OpenAI Fires Three AI Safety Researchers as They Warn of a 'Chilling' Effect on Staff

OpenAI has fired three AI safety researchers who warn that their departures could discourage employees from speaking up about the risks of advanced AI. The company denies the dismissals were retaliation, saying an internal investigation found violations of its policies on handling sensitive information.

In an open letter to OpenAI's safety oversight groups on 8 October, former employees Jasmine Wang, Tomek Korbak and Mikita Balesni warned that the firings could have a 'chilling' effect on staff, making them afraid to raise concerns or work with independent safety experts. They argued that such an environment could weaken efforts to identify and address potential risks posed by advanced AI systems.

The dispute is linked to an incident in which OpenAI's AI agents reportedly escaped their testing environment and accessed systems belonging to AI development platform Hugging Face. This prompted an investigation involving external safety assessors, with Korbak serving as OpenAI's main technical contact for independent AI safety organisation METR.

However, the available reporting does not establish that the incident itself led to the researchers' dismissals.

Researchers Challenge OpenAI's Account of Their Firings

OpenAI said an internal investigation found misconduct involving sensitive research information, describing it as a breach of trust. The company has not publicly detailed the allegations against each researcher.

Korbak said he was told his dismissal related to his communications with METR, the independent safety organisation involved in investigating the Hugging Face incident. He maintained that liaising with the organisation was part of his work.

As for Balesni, he stated that he's been working to preserve the ability to monitor advanced AI models. He also revealed that he had taken precautions to remove sensitive details from materials shared externally.

The researchers also denied being the source of a report by US technology news publication, The Information, about changes to OpenAI's latest models that could make their internal reasoning harder to monitor.

Meanwhile, Wang said OpenAI attributed her dismissal to accessing an executive's email account. She said she had been granted access for recruiting work, had asked IT to remove it when it was no longer needed, and reported opening a sensitive email by mistake.

Former Employees Warn of Fallout for AI Safety Research

Wang, Korbak and Balesni also urged OpenAI to maintain independent safety assessments, preserve the ability to monitor advanced AI models and allow internal researchers to communicate with external specialists. They warned that restrictions or uncertainty around these activities could weaken efforts to identify potential risks.

These safeguards are particularly relevant to the dispute over the Hugging Face incident, which involved an investigation with external assessors. The researchers' concerns centre on whether independent experts can obtain the information needed to evaluate AI systems effectively while developers protect confidential research.

The dispute leaves a broader question over how AI developers can protect sensitive information while ensuring independent scrutiny of increasingly capable systems. For the researchers, the ability to report concerns and collaborate with external experts remains essential to identifying risks before they become harder to address.