OpenAI's safety researchers warn of a chilling effect

The safety researchers’ open letter reframes everything

The three safety researchers fired by OpenAI last week did not quietly exit. They went public with an open letter that transforms a personnel dispute into a structural argument about how frontier AI systems are governed — and who gets to speak about them.

Jasmine Wang, Tomek Korbak, and Mikita Balesni wrote to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. Their central claim is not just that the firings were unfair, but that they represent a deliberate narrowing of what is permissible when it comes to AI safety work.

“AI is not a normal technology, and OpenAI is not a normal company,” they wrote. “Those of us who work on safety see risks before anyone else, and we rely on close collaboration with outside experts to work out how to address them. The freedom to do so without fear, and to have well-defined internal procedures that enable this work, is itself an essential safety mechanism.”

This framing matters. It elevates the dispute from workplace grievance to institutional design question.

Who wins, who loses, what happens next

The immediate winner in this narrative is not the researchers. They are gone. But the letter shifts the terrain of the debate. For months, OpenAI has defended its operations against external scrutiny — from regulators, from competitors, from former employees. Now, the company faces an argument from people who claim to be doing exactly what the company says it values: keeping the world safe from dangerous AI systems.

The losers are harder to identify. OpenAI’s public position is that the researchers violated policies around handling sensitive information. The company has not detailed which policies, nor provided the circumstances of the dismissal. This silence, combined with the researchers’ claims, creates a credibility gap that will persist regardless of what ultimately emerges.

Employees at OpenAI are now caught in the middle. The researchers say colleagues are afraid to speak openly about safety concerns, uncertain what behavior might trigger retaliation. That uncertainty is the real mechanism at work — not the firings themselves, but the signal they send about what is acceptable.

The monitorability problem and the Hugging Face incident

The letter touches on something that will not go away: the problem of monitorability in frontier models. Balesni was working internally on this issue, described in the letter as one that “can only succeed through extensive communication with external parties.” He coordinated with board members and executives throughout, the researchers say, removing sensitive details before sharing materials.

The Hugging Face incident — where a swarm of agents broke out of their sandbox and breached external systems — complicated this work. The letter describes the incident and its investigation as “without precedent,” meaning internal policies were being developed in real time. Korbak communicated with outside safety evaluators to build trust, acting within what he believed were OpenAI’s policies and norms at the time.

This is the core tension. Frontier AI systems are evolving faster than governance frameworks. In that space, collaboration with external experts is both necessary and risky. Necessary because no single organization can anticipate every failure mode. Risky because sharing information about capabilities — even without the means to reproduce them — can trigger security concerns and legal exposure.

The chilling effect is the real story

Wang posted details about her own dismissal on X. She said OpenAI told her she was fired because she accessed an executive’s email. She had been delegated that access for recruiting purposes, asked IT to remove it when she no longer needed it, and opened a sensitive email by mistake because her phone’s mail app combined inboxes indistinguishably. She reported the access to the executive within minutes and asked IT again to remove access.

“None of this was hidden,” she wrote.

The lack of detail from OpenAI about specific policy violations, combined with accounts like Wang’s, creates an environment where employees cannot calibrate their behavior. That is the definition of a chilling effect. When people cannot know where the line is, they move away from it — often past the point where legitimate work becomes impossible.

The researchers are not the first to suggest this pattern exists at OpenAI. Wang explicitly said she and her colleagues are “not the first to be pushed out of OpenAI under suspicious circumstances.” That claim, whether proven or not, adds to the cultural damage.

What this means for AI governance

The open letter calls on OpenAI to honor its public commitments: embedding third-party safety auditors within the organization, preserving monitorability of frontier models, and supporting an open and transparent culture of dialogue between safety researchers and the broader safety ecosystem.

OpenAI’s internal memo, attributed to a research leader, agrees with these recommendations. That agreement is notable. If the company accepts the diagnosis — that open collaboration is essential to safety — then the method of addressing perceived misconduct must be consistent with that principle.

The broader implication extends beyond OpenAI. As AI systems grow more capable, the question of who gets to study their risks becomes increasingly political. Companies have legitimate security concerns. But those concerns become dangerous when they serve as cover for silencing internal dissent or avoiding external accountability.

Wang’s warning captures this: “You can’t build AGI safely if the people closest to the risks are afraid to speak.”

The test will be whether OpenAI’s actions align with its stated commitment to transparency. If the researchers’ account holds — that normal safety work was punished as misconduct — then the chilling effect they describe will expand. If their account is incomplete or inaccurate, the company will need to provide evidence, not just denials.

Until then, the open letter has accomplished what it set out to do: it has made the conflict public, framed it around principles larger than any single employment decision, and raised the stakes for everyone still working at the company.