Product1 publisher3 min readPublished
OpenAI's policy chief says the top three labs have coordinated on safety for weeks
OpenAI says the joint work with Anthropic and Google DeepMind has been running for weeks. Its policy chief does not expect to need an antitrust waiver, Anthropic's CEO has asked Washington for one, and the FTC chairman says his alarm bells are going off.
The Product Desk · Product desk
What happened
- Lehane said he does not anticipate needing an antitrust waiver, unlike Amodei, whose essay asked Washington for legal cover for the same joint work.
- FTC chairman Andrew Ferguson told a Georgetown University audience that companies coming to Washington to ask for both regulation and an antitrust exemption set off all of his alarm bells.
- OpenAI is backing the FRONTIER Act, which would create a federal AI safety framework and require only the top labs to allow audits by independent verification organizations.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Multi-provider redundancy is priced against the risk that one lab stumbles alone. Three labs comparing safety notes makes those suppliers correlated, so a fallback from one frontier lab to another covers less than the design assumed.
- contradiction OpenAI's policy chief says the work needs no waiver while Anthropic's CEO asks Washington to enable it, so a buyer cannot plan around the two labs' legal exposure as though it were one position.
- exposure Audit duties written to bind only the top labs never enter the customer's compliance queue. They arrive as supplier release dates the customer cannot appeal or explain to its own users.
- precedent A carveout for safety coordination would let three labs set terms that smaller model vendors have to match, and Cohere's Gomez has already put the cartel objection on the record for anyone buying from those vendors.
This shows up in an architecture document, in the row where a primary model sits beside its fallback. If that row names any two of OpenAI, Anthropic and Google DeepMind, it assumes the two make their release and safety calls apart from each other. OpenAI's global head of policy, Chris Lehane, told Bloomberg that the top AI labs have been working together to "prioritize safety" [2]. OpenAI says that work has been under way for several weeks [1].
The reporting does not say what the three labs agreed to. Dario Amodei's weekend essay, which Sam Altman supported, calls on the industry to slow down [3] and asks the U.S. government to "mediate or at least enable" the cooperation through a standards body or the issuance of waivers, warning of antitrust trouble otherwise [4]. Lehane told Bloomberg he does not anticipate needing a waiver [5].
An industry-wide coordinated pause could amount to "output restriction," which violates the Sherman Act [6]. Lehane's counter is that safety coordination already fits existing law, citing the airline industry's permission to share sensitive information tied to safety and the Obama administration's clarification that companies can share information to defend against cybersecurity risks [7]. Congress has drafted the other route: in July, Sen. Jim Banks and Sen. Adam Schiff introduced the Collaboration on Adversarial Threats and Security Risks Act, which would give antitrust carveouts to AI labs to ease safety coordination [8].
Tuesday went badly for that route. FTC chairman Andrew Ferguson said at a Georgetown University event that he was "deeply suspicious" of the antitrust exemptions [9]. "I will say generally, though, that if companies are simultaneously coming to Washington and asking for a host of regulations and an antitrust exemption, all of my alarm bells go off," Ferguson said [10]. Attorney General Todd Blanche said he would not want to prejudge an exemption before consulting the rest of the Administration [12]. Cohere CEO Aidan Gomez, writing against the whole exercise, said of the incumbent labs: "A wolf in sheep's clothing, a cartel by another name" [15].
The audit language is the piece of this with a date attached to it. The FRONTIER Act, which OpenAI is backing, would create a federal framework for AI safety and force only the top AI labs to allow audits by independent verification organizations [16]. That duty lands on the supplier's release schedule. The buyer sees a launch date that moves for reasons nobody will put in writing.
Two things sort the exposed features from the rest at each place a product calls a model: whether the feature needs a current frontier model or would hold up on a checkpoint that shipped a year ago, and whether its fallback is a different lab or a different model at the same lab. The features that need frontier capability and fall back to a second lab are the ones exposed to a decision taken by a group all three named labs belong to [19]. The rest can sit out a pause without anyone filing a ticket. OpenAI is also backing a proposal under development by John Thune, Ted Cruz and Amy Klobuchar that would reportedly give the Commerce Secretary power to deploy government auditors to test AI models [17].
What to watch
- Whether the Administration takes a position on an antitrust exemption after Blanche consults it, and whether Ferguson's FTC does anything beyond commenting.
- Whether the FRONTIER Act's audit-by-independent-verification language survives the House AI Commission's pushback, and what testing timelines the Thune-Cruz-Klobuchar draft sets for Commerce.
- Whether Google DeepMind describes the joint safety work in its own words, or any of the three publishes a scope for it.