Anthropic and OpenAI Require Approval Before Evaluation Reports Are Published, Exposing Contradiction in Independence Commitments

Reuters reported on October 13, 2026, that Anthropic, at least until October 5, still required newly partnered evaluation organizations to obtain its approval before publishing reports, and OpenAI took a similar position.

Factual Reconstruction

According to a Reuters exclusive, Anthropic and OpenAI were concerned that external evaluations could leak trade secrets, so they sought significant control over what embedded evaluation teams published. Both companies had previously publicly stated that they supported editorial independence for evaluation organizations, but negotiation details showed differences in actual practice. After October 8, discussions between Anthropic and some evaluation organizations showed signs of softening.

The report noted that Anthropic responded that Reuters had not accurately described all of Anthropic's discussions with independent evaluation organizations. In current talks with evaluation organizations it has worked with, it has made clear that external reviewers control the methodology, standards, conclusions, and publication content. For new evaluation companies, Anthropic reserves the right to comment on and approve reports. OpenAI said it does not seek the right to approve the full text of reports; its standard practice is to remove confidential information and allow evaluators to disclose material redactions.

Mechanism Breakdown

Negotiations focused on how to prevent intellectual property leaks during embedded evaluations. Measures under discussion included having lab employees oversee the work of embedded teams and prohibiting the same evaluator from working at multiple labs at the same time. Content that labs regard as key intellectual property often overlaps with information relevant to judging model risk, leading to disputes and redactions.

Contract language already explicitly prohibits evaluation companies from publishing confidential information in reports without company consent. In the absence of unified standards and legislation, companies building the most powerful models may determine the review method, while evaluators rely on labs to provide access.

Industry Impact

This incident reveals the trade-off AI labs face between protecting trade secrets and disclosing the potential risks of powerful AI as they pursue trillion-dollar valuations and potential IPOs. The credibility of independent evaluation is directly affected; if labs have approval rights over reports, the transparency of third-party assessments will face systemic challenges.

Evaluation organizations such as Faculty stress that evaluations are only useful when people trust the reports, making independence and transparency crucial. Experts note that without standards, labs' control over access gives them leverage over report content.

Strategic Judgment (Analysis, Not Fact)

Judging from the current negotiation process, AI labs may continue to balance disclosure needs with secrecy protection through contract terms and access controls. This will push the industry toward clearer prior agreements, but in the short term, evaluation independence will still depend on the specific compromises each party makes.