Anthropic and OpenAI Sought Control Over What Outside AI Evaluators Can Publish, Reuters Reports

Anthropic and OpenAI pushed for significant control over what outside evaluators embedded inside their companies can publish, Reuters reported October 13, citing two people familiar with the talks and a contract the news agency reviewed. The concern, per Reuters' sourcing, centers on evaluators who work inside the labs with employee-level access: both companies reportedly worry those reviewers could carry proprietary information to a rival lab, and have proposed safeguards including having lab staff observe evaluation teams and barring an evaluator embedded at one lab from working at another at the same time.

The reported push sits awkwardly against both companies' own words from weeks earlier. In September, Anthropic chief executive Dario Amodei wrote that external evaluators should be able to publish their main findings "without editorial control" by the company being evaluated. OpenAI said much the same days later, citing editorial independence as its own standard. Reuters' sourcing says that as late as October 5, both labs were still seeking strict control over what reviewers could publish -- a position neither company's public statement a few weeks earlier described.

Both companies disputed Reuters' framing when asked. Anthropic said the description of seeking "significant control" didn't accurately capture all of its discussions, and that its current and prior evaluators control their own methods, criteria, conclusions and publication decisions -- adding that new evaluator firms face a review-and-approval period while they learn how its models work. OpenAI said it does not seek approval over a report's full text, that redacting confidential information is standard practice, and that evaluators are permitted to disclose when a significant redaction was made.

There is already a real example of how that redaction process plays out. After autonomous agents powered by OpenAI's models hacked Hugging Face's infrastructure, the nonprofits METR and Redwood Research investigated inside OpenAI -- a probe planned to take two days that instead ran six, over three visits, after the researchers flagged gaps in the transcripts they'd been given. The need for three separate visits is itself informative: an evaluator with genuinely equivalent internal access wouldn't have had to come back twice to get a complete record of what happened. OpenAI redacted nonpublic information from the resulting report, and METR and Redwood say they also edited its structure, emphasis, clarity and tone based on OpenAI's feedback. The two organizations say nothing that changed their conclusions was removed and stand by the findings -- but they also said concern about discouraging labs from allowing outside evaluators in after future incidents shaped how they handled the editing. That's a disclosed conflict of interest in their own process, stated plainly rather than hidden, which is itself a form of the transparency this whole dispute is nominally about.

The dispute is really about who holds the pen at one specific moment -- after an evaluator has found something, before the public sees it.

Marc Warner, chief executive of Faculty -- the Accenture-owned firm Anthropic named as its first embedded evaluator in September -- told Reuters evaluators are "only useful if people can trust what they report, so independence and transparency are critical." Rumman Chowdhury, the former U.S. Science Envoy for AI and now chief executive of Humane Intelligence, was more pointed about the structural bind embedded evaluators sit inside. More than 200 researchers, including Geoffrey Hinton, signed a public letter this fall setting minimum conditions for credible embedded evaluation -- independent control of findings, broad lab access, and protection from retaliation among them. The one-page White House Accord on Superintelligence, which President Trump signed in late September and called "morally binding," lists independent evaluation as one of four safety measures, without specifying how a review should actually be conducted or published.

“Evaluators have this very privileged access, but whether explicitly or implicitly, they know that it comes at a cost.” — Rumman Chowdhury, CEO of Humane Intelligence PBC, former U.S. Science Envoy for AI

Reuters' reporting also places the dispute against a specific financial backdrop: both companies are pursuing valuations in the trillions of dollars, with OpenAI's and Anthropic's IPO ambitions reportedly part of the calculus executives weighed in the publication-control talks. A finding that spooks regulators or investors ahead of a public offering is a commercially different event than one that doesn't -- which is exactly the kind of incentive the independence conditions researchers published this fall were written to guard against, and exactly what neither company's statement to Reuters addresses.

None of this is settled by a Reuters report or a company's pushback on it. The accord both companies signed treats independent evaluation as self-evidently a safety measure; this week's reporting is a reminder that the two labs being evaluated are also the ones currently deciding what independence is allowed to mean -- and that the first real test of which definition wins will be the first finding one of them doesn't want made public.