OpenAI and Anthropic are facing questions over how independent their outside safety reviews can really be after discussions revealed tensions over what evaluators are allowed to publish.
Both companies have publicly supported stronger third-party scrutiny of advanced AI systems. But negotiations with external evaluators have included proposed restrictions designed to protect trade secrets and other confidential information.
The dispute goes to the heart of a growing AI governance debate: can a review be truly independent if the company being examined can influence what the public eventually sees?
FCRF Launches CP-FRM to Build India’s Next Generation of Fraud Risk Professionals
AI companies want outside scrutiny but fear trade-secret leaks
Anthropic CEO Dario Amodei said last month that external evaluators should be able to publish key findings without company editorial control.
OpenAI later backed the principle of editorial independence for outside assessors.
Behind those statements, however, the practical rules are still being negotiated.
One concern is “embedded evaluation”, where independent reviewers work inside an AI company and receive access similar to employees. That deeper access can help them understand how advanced models behave, but it may also expose proprietary technology.
AI companies fear that evaluators could accidentally reveal commercially sensitive information or carry knowledge from one laboratory to a rival.
Measures discussed have included allowing company employees to observe embedded reviewers and preventing evaluators from working inside competing AI laboratories at the same time.
The difficulty is that information companies regard as a trade secret may also be exactly what an evaluator needs to explain why a model could be dangerous.
That creates a direct conflict between commercial secrecy and public-interest safety reporting.
What an embedded AI evaluation actually involves
Traditional outside testing usually gives researchers limited access to a finished model.
Embedded evaluation goes much further.
Reviewers may work inside the company and examine models during development. They can study internal safeguards, testing methods and how the system behaves under conditions ordinary users would never see.
The goal is similar to allowing an independent safety inspector into a factory instead of asking them to judge the factory only from the finished product.
Anthropic recently announced that Faculty, Accenture’s specialist AI business, would become its first embedded evaluator.
The company said such reviewers could help test frontier models, assess safeguards and challenge assumptions made by internal teams.
OpenAI has separately said third-party assessors should receive deep access to training, evaluation and deployment processes and should be able to reach their own conclusions.
The unresolved question is how much of those conclusions can be made public without company approval or redaction.
Contracts and redactions raise independence concerns
Some agreements used by AI companies already restrict evaluators from publishing confidential information without consent.
OpenAI says it is not seeking the right to approve the full text of independent reports.
It says its normal approach is to remove sensitive information while allowing evaluators to disclose whether important redactions have taken place.
Anthropic has said outside reviewers should control their own methods, criteria and conclusions.
However, it has also said that when working with new evaluator firms, it may initially reserve the right to comment on or approve reports while those organisations learn how its systems operate.
That distinction worries some experts.
An evaluator may technically be independent, but the AI laboratory still controls the access needed to conduct the review.
If publication of uncomfortable findings risks losing future access, critics argue that independence can become weaker in practice than it appears on paper.
White House AI accord leaves key questions unanswered
The debate has become more important since six major technology companies signed the White House Accord on Super Intelligence in September.
The agreement was signed by OpenAI, Anthropic, Google, Meta, xAI and Nvidia.
It calls for internal safety controls, independent external evaluations and board-level oversight of frontier AI systems.
But the accord is voluntary.
It does not establish detailed rules for how external reviewers should operate, what information they must receive or what findings must be disclosed publicly.
More than 200 researchers have separately called for minimum standards that would give independent evaluators broad access, control over their findings and protection from retaliation.
That pressure is growing as governments consider whether voluntary arrangements should eventually become law.
The argument is no longer simply about whether AI companies should permit outside reviews.
The harder issue is who controls the reviewer, who decides what is confidential and whether the public can see findings that reveal serious safety problems.
Without common standards, each AI company can still negotiate those conditions separately.
That could leave some of the most important decisions about transparency in the hands of the same companies whose systems are being examined.
What this means for you:
Independent AI reviews matter only if outside experts can report serious risks without being silenced by commercial agreements. The rules being negotiated now could determine how much the public learns when powerful AI systems fail safety tests.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
