google DeepMind on AI safety for several weeks, Bloomberg first reported. Lehane, who was in Washington to meet lawmakers about catastrophic AI risks, said OpenAI would support any bipartisan legislation aimed at mitigating those risks. Reuters, which covered the Bloomberg report, noted it could not independently verify the account — and that OpenAI, Alphabet and Anthropic did not immediately respond to its requests for comment.
It was a remarkable admission hiding in plain sight. The public story of the last ten days has been a weekend show of unity — Anthropic CEO Dario Amodei's Saturday essay, "We Must Pace the Frontier," calling the industry to slow the frontier's advance; Sam Altman, Demis Hassabis and Elon Musk publicly backing it. Lehane's revelation suggests the unity was not improvised over the weekend at all. The labs were already coordinating before the public ever saw the essay. The handshake came first; the press release came later.
"It's better to try to work together to prioritize safety," Lehane said at the Washington briefing, according to Bloomberg.
Why lawyers are nervous
Here's the problem: in American law, when the three dominant firms in an industry sit down together to coordinate their behavior, there is a word for it. Cartel.
Several legal experts have warned that an industry-wide coordinated slowdown could amount to "output restriction" — a classic antitrust violation under the Sherman Act. The concern is not abstract. If the three leading frontier labs agree, even informally, to throttle the pace at which they release their most capable systems, that is a coordination among direct competitors over the timing and quantity of their products. Antitrust law was written for exactly this shape of conduct, even when the motive is public-spirited.
Amodei seemed to sense the trap. His essay proposed that the US government "mediate or at least enable" the safety coordination — through a standards body or a narrow waiver shielding the labs from antitrust liability. Lehane, notably, disagreed. He reportedly told reporters the firms don't need a waiver at all.
That disagreement matters more than it looks. If the labs proceed with coordination without legal cover and a future administration — or a state attorney general, or a private plaintiff — decides the coordination crossed the line, the liability could be severe. Treble damages under the Sherman Act are not a theoretical risk; they are the standard remedy. The labs are betting, in effect, that no enforcer will want to be the person who sued the companies for trying to make AI safer. It is a bet about politics as much as law. The labs' fear is not abstract: the rogue-agent swarm that hit Hugging Face this summer is exactly the kind of incident they say coordination could prevent.
The DOJ's answer: nobody has asked for a meeting
The Justice Department, for its part, says the labs haven't even knocked on the door. Associate Attorney General Stanley Woodward, who leads the DOJ's antitrust division, told a Fordham University panel in New York on Thursday that frontier AI labs have not requested a meeting on antitrust concerns related to safety coordination, according to Reuters. The remark lands as a quiet rebuke to a week of public hand-wringing: the industry has been debating whether regulators would bless coordination, while regulators say no one has asked.
That leaves the labs in an awkward spot. Anthropic's Dario Amodei called last week for explicit antitrust permission to coordinate a slower pace of AI development on safety grounds; OpenAI's Chris Lehane countered that no waiver is needed. Woodward's answer cuts through both positions — there is, as yet, no request to grant or deny. For readers, the takeaway is that the "private safety talks" remain exactly that: private, informal, and outside any regulatory framework, which is precisely the arrangement antitrust lawyers find most uncomfortable.
The FTC just weighed in
Since this piece was published, the pushback has gone on the record. At a Washington event on September 15, Federal Trade Commission Chair Andrew Ferguson said he was "highly suspicious" of the labs' request for antitrust exemptions, characterizing it as an attempt to build a moat blocking new competitors, per reports by the Korea Herald's business desk and PYMNTS. Ferguson framed the remarks as his personal view, not FTC policy — but coming from the sitting antitrust chief, they sharply narrow the labs' room to maneuver. The same week, David Sacks, co-chair of the President's Council of Advisors on Science and Technology, rejected Anthropic CEO Dario Amodei's waiver request outright, telling the labs they need nobody's permission to slow down unilaterally — and President Trump, asked whether AI development should slow or face more regulation, said the United States leads China and intends to keep it that way. Meanwhile AFP reported Tuesday that the three labs' planned standards body is modeled on FINRA, the US broker-dealer self-regulator — a concrete detail the labs hadn't disclosed.
The cartel objection
Not everyone is applauding. Cohere CEO Aidan Gomez, whose company competes with the three labs in enterprise AI, published an essay calling the coordination effort "a wolf in sheep's clothing, a cartel by another name." His argument: the market-dominant labs are using the language of safety to win permission to "define the rules and safety standards" of the industry — a permission that would let incumbents set the terms of competition for everyone else.
There is a real version of this concern. Safety standards are never neutral. If the three frontier labs agree among themselves on what counts as a safe evaluation protocol, a safe disclosure norm, or a safe deployment practice, those standards will be written around the capabilities, architectures and business models of the companies in the room. Smaller labs, open-weight developers and foreign competitors would then face a choice between adopting standards they didn't write or being labeled unsafe. Gomez's essay is self-interested — Cohere is a direct competitor — but self-interest doesn't make the point wrong.
Congress has noticed the tension. In July, Senator Jim Banks of Indiana and Senator Adam Schiff of California introduced the "Collaboration on Adversarial Threats and Security Risks Act," which would create antitrust carveouts specifically to ease safety coordination among AI labs. The bill's existence is an admission that the current legal framework does not cleanly accommodate what the labs say they need to do. Whether it moves is another question — and early signals from the Trump cabinet, as Gizmodo reported, suggest the administration is skeptical of waiver-assisted collaboration, preferring to keep the industry's competitive pressure intact.
The waiver that won't come
The legislative fix the labs might want is going nowhere. Senators Jim Banks and Adam Schiff introduced the Collaboration on Adversarial Threats and Security Risks Act in July to create antitrust carveouts for AI safety coordination — but the bill has stalled, and early signals from the Trump administration suggest skepticism of any waiver-assisted collaboration. Meanwhile Cohere CEO Aidan Gomez took his "cartel" argument to Bloomberg Tech on September 14, arguing safeguards should be shaped by a broader group than Anthropic and OpenAI alone. The labs are discovering that the permission structure they want — coordinate first, get legal cover later — has no buyers in Washington right now.
The FINRA-style standards body
According to The Information, the three companies have been working together to create an industry standards body — and Altman reportedly told staff it would need to happen without the support of the US government. That detail, reported by TechCrunch, is worth pausing on: the labs want a coordination mechanism, but they appear to have concluded that waiting for Washington to bless it would take too long, or come with strings they don't want.
The idea is not new. In July, Hassabis publicly called for the United States to establish a standards body that could act as an AI watchdog — with the power to screen the world's most advanced models and coordinate industrywide slowdowns if dangers progressed. What was then a proposal from one lab chief has now, apparently, become a joint project of three.
What would such a body actually decide? The honest answer is that nobody knows, because the hard questions are all unresolved: what counts as a dangerous capability, who gets to evaluate it, what evidence triggers a pause, and — the question that will decide whether this is governance or theater — whether a pause decision is binding on anyone. A standards body that can only recommend is a trade association. A standards body that can halt a $100-billion training run is a regulator. The labs have not said which one they are building. The economics are staggering — the labs are discussing pausing $100-billion training runs even as OpenAI's $1.2 trillion pre-IPO round values the prize of continuing.
What Washington actually wants
Lehane's trip to Washington was not only about the lab coordination. He told reporters OpenAI would back bipartisan legislation targeting catastrophic AI risks — a position that puts the company, somewhat awkwardly, on both sides of the regulatory debate at once: asking Congress to legislate, while building a private coordination mechanism that might not need Congress at all.
The awkwardness is the point. The labs are operating in a vacuum that Washington created. There is no federal AI safety law, no licensing regime, no empowered agency with clear jurisdiction over frontier models. The executive order regime — voluntary pre-release testing, reporting frameworks — asks for cooperation without compelling it. Into that vacuum, the three labs that can most credibly claim to understand the technology are writing the first draft of the rules themselves, in private, and daring the legal system to object.
There is a generous reading of all this: the people closest to the technology are moving fastest to contain its risks, and legal formalities can catch up later. There is also a cynical reading: the labs that spent a decade racing each other have discovered that coordination is useful, that safety is the one justification antitrust law might respect, and that whoever writes the standards first wins. Both readings can be true at the same time. In Washington this week, the question is no longer whether the labs are coordinating. It's whether the law will let them keep doing it — and who gets to set the terms. The question may get its first real test at the October Trump–Xi AI summit, where US–China AI competition meets safety policy.
Update, Sept. 17: The antitrust question behind this story got fresh air this week. Abigail Woodward, the DOJ's principal deputy assistant attorney general for antitrust, said Tuesday at an FTC conference that she was not aware of any frontier AI lab asking the department to meet about how to coordinate on AI safety without violating competition law — leaving the door open for such a conversation. For the labs coordinating privately, the takeaway: Washington will talk, but no one has asked yet. (Reuters, Sept. 17.)
Sources: Bloomberg (via Reuters, Sept. 15, 2026); Reuters (Sept. 15, 2026); TechCrunch (Sept. 15, 2026); Gizmodo (Sept. 15, 2026); Neowin (Sept. 15, 2026). For the weekend slowdown agreement that preceded this revelation, see our earlier coverage of the Amodei essay and the industry pact.
Comments
Post a Comment