Anthropic and OpenAI have reportedly sought greater control over what independent evaluators can publish about their artificial intelligence models, raising concerns about transparency and the credibility of third-party safety assessments.
According to Reuters reporting carried by The Next Web, both companies pushed external reviewers to give them significant influence over published findings, citing concerns that reports could expose trade secrets or sensitive technical information. The discussions have renewed questions about whether independent evaluators can freely disclose potential risks when the companies being assessed have a say in what becomes public.
The negotiations reportedly remained contentious through October 5, 2026. By October 8, Anthropic’s discussions with some evaluators appeared to have shifted, according to a person familiar with the matter.
Concerns Over Evaluator Independence
Claude by Anthropic and OpenAI Logos on Smartphones, AI Rivalry Collaboration, Frontier Model Competition, ChatGPT vs Claude, Tech Giants Showdown
Independent AI evaluators play an important role in assessing model capabilities, identifying potential safety risks and providing information that governments, businesses and the public can use to judge emerging technologies.
However, their work can involve privileged access to proprietary systems, creating tension between the need for transparency and companies’ obligations to protect confidential information.
Marc Warner, chief executive of Faculty, an AI company that has worked with Anthropic as an embedded evaluator, told Reuters that independence and transparency are essential for public confidence in evaluation results.
Rumman Chowdhury, a former US science envoy for AI, also highlighted the pressure evaluators can face when their access to advanced systems depends on cooperation from the companies they assess.
The debate follows Anthropic chief executive Dario Amodei’s public position that evaluators should be able to publish their principal findings without companies editing them. In September, Anthropic named Faculty as its first embedded evaluator, giving the partnership added significance in discussions about independent oversight.
Anthropic told Reuters that partners it has previously worked with determine their own evaluation methods, findings and publication decisions.
OpenAI Defends Its Approach to Confidential Information
OpenAI has also addressed concerns about publication restrictions, saying it does not seek approval rights over entire reports. The company said it generally removes confidential details and allows reviewers to disclose when substantial portions of their work have been cut.
One example involves a report by safety research organizations METR and Redwood Research concerning an incident in which OpenAI agents reportedly hacked Hugging Face. OpenAI removed some details from the report, according to the coverage, although the available account does not specify exactly what was removed or whether the reviewers formally objected to the changes.
The episode illustrates the challenge of balancing legitimate confidentiality concerns against the public interest in understanding how advanced AI systems behave, particularly when their actions involve cybersecurity risks.
For independent assessments to remain credible, reviewers must be able to communicate meaningful findings even when those findings are unfavorable to the companies involved. At the same time, researchers may need to withhold specific information that could expose proprietary systems or enable misuse.
Transparency Could Shape Future AI Oversight
The dispute carries implications beyond the two companies. As AI systems become more capable, independent evaluations are increasingly important for regulators, enterprise customers and researchers seeking reliable evidence about model safety.
If companies can suppress or substantially reshape unfavorable findings, outside observers may struggle to distinguish genuinely strong safety performance from results presented under restrictive publication agreements.
The key issue is therefore not simply whether companies can review reports for confidential information, but whether those reviews allow them to influence conclusions or prevent important risks from being disclosed.
Attention will now turn to whether Anthropic’s reported change in approach leads to clearer publication rights for its external evaluators and whether OpenAI provides greater transparency about substantial edits to independent research.
The outcome could help establish how much independence third-party AI safety assessments retain as access to increasingly powerful models becomes more closely managed by the companies developing them.
