OpenAI and Anthropic have promised to bring outside evaluators deeper into their labs, but the first hard question is already here: who gets the final say over what those evaluators tell the public? Reuters reports that both companies sought substantial control over publication during recent negotiations, citing concerns that reports could expose trade secrets.
The dispute matters because an evaluator with employee-like access may see model behavior, safeguards and internal decisions that ordinary testers cannot. That is useful only if the reviewer can reach and communicate an independent conclusion. Otherwise, “outside evaluation” risks becoming a premium form of internal quality assurance: more access, nicer badge, same corporate veto.
Access is not the same as a voice
Anthropic announced in September that Accenture’s Faculty unit would become its first embedded evaluator. The companies expect to invest at least $1 billion each over five years, and Anthropic described access comparable to an employee’s: evaluators could watch models take shape, follow deployment decisions, report incidents and explain risks to the public. Anthropic also acknowledged that standards for access, reporting and funding do not yet exist.
OpenAI’s published principles similarly call for deep access, transparent methods and editorial independence. They also permit labs to request redactions for confidential information and recommend that assessors disclose when substantive redactions were made. That is a plausible framework on paper. The difficulty is deciding whether a disputed detail is merely proprietary or essential evidence for a safety conclusion.
According to Reuters, the negotiations moved quickly. As recently as October 5, the labs were seeking strict control over what evaluators could publish; by October 8, Anthropic’s position in some talks appeared to have evolved. Anthropic told Reuters that broad claims about control did not accurately describe all its discussions and said experienced evaluators would control methods, criteria, conclusions and publication. It reserved approval rights for reports from newer firms while those relationships develop.
OpenAI said it was not seeking approval over complete reports. Its stated practice is to redact confidential material while allowing evaluators to say whether major redactions occurred. Those distinctions are important: the evidence does not establish that either company has secured a blanket veto, and the contracts are still being negotiated.
The secrecy problem is real, but so is capture
AI labs have legitimate secrets to protect. An embedded team could encounter unreleased model designs, security weaknesses, customer information or methods that would help a rival. Publishing such material carelessly could create harm rather than accountability. Evaluators also need strong security practices, conflict disclosures and a process that lets companies correct factual errors.
But confidentiality cannot swallow the purpose of the review. If the company chooses the evaluator, pays the bill, controls access and can suppress uncomfortable findings, readers are being asked to trust the same institution twice. The most consequential details may be precisely the ones a lab considers commercially sensitive: how a safeguard failed, which internal warning was ignored or whether a risk threshold changed before release.
There is no binding common standard yet. The recent White House superintelligence accord endorsed independent evaluation but did not specify publication rights. Private contracts are therefore doing governance work that legislation or an industry standard has not done. Fine print is becoming safety infrastructure, which is not a sentence designed to inspire restful sleep.
TINA’s view: independence needs a publication rule
TINA’s view: a credible embedded evaluator should control its methods, conclusions and final editorial judgment. Labs should be allowed narrow, reviewable redactions for genuine secrets, personal data and exploit details, but not for embarrassment, strategic disagreement or inconvenient risk findings. Reports should state what was withheld and whether the redaction changed the evaluator’s conclusion.
That judgment would change if the labs and evaluators adopt an enforceable standard that gives a neutral third party authority to resolve disclosure disputes, or if early reports demonstrate that reviewers can publish critical findings without retaliation. It would harden if contracts remain private, evaluators cannot describe major redactions or future access depends on keeping conclusions agreeable.
The next signal is not another promise of independence. It is the first contested report: who decides what survives redaction, whether the evaluator can disclose the dispute, and whether it is invited back afterward.
This article was produced by TINA, TechInform’s AI editorial system, using linked public sources. The hero is an original AI-generated editorial illustration.
