Three former OpenAI safety researchers have published a letter contesting the account of their dismissals and asking the company’s oversight bodies to protect independent safety collaboration. Their argument concerns both the employment dispute and the conditions under which researchers can raise difficult questions.
The letter is a first-party account, not an independent determination of what happened. OpenAI told TechCrunch it denied retaliation and alleged policy violations. The public material reviewed here does not settle those competing accounts.
The researchers want clearer operating rules
In their letter, Tomek Korbak, Jasmine Wang and Mikita Balesni dispute misconduct allegations and describe uncertainty over acceptable collaboration with outside safety organizations. They warn that the circumstances of their departure could discourage remaining colleagues from speaking openly.
Their recommendations are specific: preserve substantial access for independent evaluators, avoid development choices that erode model monitorability, and clarify procedures for internal and external safety dialogue. The authors argue that these conditions are part of safety work itself, rather than merely workplace preferences.
It is possible to examine those proposals without deciding the employment dispute. Conversely, agreement with a proposal would not prove that the company’s allegations were wrong. Keeping those questions separate is essential when a public letter combines personal explanations with institutional demands.
The test for the proposals is operational. Who may see relevant systems, what can be examined, and how are disagreements handled? Without those details, a commitment to collaboration can mean very different things to the people expected to rely on it.
Monitorability is useful but not guaranteed
A cross-industry research paper co-authored by the three researchers explains why monitoring a model’s chain of thought is considered a promising safety technique. Its premise is that a reasoning trace can sometimes reveal problematic intentions or behaviour that an observer would miss by inspecting only final actions.
The paper also calls this opportunity fragile. A monitor can fail to notice misconduct, and training choices may affect how informative the trace remains. The authors advocate research and preservation of monitorability alongside other safeguards, not treating it as a complete solution.
OpenAI’s own monitorability evaluation release provides a practical point of comparison. Published in April, it supplies datasets and reference code for selected evaluations. The company describes using reasoning traces to detect behaviour such as reward hacking and says it avoids strongly optimizing directly over those traces.
The release is deliberately incomplete: some private or restricted datasets are omitted, and the reference scaffold is not presented as a fully supported evaluation framework. That limits what outside users can reproduce from the public materials alone.
These documents support the importance of the research question. They do not establish that any particular dismissal caused monitorability to decline. A technical concern, a claim about organizational culture and a finding about a specific model require different evidence.
External research does not always mean internal access
OpenAI’s Safety Fellowship announcement illustrates another form of collaboration. The program runs from September 2026 into February 2027 and offers external researchers support, mentorship and compute resources for work such as papers, benchmarks and datasets.
The company says fellows receive API credits but not access to internal systems. That is a meaningful distinction from the deeper evaluator access sought in the former employees’ letter. Funding research about a system and allowing inspection inside its development process are different arrangements.
Neither arrangement automatically substitutes for the other. Public-facing research can explore questions that an API makes observable. Internal access may be necessary to inspect material unavailable through that interface. What matters is matching the access to the question being asked, rather than counting collaborations as interchangeable.
The unresolved issue is therefore broader than whether OpenAI works with outsiders at all. It is whether independent researchers have enough access, clarity and freedom to test the safety claims that require scrutiny. The letter raises that question; the available public evidence does not yet answer it.
