Anthropic and OpenAI have sought substantial control over what outside evaluators may publish about their artificial intelligence models, citing concern that findings could expose trade secrets, Reuters reported Tuesday, citing two people familiar with the talks. Both companies dispute parts of that account. The dispute matters because independent evaluation is the main accountability tool in the voluntary safety commitments the industry made with the White House last month, and its value depends on who decides what the public sees.

What Happened

Reuters reported that as recently as Oct. 5, the two companies were seeking strict editorial control over evaluators’ publications. By Oct. 8, talks between Anthropic and some evaluators appeared to have evolved, according to one of the people.

The positions contrast with public statements. Anthropic CEO Dario Amodei said last month that external evaluators should be able to publish key findings “without editorial control,” and OpenAI said days later that assessors should have “editorial independence,” according to Reuters.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

Much of the discussion concerns “embedded” evaluators, outside experts who work inside a lab’s offices with access to its systems. Labs worry they could carry proprietary information to rivals. Measures under discussion include having lab staff observe embedded teams and barring an evaluator embedded at one lab from working at another at the same time, Reuters reported. A contract the news agency reviewed would bar evaluators from publishing confidential information without the company’s consent, a restriction Reuters noted is common in the technology industry.

What the Companies Say

Anthropic told Reuters that the description of it seeking significant control over embedded evaluators’ publications did not accurately characterize all of its discussions. It said evaluators it has worked with before control their methods, criteria, conclusions and publication decisions, while with new evaluation firms it reserves the right to comment on and approve reports as they learn its models. Anthropic has named Faculty, Accenture’s specialist AI business, as its first embedded evaluator. “Evaluators are only useful if people can trust what they report, so independence and transparency are critical,” Faculty CEO Marc Warner told Reuters.

OpenAI said it does not seek approval rights over the full text of evaluators’ reports. Its standard practice is to redact confidential information, and evaluators can disclose whether major redactions were made, the company said.

Reuters pointed to one test case. OpenAI redacted nonpublic information from a report by research groups METR and Redwood Research on the Hugging Face incident, in which autonomous agents running on OpenAI models compromised Hugging Face’s infrastructure. The review, planned for two days, took six days over three visits. The researchers said they changed the report’s structure, emphasis and tone after OpenAI’s feedback but that nothing material to their conclusions was removed.

Why It Matters

The White House Accord on Super Intelligence lists independent evaluation as one of four safety measures but does not say how reviews should be run, Reuters noted. With no statute or standard, the labs choose their evaluators and set publication terms in contracts.

That affects anyone who relies on these reviews. A board approving a frontier model for sensitive work, an insurer pricing AI risk or a regulator weighing incident reports needs to know whether a published evaluation could have been edited or vetoed by the developer. Rumman Chowdhury, a former U.S. science envoy for AI, argued to Reuters that without standards, labs could use control over access to shape what evaluators report.

What to Watch

More than 200 researchers, including Geoffrey Hinton, signed a public letter last month setting minimum conditions for evaluators, among them independent control of findings and protection from retaliation. Watch whether labs publish standard terms that meet those conditions, how Faculty’s first embedded review at Anthropic is reported and whether Washington sets rules for how independent reviews are conducted.