Fired OpenAI Safety Researchers Demand Inspectable AI

Jasmine Wang, Tomek Korbak, and Mikita Balesni dispute leak allegations and call for a freeze on models that hide their reasoning steps.

The sudden firing of three OpenAI safety researchers earlier this month brought an simmering dispute into the open. Jasmine Wang, Tomek Korbak, and Mikita Balesni were dismissed over alleged violations of confidential data policies. Their response arrived quickly: a public letter to OpenAI’s board urging the AI industry to freeze the deployment of frontier models whose reasoning traces cannot be inspected by human evaluators. The conflict highlights growing friction between release deadlines and technical oversight, according to reports from the Associated Press, the BBC, and TechCrunch.

Conflicting accounts behind the dismissals

OpenAI framed the dismissals as a clear disciplinary issue. In statements to journalists and on X, company representatives maintained that the researchers committed a significant breach of trust by mishandling proprietary materials. The company explicitly rejected claims that the firings had anything to do with internal safety debates or technical dissent.

The researchers paint a starkly different picture. They argue they were dismissed in retaliation for flagging security gaps and maintaining technical dialogue with METR (Model Evaluation and Threat Research), a non-profit risk evaluation group that has evaluated OpenAI models in the past. Korbak posted on X that managers verbally linked his termination to communications with METR. In their joint letter, titled “OpenAI cannot make AI safe on its own”, the group warned that abrupt dismissals create an aggressive chilling effect across the entire laboratory.

The technical hazard of opaque reasoning

The core technical issue at stake involves how next-generation frontier models think. Korbak and Balesni recently co-authored research alongside teams at Google DeepMind and Anthropic examining the monitorability of chain-of-thought reasoning. Current reasoning models write out intermediate thoughts in natural language before delivering a final answer. Those legible scratchpads give safety researchers an essential window to spot deception, logical fallacies, or misaligned behavior before deployment.

That window may close fast. Experimental architectures using recurrent depth or loop transformers, as detailed by The Information, run multiple reasoning loops across internal latent layers without outputting text tokens. When computation happens entirely inside latent space, evaluators cannot see how the system arrived at a conclusion or whether it masked a dangerous failure mode during the process.

What the OpenAI safety researchers demand

To prevent unreadable models from reaching production, the researchers outlined three explicit demands for OpenAI and the wider AI sector:

Preserve inspectable reasoning: halt deployments of frontier systems whose internal problem-solving steps cannot be read by human evaluators.
Embed independent auditors: integrate third-party evaluators directly into laboratory development cycles rather than relying on internal reviews.
Protect research collaboration: keep secure channels open with external safety institutions without threat of reprisal.

A recurring pattern in frontier labs

The dismissals are part of an ongoing exodus from OpenAI’s safety infrastructure. Since mid-2024, the lab has seen key departures, starting with the dissolution of its Superalignment team following the resignations of Jan Leike and Ilya Sutskever. Senior researchers Leopold Aschenbrenner and Daniel Kokotajlo left shortly after, as tracked by NPR. With commercial competition driving laboratories toward deeper reasoning models, the friction between shipping fast and keeping systems transparent is becoming harder to paper over.