OpenAI should commission an independent review of its safety researchers’ dismissals

“We have become concerned that internal and external communications around our firing have made our former colleagues afraid to speak and operate in ways that, until last week, were an integral part of working at OpenAI,” the researchers wrote in an open letter reported by Engadget.

That is a serious concern, but it is the former researchers’ assessment. The material available does not include a current employee account confirming that people have become less willing to raise safety issues, or data showing a change in reporting or outside collaboration. A warning about a possible chilling effect should not be mistaken for proof that one has occurred.

Korbak said he was told his dismissal concerned how he communicated with METR, an independent nonprofit that evaluates AI systems and was brought in to investigate a separate incident. He said he received no further details. That account does not establish what was communicated, what rule applied or whether the rule was broken.

OpenAI, for its part, says its investigation found violations of its policies for handling sensitive information. The company also says the dismissals were not about safety concerns or speaking out. Those statements establish the company’s position; without a more specific account of the finding and its basis, the public cannot assess how it fits with the researchers’ version.

Confidentiality and safety work need clear boundaries

The strongest case for OpenAI’s position is straightforward: companies need to protect sensitive information, and employees may face legitimate limits on what they can share outside the organization. Working with an independent evaluator does not automatically make every communication appropriate. If the company found a genuine policy violation, it should be able to address it.

But those boundaries matter especially in safety research, where external evaluation and communication with outside specialists can be part of identifying risks. Staff need to know in advance what information they may share, with whom, and through what channels. If a rule is applied after the fact in a way employees could not reasonably anticipate, the result may be uncertainty even without retaliation.

The public account does not resolve that question. It does not establish what conduct the investigation examined, which provisions were applied or how OpenAI communicated its expectations to employees working with external groups. Nor does the dispute establish that the researchers’ conduct was consistent with those rules. Both points require evidence, not inference from the fact of dismissal alone.

That distinction matters beyond this personnel dispute. OpenAI’s ability to enforce information safeguards and its ability to hear good-faith safety concerns are not competing goals. Clear rules and a credible process are how a company can pursue both without asking employees or the public to take its assurances on trust.

A process-level review would help

Transparency need not mean releasing confidential material or publishing private employment records. OpenAI could explain the relevant rules, how they were communicated, what general category of conduct the investigation found, and what review process was used—while protecting sensitive details and individual privacy.

A reviewer independent of the decision-making chain should also examine whether the rules governing outside safety work were clear and whether employees have workable channels for raising concerns. OpenAI could publish a summary of that review, including any changes it plans to make. An account of the process would not automatically prove or disprove retaliation, but it would give employees and the public a basis to judge the company’s safeguards.

The researchers have raised a question about the conditions for safety work, not demonstrated that colleagues have fallen silent. OpenAI should answer that question with an independent review and a public, appropriately redacted account of its findings. That is a proportionate response to a disputed decision—and a better basis for trust than asking the public to choose between two accounts it cannot verify.