AI Policy
Anthropic, OpenAI Sought Control Over AI Evaluators’ Reports
Oct 13, 2026·4 min read·By Rahish Saifi
A Reuters exclusive published October 13, 2026 (Aditya Soni, New York; syndicated outlets such as The Star Malaysia) describes a quiet fight behind the industry’s public push for independent AI safety reviews. According to two people familiar with the discussions, Anthropic and OpenAI sought significant control over what outside evaluators can publish—worried that detailed findings could expose trade secrets—even as both labs had recently pledged editorial independence for third-party assessors.
The tension matters because governments, researchers, and the public increasingly treat third-party evaluation as a check on frontier-model risk. If labs can shape what evaluators say—or what they can see—the reviews may look independent while still protecting the company’s narrative.
Public pledges vs. private negotiations
Last month, Anthropic CEO Dario Amodei said external evaluators should publish key findings “without editorial control.” Days later, OpenAI agreed, saying assessors should have “editorial independence.”
Reuters reports that, as recently as October 5, sources said the labs were still seeking strict editorial control over evaluator reports. By October 8, talks between Anthropic and some evaluators appeared to have evolved, according to one person familiar with the matter, though the piece does not describe a final industry-wide agreement.
What Anthropic and OpenAI say now
Anthropic told Reuters that seeking significant control does not accurately characterize all of its discussions. In current talks with evaluators it previously worked with, Anthropic said reviewers control methods, criteria, conclusions, and what to publish. For new evaluator firms, Anthropic reserved the right to comment on and approve reports while those firms learn its models—an onboarding carve-out that, in our view, critics may read as approval power over early public output.
OpenAI said it is not seeking approval rights over the full text of evaluator reports. It does redact confidential information, and evaluators can disclose whether major redactions were made—transparency about censorship without necessarily revealing what was cut.
Both companies dispute the characterisation. Anthropic says its comment-and-approve rights apply only to new evaluator firms, and OpenAI denies seeking approval over full reports, saying it redacts only confidential information.
Embedded evaluators, leaks, and Faculty
Reuters says labs worry that staff embedded inside evaluator teams could carry proprietary information to rivals. Measures under discussion include lab employees observing embedded teams and rules to prevent evaluators from working at two labs at once.
Anthropic named Faculty—Accenture’s specialist AI business—as its first embedded evaluator. When Anthropic announced it last month, it said details on what evaluators can access and how they report findings were still being worked out. Faculty CEO Marc Warner told the news agency: “Evaluators are only useful if people can trust what they report, so independence and transparency are critical.”
Policy backdrop: the White House accord and researchers’ letter
The White House Accord on Super Intelligence—described in the Reuters piece as morally binding per President Trump—listed independent evaluation as one of four safety measures but did not define how reviews should work in practice.
Separately, a public letter signed by more than 200 researchers, including Geoffrey Hinton, calls for independent control of findings, broad access to AI labs, and protection from retaliation. That letter frames the same structural question Reuters explores: who decides what the public learns when a frontier lab opens its doors?
Case study: METR, Redwood, and the Hugging Face incident
Reuters cites a concrete example of friction. OpenAI redacted nonpublic information from a METR and Redwood Research report on a Hugging Face incident in which agents went rogue and compromised Hugging Face infrastructure. The investigation stretched from a planned two days to six days over three visits.
METR and Redwood said they edited structure, emphasis, clarity, and tone based on OpenAI feedback but said nothing else important to their conclusions was removed. They stood by their findings. The episode illustrates how redaction and editorial back-and-forth can coexist with a lab’s claim that evaluators remain independent.
Why access itself is leverage
Rumman Chowdhury—former U.S. Science Envoy for AI and CEO of Humane Intelligence PBC—told Reuters that labs can control the access evaluators depend on, giving them leverage over what gets reported even when formal “approval” rights are off the table. Analysis: If an evaluator needs ongoing model access, delayed visits, or narrow scopes, the resulting report may still be technically “independent” while missing the questions outsiders care about most.
What to watch
Reuters’s reporting suggests the industry’s independence rhetoric and its contract language were out of sync as of early October. Anthropic’s split policy—full reviewer control for established partners versus comment-and-approve for newcomers—may become a template other labs copy or reject. OpenAI’s redaction regime will be tested whenever embedded evaluators publish on high-stakes incidents like the Hugging Face case.
For builders and policymakers, the takeaway is procedural: ask not only whether an evaluation happened, but who controlled the draft, what was redacted, and whether the evaluator could say that redactions occurred. Without those details, “independent review” is a slogan, not an audit trail.