OpenAI says it will let third-party evaluators inside its development teams

The vendor review form has a row for independent pre-deployment testing: yes, no, evidence attached. For OpenAI, the evidence on offer this week is a press briefing in Washington and a chief executive agreeing with a rival's blog post ^[1][5]. Chris Lehane, OpenAI's global policy chief, told reporters on Tuesday that the company had been working with Anthropic and Google DeepMind on safety for weeks ^[1]. He did not elaborate on what that work involves, according to SiliconANGLE ^[3].

The two accounts of Altman's commitment do not describe the same arrangement. SiliconANGLE reported that Altman said OpenAI is willing to join Anthropic's initiative and embed third-party evaluators within its development teams ^[5]. TechCrunch reported him saying OpenAI would embed third-party evaluators into the company ^[6]. An evaluator sitting with a training team sees checkpoints. An evaluator sitting in the building reads documents. Neither report names an evaluator organisation or a start date ^[23].

Amodei published his essay on Saturday, calling on the industry to slow the pace of frontier AI and avoid what he called "catastrophic risks" ^[4]. Lehane briefed reporters three days later and dated the talks to several weeks, so the cooperation was under way before the public call for it ^[22]. Altman, Demis Hassabis and Elon Musk all backed the essay once it was out ^[20]. The Information reported that the three companies have been working on a standards body for the industry, and Altman reportedly told staff it would have to happen without the support of the US government ^[18].

The part a customer will eventually be able to cite is legislative. Lehane said OpenAI supports a FRONTIER Act provision that would force top frontier labs to allow "independent verification organizations" into their companies ^[16], and the bill would also require labs to submit new models for audits by independent evaluators ^[15]. Reuters reported that Majority Leader John Thune is working with Ted Cruz and Amy Klobuchar on a separate bill giving the Commerce Secretary power to order safety audits before release ^[17].

Amodei asked the government to consider a waiver so safety coordination would not fall foul of antitrust law; Lehane said the firms do not need one ^[7]. Coordination of this kind can be read as "output restriction" under the Sherman Act ^[8], and Senators Jim Banks and Adam Schiff have already proposed a Collaboration on Adversarial Threats and Security Risks Act to let labs act on safety where doing so contravenes antitrust law ^[9]. Cohere chief executive Aidan Gomez was blunter. "Once again using fear under the pretext of protecting the public, these oligopolies are now requesting to bend competition rules and be permitted to dictate the terms for everyone else," Gomez said in a post responding to Amodei ^[11]. "A wolf in sheep's clothing, a cartel by another name" ^[12]. Trump rejected Amodei's call in a Truth Social post, saying a slowdown would let China overtake the US ^[13], and his AI adviser David Sacks, an investor with stakes in many AI firms, has said existential-risk fears are overblown ^[14].

Who is the evaluator, and who pays them? What stage do they see: a training checkpoint, a release candidate, or a shipped model? What can they stop? Which part of their findings does a customer get to read? On the current record the first has no answer, and the second is the one the two reports disagree about, so the pledge belongs in the risk register. The tradeoff in waiting is real: anyone who wants evaluator access in a 2027 contract has to draft the clause with no published standard to point at, and will be writing the scope themselves.