Anthropic, OpenAI Sought Control Over Reviewer Publications

Anthropic and OpenAI pushed outside evaluators of their AI models to give the labs a big say over what those reviewers can publish, Reuters reported via The Next Web, citing two people familiar with the talks. The companies argued that findings could expose trade secrets, with particular concern about reviewers who sit inside the labs and could take proprietary information to rivals.

Until at least October 5, both labs wanted strict control over what reviewers could publish; by October 8, Anthropic's talks with some reviewers "seemed to have moved on," one source told Reuters. The shift comes weeks after Anthropic boss Dario Amodei said evaluators should be able to publish their main findings without the company editing them, and a month after the company named Faculty, Accenture's AI arm, as its first embedded evaluator in September.

"Evaluators are only useful if people can trust what they report, so independence and transparency are critical," Faculty chief executive Marc Warner told Reuters. Rumman Chowdhury, a former US science envoy for AI under Joe Biden, put it more plainly: "Evaluators have this very privileged access, but whether explicitly or implicitly, they know that it comes at a cost."

The labs' public positions diverged. OpenAI said it did not want to approve whole reports, that it usually removes secret details, and that reviewers can note when large cuts were made. It has already cut details from a report by safety groups METR and Redwood Research about the time its agents hacked Hugging Face. We covered METR a day earlier, when CNBC profiled the firm as one of the industry's safety gatekeepers. Anthropic told Reuters that for reviewers it has worked with before, those partners "decide their own methods, findings, and what to publish."