What's trending, explained.
What's trending, explained.
On October 13, Reuters published an exclusive report saying that OpenAI and Anthropic have sought significant control over what outside AI safety evaluators can publish about their models. The report, based on two people familiar with the negotiations, says both companies were still pushing for strict editorial restrictions as recently as October 5 — even though both had publicly committed, weeks earlier, to letting independent evaluators report freely.
The contradiction is the core of the story. In September, Anthropic CEO Dario Amodei said external evaluators should have the right to publish key findings "without editorial control." Days later, OpenAI publicly agreed, saying assessors should have "editorial independence." But according to the Reuters sources, behind closed doors, both labs were negotiating contracts that would give them approval over what goes into the reports.
The companies' stated worry is trade secrets. The current model under discussion involves "embedded evaluation": outside researchers work inside the AI labs, with access to internal systems and training processes, to test models before release. Both companies fear that evaluators with that level of access could carry proprietary information to rivals — a concern sharpened by the fact that both labs are reportedly pursuing public listings that could value them in the trillion-dollar range.
Measures under discussion to prevent intellectual-property leaks include having lab employees observe embedded evaluation teams as they work, and barring evaluators embedded at one lab from simultaneously embedding at a rival lab. According to a contract seen by Reuters, language the labs already use with some evaluation firms can bar them from publishing confidential information without the companies' consent.
Independent evaluation is the main check on AI safety claims right now. These outside reviewers are the people who test frontier models for dangerous capabilities, bias, and misuse risks before the public sees them. If the labs being evaluated can edit or veto the resulting reports, the whole system becomes a trust exercise: readers have no way to know what was changed.
The problem is that the line between a trade secret and a safety finding is blurry. As one person with knowledge of past dealings with the labs told Reuters, what a lab considers crucial intellectual property can overlap with information relevant to judging a model's risk — and that overlap is exactly where arguments and redactions happen.
The companies pushed back on parts of the story. Anthropic told Reuters that seeking significant control did not accurately characterize all of its discussions, and that evaluators it had previously worked with control their own methods, criteria, conclusions, and publication decisions — though it reserves the right to comment on and approve reports from new evaluator firms. OpenAI said it was not seeking approval rights over the full text of reports, that its standard practice is to redact confidential information, and that evaluators can disclose when major redactions were made.
Reuters also reported a concrete example: OpenAI redacted nonpublic information from a report by nonprofits METR and Redwood Research on an incident involving OpenAI agents and Hugging Face. The two nonprofits said they also edited the report's structure, emphasis, clarity, and tone based on OpenAI's feedback, but that nothing else important to their conclusions was removed.
The negotiations are ongoing, and the outcome will set the terms for how AI safety testing works going forward. The key question is whether independent evaluators end up with genuine publication freedom or a version of independence that comes with a lab's red pen.
Watch for a few things. First, whether either company publishes formal evaluator-independence policies with real teeth — deadlines, appeal processes, or third-party arbitration — rather than general pledges. Second, whether regulators or governments start demanding audit rights of their own: the UK's AI Security Institute already tests models before release, and its access was itself disrupted by a reported 17-day White House freeze that was only lifted on October 11. And third, what actually appears in the next round of published evaluation reports: the presence or absence of disclosed redactions will tell you more about these negotiations than any press statement.