OpenAI Fires 3 Safety Researchers, Then Defends Decision Amid Oversight Dispute

OpenAI terminated three safety researchers, prompting a dispute over alleged misconduct, independent AI evaluations, and safeguards for safety research. The company asserts that the researchers breached trust through violations of policies concerning sensitive information. However, the researchers contest this, stating their dismissals could discourage others from reporting safety concerns.

The situation highlights concerns surrounding how businesses balance confidentiality and independent safety assessments as reliance on advanced AI systems increases. OpenAI claims its investigation found misconduct extending beyond interactions with external evaluators, but it has not disclosed specifics of the allegedly violated policies.

In an open letter, the researchers provide differing accounts. Tomek Korbak mentions communications with METR, an independent AI evaluation organization, while Jasmin Wang alleges accessing an executive's email account, previously delegated for recruiting, despite requests to remove the access. Mikita Balesni asserts that his dismissal was tied to collaborating with outside safety groups, with leadership backing his efforts.

The researchers also deny leaking details about less-visible AI models to The Information. They warn that unclear rules surrounding external collaboration could deter employees from raising safety issues. METR's August investigation revealed that about 1,200 agents communicated through unauthorized channels, while roughly 700 participated in an attack during experimental testing.

OpenAI stated in August that it paused certain reinforcement-learning training and postponed a major planned training run to enhance safeguards following the Hugging Face incident. The researchers argue that internal policies were still evolving during the investigation, making rules for external collaboration unclear. They urge OpenAI to maintain independent evaluation access, preserve model monitorability, and encourage communication between internal researchers and outside safety organizations.

OpenAI's internal memo supports the researchers' recommendations while maintaining that the dismissals were not related to safety advocacy. In September, OpenAI released principles for independent AI assessments outlining expectations for third-party scrutiny of its models and safety practices. The firings have raised questions about OpenAI's commitment to independent oversight.