OpenAI defends firing three AI safety researchers

OpenAI has dismissed three safety researchers — Jasmine Wang, Tomek Korbak, and Mikita Balesni — for what the company calls a "significant breach of trust," according to a Friday post on X. The company is doubling down on the firings after the group publicly appealed for transparency about their removal.

The post, attributed to OpenAI's Newsroom account, says the three violated "clear policies on handling sensitive information." OpenAI insists the dismissals had nothing to do with the researchers speaking out about AI safety. The statement is a direct rebuttal of an open letter the researchers circulated one day earlier.

What is OpenAI's stated reason?

OpenAI's X post frames the dismissals as the result of an internal investigation. The post does not specify which policies the three researchers allegedly violated, nor does it describe the evidence behind the finding. The phrase "significant breach of trust" — a term that recurs in HR (human resources) disputes over employee misconduct — is offered without further definition.

The company makes two specific assertions in the post:

• The decision followed an investigation into a "significant breach of trust."
• The decision was not retaliation for the researchers' public concerns about AI safety.

OpenAI does not name the cleared policies. It does not publish the investigation's findings. It does not say when the alleged conduct occurred, or who else inside the company knew about it. The post also does not say whether the three were given an opportunity to respond to the findings before being dismissed.

What did the researchers say in response?

On Thursday, the three published an open letter, hosted at mikitabalesni.com, urging the company to be more transparent about the decision. Each researcher also posted individually on X the same day. Wang, Korbak, and Balesni argued that the firings followed their public criticism of OpenAI's safety practices, not any misconduct on their part.

The open letter sets out the group's account of events. It is, by its own framing, a request for OpenAI to publish a fuller version of those events. The dispute now sits at the stage of competing narratives.

How do the two sides frame the disagreement?

OpenAI says the firings were about policy. The researchers say the firings were about speech. The two accounts contradict each other directly, and OpenAI has not yet released evidence that would let outside observers adjudicate between them. The company has, in effect, asked the public to take its word that an internal review found a breach — without showing the review.

A skeptical reading: when a company fires researchers who have publicly criticized it, and then says the firings had nothing to do with that criticism, the burden of proof sits with the company. OpenAI has not yet met that burden.

What is missing from the public record?

Several pieces of context are absent from both OpenAI's post and the researchers' public statements:

• The specific policies the three allegedly breached
• The dates on which the alleged violations occurred
• The chain of custody for any sensitive information involved
• The reporting structure of OpenAI's safety team
• The status of any internal appeals the researchers may have filed

OpenAI's post also does not address whether the alleged breach involved model weights (the numerical parameters that shape a model's outputs), training data, or internal evaluations. Each of those categories would imply a different severity, a different set of affected parties, and a different set of mitigations. The phrase "sensitive information" could plausibly cover any of them — or none of them, in a way that lets OpenAI defer the question.

The researchers' open letter, as excerpted in public coverage, is itself partial. It sets out their account but is also a call for OpenAI to publish more. Neither side has yet produced a documentary record that a third party could check.

Why does this matter?

The contested firings arrive at a moment when external reviewers — including regulators, commercial partners, and outside auditors — already lack direct visibility into how AI labs investigate safety claims internally. When the company and its safety researchers publicly disagree about what just happened, neither party can easily substitute for the other. The unresolved question — what specifically did the three researchers do, and what evidence supports that finding? — sits in that gap.

The stakes go beyond personnel. AI safety research depends on a small number of labs that can both train frontier models (the largest and most capable AI systems in active development) and review their own behavior. When those labs dismiss their own safety researchers, the pool of people who have seen the relevant evidence shrinks further. Outside auditors cannot replicate what they cannot see.

For commercial partners evaluating OpenAI's safety posture, the unanswered questions carry weight. A researcher dismissed for cause — defined here as misconduct or policy violation — is one kind of personnel story. A researcher dismissed for speech is another. Both carry different implications for the credibility of OpenAI's public safety commitments, and for how partners should price the risk that internal review at the lab can be trusted.

The bottom line: contested departures from major labs reshape the public record of who can be believed about internal safety practices, and each case leaves the burden of explanation higher for the next one.

What happens next?

OpenAI has not announced an independent review of its decision. The three researchers have not, according to the source coverage, filed a public legal claim. The open letter's call for transparency has not yet received a substantive reply. The dispute sits at the stage of competing public statements, with the underlying facts still held inside OpenAI's internal channels.

The next inflection will arrive when one of the two sides produces a document the public can check. Until then, the dispute is a contest of credibility rather than a contest of evidence. For a company whose commercial value rests substantially on trust in its safety work, that is the worst place to leave a fight with its own researchers.