Anthropic and OpenAI deny seeking veto rights – but OpenAI admits removing details
A Reuters investigation published on October 13 shows that both Anthropic and OpenAI, as recently as October 5, reportedly demanded strict editorial control over what safety evaluators could publish – less than three weeks after the companies publicly pledged the opposite. The information rests on two anonymous sources familiar with the private negotiations, as reported by TechTimes. Both Anthropic and OpenAI dispute parts of the account, but the story lands in the middle of an ongoing process in which OpenAI says it is in the process of finalizing contracts with external evaluators, and in which the framework for who pays – and thus who controls – future AI evaluation remains unsettled.
What Reuters claims, and what the companies say
In September 2026, Anthropic CEO Dario Amodei wrote in a widely read essay that external evaluators should be able to publish their core findings without editorial control from the company – a pledge at the heart of the embedded evaluator framework he himself proposed (TechTimes).
According to Reuters' investigation, based on two sources familiar with the private conversations, however, both Anthropic and OpenAI as recently as October 5 – less than three weeks after the public pledges – sought "strict editorial control" over what safety evaluators were allowed to release publicly.
Both companies have responded to the account:
- Anthropic disputed the characterization to Reuters, saying that the reported pursuit of significant control did not give an accurate picture of all the conversations the company had had with independent evaluators. In ongoing conversations with evaluators it has previously worked with, the company said in its statement that it has made clear that evaluation methods, criteria, conclusions and publication decisions rest with the evaluators themselves.
- OpenAI said it did not seek to approve entire reports, but acknowledged having removed certain details from at least one evaluation – while evaluators would be allowed to flag where such editorial interventions had occurred.
The story is fresh, and it is worth holding on to what is actually documented: the claims about the veto demands come from two anonymous sources, and neither company has confirmed the demands in full. At the same time, neither fully denies editorial considerations – OpenAI, as noted, concretely acknowledges having removed details from at least one evaluation.
The embedded evaluator and the conflict of interest the company itself points to
The context is a framework launched on September 18: Anthropic announced a partnership with Accenture – specifically Faculty, Accenture's AI business – on the first embedded evaluator team. Each company committed to investing at least $1 billion over five years, with Faculty's team set to carry out the model testing.
The same day, the AI Evaluator Forum published an open letter, which grew to more than 200 signatories – including Geoffrey Hinton – laying out concrete independence standards for any such framework. The letter warned specifically that evaluators funded directly by the labs they assess would face structural pressures that no amount of good faith can fully counteract.
Anthropic acknowledged the underlying problem in its own press release: funding the work one commissions oneself, the company wrote, "creates an apparent conflict of interest." The company said that in the long run it envisions evaluations being funded from pooled funds or public sources. But such a funding model is so far only a vision – not an established arrangement.
Faculty CEO Marc Warner put the stakes this way to Reuters: "Evaluators are only useful if people can trust what they report, so independence and transparency are critical."
Voluntary oversight without enforcement
The framework around the evaluator arrangements is fundamentally voluntary. In late September, President Trump presented a voluntary accord urging AI companies to "partner with an independent external auditor or evaluator." There is thus no formal oversight body that defines what independence entails, or that can enforce it – the power of definition lies with the companies and the evaluators they negotiate with.
In this context, OpenAI said in a post on X, as reported by CNBC, that the company is "in the process of finalizing contracts with third-party safety assessors and will announce details in the coming weeks." That means the veto question is not a historical side note: the terms of these very contracts will determine how much the voluntary arrangement is actually worth.
An adjacent conflict – but not the same story
The debate over evaluation independence is unfolding in parallel with a separate OpenAI matter. The company recently fired three employees – Mikita Balesni, Tomek Korbak and Jasmine Wang – who claimed they were punished for speaking out about safety or for working with external researchers. OpenAI disputes the account.
Korbak himself stated on X that he was told he was being let go because of the way he communicated with METR – an independent nonprofit evaluation company that OpenAI brought in to examine an incident related to Hugging Face (AP News; NPR). The fired employees also expressed concern that the company could roll back a recent safety commitment.
That matter concerns communication with an external evaluator, not the veto demands in the Reuters investigation, and the two stories should be kept separate. They do, however, point in the same direction: the relationship between the labs and those meant to assess them is, in practice, conflict-ridden and negotiated behind closed doors.
What remains unresolved
Several central questions remain open. The scope of the reported veto demands is not fully documented, and it is unclear whether Anthropic's reportedly softer stance in the later conversations reflects a genuine policy change, applies only to certain relationships, or is a response to Reuters' inquiries. The funding model – pooled funds or public money – remains, for now, only a proposal. And the details of OpenAI's forthcoming evaluator contracts will show whether the public September pledges are reflected in the actual terms.
Until then, what is established is simply this: the pledges of editorial independence were made publicly in September, and two sources tell Reuters that both companies were negotiating for something substantially different as of October 5. What ends up in the final contracts will show which version was the real one.