OpenAI confirms weeks of safety coordination with its rivals — and says no antitrust waiver is needed

Chris Lehane, OpenAI's chief global affairs officer, told reporters in Washington on Tuesday that his company had been working with Anthropic and Google DeepMind on AI safety, and that the engagement had been under way for several weeks. The three firms do not need an antitrust waiver to coordinate on safety matters, he said — a claim Bloomberg Law recorded as OpenAI's position rather than as a legal ruling. His own wording was plainer: "It's better to try to work together to prioritize safety."

That much is checkable. A named executive at the largest AI lab put his name to coordination with two named competitors, and the duration was weeks rather than hours. The rest of Tuesday's news was thinner than the headline count suggests: an endorsement of one provision in a bill that has not passed, and a rival's proposal nobody has accepted.

What OpenAI put its name to

Politico reported the same morning that OpenAI is backing a key provision of the FRONTIER Act, the bipartisan bill from Representatives Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) that would create the first federal framework for AI safety. The provision would require top labs to allow "independent verification organizations," or IVOs, into their companies to check that models are developed safely. Politico calls it the first time OpenAI has backed a specific federal mandate requiring third-party assessors to review the largest developers' work.

Lehane's account of how the support arrived is worth taking literally. In a Monday meeting on Capitol Hill, he said, he sat down with one of the bill's sponsors and "specifically talked about how we think about IVO and made clear that we can support that." The progression behind it is on the record: OpenAI endorsed a California bill establishing accreditation and vetting for IVOs just as Governor Gavin Newsom signed it into law, and the same day Lehane published a blog post making independent verification a pillar of the "mandatory national AI safety requirements" he wants Congress to pass.

On Saturday, Sam Altman pledged that OpenAI would voluntarily embed third-party evaluators, answering Dario Amodei's essay published hours earlier — the plan this site examined when it appeared, and one that followed an Anthropic researcher's resignation.

The hole where the IVO should be

Neither report names a body empowered to certify the certifiers. There is no threshold figure, no accreditation authority, no timeline for a first assessment, and no answer to the question a working system would settle immediately: who decides an evaluator is independent enough to be let inside a training run, and what happens to a lab that disputes the finding? Lehane can support the provision precisely because it does not yet say who administers it.

An anonymous AI staffer in Trahan's office called the timing a "really strong signal that the leaders of these labs jumped on board with Dario's embedded evaluators idea," then supplied the sentence that matters more: the next few weeks "will be critical to see sort of how those firm commitments translate to actual legislation."

One detail has drawn less attention than it deserves. The requirement would apply only to companies meeting high thresholds for annual revenue and computing investment, which means the duty lands on a small club of frontier developers — the same companies whose lobbyists are now helping shape the text. Congress chose a threshold, and it happens to describe the firms asking for the rule.

Two mechanisms, a scoreboard and a houseguest

Elon Musk offered a different instrument on Monday, at the All-In Summit in Los Angeles. xAI, OpenAI, Anthropic, Google, Meta and "three or four of the leading Chinese companies," he said, should let rivals run a "test harness" on their models before release. CNBC recorded the framing: "instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns." No rival has agreed, he conceded, and he predicted "it's probably something that China would agree to."

The two ideas are not variants of one another. An embedded evaluator is a houseguest with a contract: continuous access to training pipelines, incident logs and weights, and a mandate to verify whether a lab kept its own safety commitments. A test harness is a scoreboard run by the opposing team, at a point in time, producing a result either side can contest. The first depends on trusting the institution doing the watching; the second, on little beyond agreement over which tests count.

Each has an obvious failure mode. Rivals scoring your model have an incentive to over-report, which turns every finding into a negotiation about methodology. An auditor living inside a company can be captured so quietly that nobody notices. Neither mechanism needs a statute to begin — the strongest thing that can be said for either — and neither names who holds the last word when a lab rejects the verdict.

The antitrust question nobody has answered

TechCrunch notes that some in the industry, Altman among them, have said coordination of this kind could invite antitrust liability if it is found to suppress competition. Amodei's essay proposed a narrow government waiver; Lehane says none is needed. Both cannot be right in the same legal universe, and which one is depends on facts neither lab has disclosed: what the three discuss, whether any of it touches pricing, roadmaps or hiring, and whether an enforcement agency is willing to say so on paper.

TechCrunch also cites The Information's reporting that the three have been working to create an industry standards body — something Altman reportedly told staff would need to happen without US government support. That is precisely the institution whose existence would force the antitrust question the labs are now answering with the word "safety."

What would settle it

Four documents would move this from a claim to a record. The FRONTIER Act's text, with its revenue and compute thresholds published and its IVO-accreditation mechanism named. At least one IVO accredited and disclosed, with its access terms. The first audit produced by a lab's embedded evaluator, including what that evaluator was not permitted to see. And antitrust guidance saying whether coordination of this shape is permitted or not.

Until those exist, the week amounts to three competitors telling the same reporter that they have been talking about safety. It is not yet a structure that could catch a lab which decided to stop talking. We covered the other half of this week's politics, from the White House's response to the market's reaction, in our September 15 story.