White House Accord Lifts AI Evaluators — Without Demands for Reporting or Transparency
In just a few weeks, the independent AI evaluation institutions have gone from footnote to central players: funding pledges measured in the tens of millions of dollars, direct mention in the White House's voluntary accord, and promises from the labs to build them into their own operations. At the same time, OpenAI fired three safety researchers — two of whom say the dismissals are connected to their contact with third-party evaluators. The ecosystem meant to scrutinize AI labs is thus being funded and formalized now, before the fundamental questions of who pays, to whom evaluators report, and whether findings become public have been answered.
The policy trigger: an accord that points, but does not bind
In late September, the president presented a voluntary accord in which he urged AI companies to "partner with an independent external auditor or evaluator" (CNBC). The language matters because it lifts third-party evaluation into the official safety policy around frontier AI.
But the accord is, as the Brookings analysis of it describes, silent on exactly what determines whether evaluation becomes meaningful. According to the analysis, nowhere does it discuss basic issues such as to whom the evaluators report, who pays them, and whether they have the ability to intervene and stop harmful practices they discover. There is also no requirement that the auditors' findings be made public — or even handed over to any responsible public interest agency (Brookings).
That means the accord creates an expectation of external scrutiny without creating any structure for it. Reporting lines, funding, intervention powers, and transparency are left to the labs themselves — the very companies being evaluated.
The labs promise to let the evaluators in
Industry has nonetheless moved quickly. Anthropic CEO Dario Amodei promised last month to embed independent evaluators in his company — a pledge OpenAI CEO Sam Altman quickly endorsed, according to CNBC (CNBC). OpenAI stated in a post on Friday that it is "actively finalizing contracts with third-party safety assessors and will announce details in the coming weeks."
So far, these are company statements, not published contracts. It is unknown who will pay, what access rights evaluators will get, whether findings can be published, and whether evaluators can speak freely. In other words: exactly the questions Brookings flags as unanswered in the accord are also unanswered in the labs' pledges — until the contracts become known.
The OpenAI firings: an early test of independence
While the pledges were being made, another event took place that may become a test of the entire arrangement. OpenAI last week fired three employees for "violating our policies on accessing and handling sensitive company information," according to a spokesperson. Two of those dismissed, Mikita Balesni and Tomek Korbak, have said they believe they were fired because of the way they communicated with third-party evaluators, according to their own posts on X as reported by CNBC (CNBC).
OpenAI rejects that account. A spokesperson told TechCrunch that the three were fired after an investigation uncovered a "pattern of misconduct" in "clear violation of our policies of mishandling research information," and that this goes beyond sharing information with an external AI evaluation group. The company has not specified which policies were violated (TechCrunch).
The two accounts cannot be reconciled from the available source material, and this article does not assume either is correct. But for the evaluation question, it is almost irrelevant who is right: the case shows that an employee's contact with external evaluators can become the center of a dismissal dispute, without the rules for such contacts being clear. Balesni and Korbak write together with Jasmine Wang in an open letter: "Given the significant safety concerns surrounding the development of AI, employees must not be left working in an environment where fear and unclear rules stymie AI safety work and weaken third-party accountability." They also warn that dismissals like theirs, "executed and communicated so abruptly," chill the open culture OpenAI has prized (TechCrunch).
If lab employees risk dismissal for passing information to evaluators, the evaluators' actual access becomes decisive — and that is decided by the contracts OpenAI now says it is finalizing. An open exchange about what is permitted does not yet exist.
Follow the money: from $13.6 million to over $100 million
However the disputes are resolved, the money is on its way. Two figures illustrate the scale.
Vals AI, a commercial evaluation company, has according to its CEO Rayan Krishnan grown from eight employees to around 30 this year, and announced in August a funding round of $40 million, according to reporting in Market Briefs (Briefs).
METR, a nonprofit organization, said in August that it had secured around $71 million in pledges over six months — compared with a total of $13.6 million in contributions for all of 2024, according to the organization's latest IRS reporting. METR says it declined payment for the work in question (Briefs).
It is worth being precise about what the figures are: these are company and organizational statements relayed through secondary sources, not publicly audited accounts. And in METR's case, it has not been stated whether the roughly $71 million comes from sources independent of the AI labs — exactly what is central to the funding-conflict question. The organization's statement that it declined payment for one specific assignment says nothing about its overall funding structure.
Andrew Freedman, who leads the policy think tank Fathom, expects an "influx of capital" into the ecosystem. "I've worked in politics and policy for the last 20 years of my life, and I've never seen an issue move so fast on so many different political spectrums," he says (Briefs).
The independence problem: who pays the judge?
All the figures above point to the same structural problem, which Suresh Venkatasubramanian, a professor of computer science at Brown University, has put into words. "To a degree, the problem, as always, is money," he told CNBC. "Who is paying for these companies to do their work? How are they going to support them? You need an ecosystem, you need a viable business model for this" (CNBC).
And: "If you want true third-party evaluation, you need true independence financially and otherwise," he says (Briefs).
The math is simple: if the labs are the natural payers — because they are the only ones with the models the evaluation concerns, and because the accord urges them, not the public, to procure evaluators — then the industry faces a principle-versus-practice dilemma. A nonprofit like METR can decline payment in individual cases, but $71 million has to come from somewhere. A commercial company like Vals AI needs customers, and the natural customers are precisely the labs. None of the sources point to any public funding scheme that could break this dependency.
There are known models for this in other industries — the audit industry's long history of independence problems is the most obvious parallel, where the client paying for the audit is also the one being scrutinized. But the audit industry has legislation, public reporting requirements, and oversight bodies. AI evaluation has, as of now, none of these: the Brookings analysis states that the accord requires neither publication nor handover to public agencies (Brookings).
What to watch
The coming weeks offer three concrete benchmarks for whether the arrangement becomes real independence or a symbol:
The contracts OpenAI announces. The company says details will come in the coming weeks. What matters is not that contracts are signed but the terms: who pays, what access evaluators get to models and internal information, whether they can publish findings without prior approval, and whether they can speak when they find something serious. If the contracts are kept secret, the accord's weakness is repeated at the contract level.
Where METR's $71 million comes from. If the pledges to a significant degree come from AI labs or from donors closely tied to them, financial independence — in Venkatasubramanian's sense — remains unfulfilled, however large the sums. If they come from foundations or public sources independent of the industry, that changes the basis.
Whether the accord gets teeth. The Brookings critique points out that without requirements for reporting and publication, evaluation remains voluntary in practice. Any move in that direction — legislation, regulation, or voluntary add-ons with reporting obligations — would be more structurally significant than more funding rounds.
In the meantime, the OpenAI firings are a reminder that independence is not only about money, but about access and culture: an evaluator without information flow from employees willing to speak has little to evaluate. Whether the dismissed researchers were right, or whether OpenAI's account of a broader pattern of misconduct holds, is unresolved. But the case shows that the question of whom the AI judges owe their loyalty — the employer that pays, or the public that needs the scrutiny — is now being decided in practice, through contracts and dismissals, long before it is answered in principle.