Every firing has two stories, and the unusual thing about this one is that the public can't test either. OpenAI says it dismissed Jasmine Wang, Tomek Korbak and Mikita Balesni after an investigation found they breached policies on sensitive information, and that the decision had nothing to do with their safety concerns or public statements. The researchers say the way the dismissals were communicated has left former colleagues afraid to speak or work freely. The investigation was run by the employer, its findings are confidential, and the material at issue is secret by definition. So the company is judge, prosecutor and sole custodian of the evidence.
I don't know what these three did, and nobody outside the company can say. Confidentiality rules at frontier labs are not frivolous. Some information, such as model weights, security vulnerabilities and dangerous-capability evaluations, really should not leak. A safety researcher is not exempt from rules because she believes she is right. But 'information policy' is the most flexible instrument an AI lab owns. In a field where nearly everything of public importance is also proprietary, almost any act of conscience can be framed as a breach. A rule that can be enforced selectively, and judged only by the enforcer, works as a discretionary power to dismiss.
The timing is hard to wave away. The firings came after the researchers called in September for a slower pace of frontier development. The company says the two are unrelated. It may be right. But OpenAI has earned less benefit of the doubt than most employers. In 2024 it was revealed that departing staff were asked to sign agreements that threatened their vested equity if they criticized the company, and the safety-focused Superalignment team dissolved amid public complaints that safety had taken a back seat to product launches. A company with that history cannot just say 'trust our investigation' and expect it to settle anything.
The researchers addressed their letter to the board and its safety committees, the same institutions the company points to as proof of its seriousness. That shows how thin the structure of accountability is. These committees are staffed and appointed from inside the corporate perimeter. They are fine for questions the company wants answered, but they are poorly suited to the one question here: was a dissent punished? An internal body can't credibly review its own company's decision to fire the people raising alarms.
There is also a bitter irony in the company's statement, which stresses its commitment to the 'monitorability' of advanced models. OpenAI wants the ability to see what its systems are thinking and to catch dangerous behavior early. A lab that wants to monitor machines but makes its human researchers afraid to talk is building the wrong kind of transparency. Whatever else the dismissals show, the chilling effect on remaining staff is real, and the people with the best view of the risks are the ones who now have the most reason to keep quiet.
This is ultimately a labor question dressed in technical language. Engineers at AI labs are well paid, but they have no collective voice, and their leverage is individual and fragile: equity, reputation and the willingness to walk out. Pay does not make up for the lack of rights. California's recent frontier-AI law added whistleblower protections for disclosures about catastrophic risk, which is a start. The principle needs to go further: a protected channel to an outside regulator, independent adjudication when a safety employee is fired, and a presumption against the employer when the dismissal follows public dissent. Without that, the public learns about AI risk only from what companies choose to release, and dissenters are treated as a legal liability.
If these three did breach rules that protect something important, an independent review would show it, and the company would benefit. If they didn't, we need to find out. In either case, no one should have to take the company's word for it.
