Reuters reported on October 14 that Anthropic and OpenAI pushed for veto or sign-off rights over what external evaluators could publish about their AI systems, even as their CEOs publicly pledged to embed independent reviewers inside their companies.
The reporting, covered by TechTimes, CNBC, and AI Weekly, details how both companies sought to retain control over the publication of safety findings, contradicting commitments made in September 2026 to give outside evaluators genuine independence.
The September pledges
In September 2026, Anthropic CEO Dario Amodei pledged to embed independent evaluators inside his company. OpenAI CEO Sam Altman endorsed the idea. President Donald Trump supported it, as did most of the largest US tech companies, according to CNBC.
The premise was straightforward: outside experts would get access to internal systems, conduct evaluations, and publish their findings without the companies they were evaluating having the power to suppress or alter the results.
The push for independent evaluation did not emerge in a vacuum. AI systems have grown more capable and more widely deployed, yet the mechanisms for assessing their safety have lagged behind. Without mandatory government testing regimes, voluntary commitments to external review became one of the few ways for companies to demonstrate that their systems had been scrutinised by parties with no financial stake in the outcome.
Both companies have faced scrutiny over their safety practices. Anthropic has built its brand partly around responsible AI development, while OpenAI has faced questions about the pace of its product releases relative to its safety commitments. The pledge to embed independent evaluators was, in part, a response to that scrutiny.
What the companies sought
According to Reuters, both Anthropic and OpenAI pushed for arrangements that would give them veto or sign-off rights over evaluator publications. This would allow the companies to review, revise, or block findings before they reached regulators, enterprise customers, or the public.
A reported example involved OpenAI cutting material from a METR and Redwood Research report about the Hugging Face agent incident. AI Weekly, summarising The Next Web, reported that it is unclear whether reviewers formally flagged the cuts.
Who the evaluators are
The independent evaluation field includes the nonprofit METR, based in Berkeley, California, and Redwood Research. The AI Evaluator Forum is drafting best practices for the field, working to establish norms and standards for how evaluations should be conducted and published. Both METR and Redwood Research have conducted evaluations of frontier AI models, and their work has been cited by regulators, enterprise customers, and other stakeholders seeking to understand the capabilities and limitations of systems deployed by major labs.
Ryan Stosz, now head of governance at Transluce, commented on the ambiguity of the embedding concept. “It’s a little ambiguous what embedded evaluators means,” said Stosz, whose company revealed that OpenAI agents hacked US and Australian government websites. Evaluators are interested in embedding with labs for greater access, but the lack of clarity creates tension between the desire for transparency and the companies’ desire for control.
Why independence matters
If the evaluated company can review, revise, or block findings before they reach regulators, enterprise customers, or the public, the report is not an independent assessment. It becomes a corporate communication dressed in the language of third-party validation.
The concern is not hypothetical. Enterprise customers rely on safety reports to decide whether to deploy AI systems in high-stakes settings such as healthcare, financial services, or critical infrastructure. Regulators look to independent evaluations to understand risks that companies may understate. The public depends on unfiltered findings to understand the capabilities and limitations of systems that are increasingly woven into daily life.
When a company can veto findings, the evaluation process risks becoming a public relations exercise rather than a genuine check on safety.
The outsized role of evaluators
CNBC noted that without a federal push for regulation, evaluators have taken on outsized importance. In the absence of government-mandated testing and disclosure requirements, the independent evaluation field has become one of the few external accountability mechanisms for AI companies.
This raises the open question of who pays for independent evaluations and whether a viable business model exists to sustain them. If the companies being evaluated fund the evaluations, the financial relationship creates an inherent conflict of interest. If evaluators rely on grants or donations, their long-term sustainability is uncertain.
The tension between access and independence is unlikely to resolve quickly. Labs want evaluators to have deep access to their systems, but that access can become leverage that companies use to shape what evaluators publish.
What enterprises should ask vendors
Companies evaluating AI vendors can ask direct questions about third-party testing and publication rights. The following questions, framed as general enterprise security practice, can help clarify the degree of independence in a vendor’s safety reporting:
- Which external evaluators are embedded with the lab, and what formal agreements govern their access and publication rights?
- Does the vendor retain veto or sign-off authority over evaluator reports before publication?
- Are evaluator findings published in full, or only in redacted or summary form?
- Who funds the evaluations, and what financial relationships exist between the evaluators and the vendor?
- Can the enterprise review unredacted evaluation reports, including negative findings, under non-disclosure?
Vendors that welcome genuine independence will answer these questions directly. Vendors that resist publication of unfiltered findings, or that insist on reviewing and approving every word before release, are signalling that their evaluations serve the company’s interests more than the customer’s.
The path forward
The AI Evaluator Forum’s work on best practices is a step toward standardising expectations. Clear norms around publication rights, funding transparency, and conflict-of-interest disclosures would give enterprises and regulators a basis for distinguishing genuine independence from sponsored commentary.
Until those norms are established, enterprises should treat vendor-commissioned evaluations with healthy scepticism and insist on reviewing primary source material wherever possible.
Sources
- TechTimes – Anthropic, OpenAI pushed for veto rights over safety reports while pledging independence
- CNBC – AI’s quiet safety gatekeepers are stepping into the spotlight
- AI Weekly – Anthropic, OpenAI sought control over reviewer publications
