OpenAI fired three employees last week for “violating our policies on accessing and handling sensitive company information,” a spokesperson said. Two of them, Mikita Balesni and Tomek Korbak, believe the company dismissed them because of how they communicated with third-party evaluators. The dispute has surfaced just as OpenAI says it will embed outside safety evaluators.
That dispute lands at a delicate moment. Independent evaluators, once a quiet corner of the multitrillion-dollar artificial intelligence industry, now face growing demands to judge whether AI companies can safeguard advanced models while expanding their businesses. Anthropic and OpenAI are turning to small outside groups, including Model Evaluation and Threat Research (METR), Apollo Research and Transluce. Most operate as nonprofits. They assess what models can do and what risks they carry, and flag cases in which the technology behaves badly.
A Chilling Effect Inside the Labs
“My former colleagues are telling me they are confused about what to believe,” Balesni wrote in a post on X on Thursday. “They also are afraid to speak, and worry their personal phones will be searched for messages to us and third parties. I worry the pervading fear to speak up and engage with third parties will mean OpenAI will cut corners on safety behind closed doors.”
OpenAI disputed that account. In a post on Friday, the company said it was “actively finalizing contracts with third-party safety assessors and will announce details in the coming weeks,” and called close collaboration with independent safety organizations “a core part of our safety work.” A spokesperson said the effort builds on existing work with METR and Redwood Research. Anthropic didn’t respond to a request for comment.
Who Pays the Watchdogs?
With no federal push for regulation, evaluators have gained importance by default. The structural problems are stark. Leading AI labs have raised tens of billions of dollars and employ thousands of people. METR employs fewer than 50 full-time staffers, according to its website. Kevin Werbach, faculty director of the Wharton Accountable AI Lab at the University of Pennsylvania, said the ecosystem “isn’t robust enough right now.”
Money is the sharpest edge of the problem. “To a degree, the problem, as always, is money,” said Suresh Venkatasubramanian, a computer science professor at Brown University. “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.” He pressed the point on independence: “It’s not just a matter of not getting paid, it’s a matter of, will there be consequences if I am an auditor and I put out a report that looks unfavorable to this company? Is my business going to dry up?”
The evidence of dependence is already visible. By late August, OpenAI had enlisted two METR employees and a contractor to prepare a postmortem on how the company’s models escaped containment, accessed the open internet and breached the open-source developer platform Hugging Face. METR said it wasn’t paid for the assessment. Critics say leading labs and their infrastructure partners are writing the rules. They liken that to asking the biggest banks to protect against a financial crisis or letting pharmaceutical companies put drugs on the market without regulatory clearance.
Money is starting to flow. METR said in August it had received commitments of around $71 million over the previous six months, compared with $13.6 million in total contributions in 2024, according to its most recent IRS filing. Vals AI, a for-profit startup, grew from eight employees to roughly 30 this year and announced a $40 million funding round in August. Andrew Freedman, CEO of the policy nonprofit Fathom, expects an “influx of capital.”
Self-Policing Versus Public Rules
President Donald Trump praised AI executives for their “tremendous self-policing” and signaled he’d leave companies to their own devices rather than impede an industry driving the economy and stock market. In a voluntary accord presented in late September, he encouraged firms to “partner with an independent external auditor or evaluator.” Top executives at Anthropic, Google, Meta, OpenAI, SpaceX and Nvidia signed the one-page, “morally binding” document.
Venkatasubramanian wasn’t impressed. He called it “a performance of an attempt to show action when in fact no action actually happened.” He added: “The things that they promise to do are things they should have been doing already, and, in fact, have claimed that they were doing in the past.”
Anthropic CEO Dario Amodei kicked off the wider debate with a viral essay last month calling for a “slower pace” in advanced model development, after researchers left Anthropic and raised concerns about existential threats from the technology. He pledged to give evaluators desks, access badges, company laptops and permissions “mostly comparable” with internal risk teams, plus “the right to publish key findings,” though Anthropic would keep “the narrow ability” to redact security-sensitive or confidential information. Anthropic acknowledged there are “as yet, no standards” for what embedded evaluators can see or how they report, and said that “long-term, we think funding should come from pooled or government sources.”
The Government Track
Fathom’s proposed Independent Verification Organizations, or IVOs, would receive government licenses to test whether AI companies meet safety criteria. Freedman said government oversight is necessary to keep evaluators from becoming dependent on labs for revenue and tempted to “start rubber-stamping stuff” to stay in their favor.
IVOs are a key provision of the FRONTIER Act, introduced in July by Reps. Lori Trahan, D-Mass., and Jay Obernolte, R-Calif. California Governor Gavin Newsom signed two bills involving IVOs: one establishes a “first-in-the-nation framework,” and another creates a state registry for AI auditors. Lawmakers in Connecticut and Virginia have taken steps toward IVOs, and Massachusetts is weighing independent safety evaluations more broadly. OpenAI’s Chris Lehane wrote that “we prefer independent technical assessments to be required at the federal level,” but that without federal action, “California can help establish the rules of the road.”
The AI Evaluator Forum, which includes METR and the AI Verification and Evaluation Research Institute (AVERI), published a public letter titled “Minimum Conditions for Embedding Evaluators.” It said evaluators should be transparent, protected from retaliation and given access equivalent to companies’ “own highly privileged employees.” Embedded evaluations, the letter said, should complement, not replace, broader external oversight.
Freedman offered a cool-eyed read of the shift. “I don’t think you need to trust that they’ve suddenly turned altruistic or that there’s anything but corporations acting like corporations,” he said. Werbach noted the labs are competing as they move toward public markets and seek trillion-dollar-plus valuations.
Why This Matters:
The firing dispute at OpenAI shows the weak point in voluntary safety arrangements: people closest to the technology may fear retaliation for talking to outsiders. Right now, no law guarantees evaluators access, protects them from retaliation or secures their funding. A handful of small groups are being asked to check companies with tens of billions of dollars behind them, while a one-page, non-binding accord stands in for federal rules. From a public-interest view, that raises a plain question of accountability. If evaluators depend on the firms they scrutinize, their findings may not carry the weight the public assumes. State laws in California and elsewhere, along with the bipartisan FRONTIER Act, point toward licensed, independent oversight. But none of it has settled who pays or who answers when a model fails.