OpenAI Stands by Firing of Three Safety Researchers, Spotlighting Limits of Outside Scrutiny

OpenAI said on October 9 that it is standing by its decision to fire three safety researchers. The statement, posted from the company’s official account, came a day after the researchers published an open letter. With the two sides offering conflicting accounts of why the dismissals happened, the question coming into focus is how far outside researchers can examine the safety of AI systems.

Why were the three researchers fired?

The dismissed researchers are Jasmine Wang, Tomek Korbak, and Mikita Balesni. All three worked on making AI behave safely and as intended, and their employment with OpenAI ended last week. OpenAI said an internal investigation found violations of its rules on handling sensitive information, along with a serious breach of trust. The company also said the firings were not because the researchers raised safety concerns or spoke publicly.

The researchers pushed back, saying they acted within the work practices in place at the time and within the authorization they had been given. They also warned that the episode could have a chilling effect on collaboration with outside organizations and on internal criticism. OpenAI said there were additional issues beyond what the open letter described, but it did not release investigative material or specify the alleged violations.

How much can outside researchers see?

The researchers’ demands center on preserving the conditions that make external safety review possible: giving outside safety evaluators access to models and spelling out clear procedures for working with them. They also called for maintaining the ability to inspect the process by which an AI arrives at its answers.

OpenAI had previously announced a commitment in principle to supporting independent evaluations by outside organizations, an approach that would let external groups look deeply into a model from training through evaluation and deployment. In its latest statement, the company said it is finalizing agreements with external safety evaluators and will share details within a few weeks. Whether those stated evaluation principles actually translate into contracts and procedures is the next thing to watch.

Why the record of how an AI reaches an answer matters

The trail an AI leaves as it works toward an answer offers clues for spotting safety problems. Researchers examine these records to look for signs of inappropriate behavior. OpenAI has itself described this kind of monitoring as useful, while adding that depending on how a model is trained, that process can become harder to inspect. That is precisely the point the researchers were targeting when they demanded that monitorability be preserved.

For people who use AI, the concern boils down to one thing: how far can outside organizations go in checking the safety of the models they rely on? Both the access guarantees the researchers requested and the external evaluation agreements OpenAI has signaled speak to those conditions for review. The key issue is how the scope of evaluator access and the procedures for collaboration are defined in OpenAI’s forthcoming announcement. As for whether the firings were justified, settling that will require more evidence to weigh the company’s claimed rule violations against the researchers’ account of what they were authorized to do.