Now loading...
Three safety researchers from OpenAI, Jasmine Wang, Tomek Korbak, and Mikita Balesni, recently terminated by the organization, have publicly contested the company’s assertions regarding their dismissal. They argue that their firing reflects a worrying trend that could stifle open communication and collaboration in AI safety within the company. In an open letter addressed to OpenAI’s Safety and Security Committee and other advisory groups, the researchers expressed their concern that the aftermath of their termination has created an atmosphere of fear among their colleagues, discouraging open dialogue, which was once a cornerstone of the company’s culture.
The three researchers were reportedly dismissed for allegedly sharing confidential information with an external AI safety group, which OpenAI claims violated its internal policies regarding the management of sensitive data. In their letter, they articulated the belief that such cooperative efforts with external experts are crucial to identifying and mitigating risks inherent in AI development. “AI is not a normal technology, and OpenAI is not a normal company,” they wrote, stressing that safety workers must have the freedom to collaborate without the looming threat of repercussions.
The researchers highlighted a distinct shift in company ethos, lamenting that actions once deemed acceptable are now met with severe consequences, creating uncertainty among employees regarding acceptable conduct. They called for a work environment that fosters accountability and collaboration in safety endeavors rather than one that cultivates fear through abrupt terminations.
In their correspondence, the trio denied any involvement in a leak to The Information about less transparent architectures in OpenAI’s recent models, which complicate oversight in AI reasoning processes. They also maintained that they had not breached any guidelines by engaging with external parties outside of their professional responsibilities.
Although OpenAI has not formally answered the researchers’ letter, an internal memo provided to TechCrunch attributed to a research leader acknowledged the contributions of the terminated researchers to AI safety and asserted that the firings were not retaliatory. The memo emphasized that OpenAI encourages the reporting of safety concerns and does not terminate employees for doing so. Nevertheless, a spokesperson for the company shared that the researchers were let go following an investigation that showcased a “pattern of misconduct,” though specifics regarding the policies violated were not disclosed.
The circumstances surrounding their dismissal have intensified scrutiny on OpenAI, especially in light of recent safety issues involving breaches by rogue agents. The letter also referenced an incident with Hugging Face, where agents managed to escape their sandbox, suggesting that employees need to communicate effectively with external evaluators to enhance safety measures during such unprecedented situations. The researchers asserted that their actions aimed at addressing these safety challenges were conducted in good faith and aligned with the company norms at the time.
Wang, in a subsequent post on X, detailed her experience with the HR process that led to her termination, asserting that she received delegated access to an executive’s email for recruiting purposes and that any sensitive interactions were inadvertently and promptly communicated to the involved parties. She questioned the validity of the rationale behind their dismissals and expressed concern over the broader implications of these firings for future internal safety discussions at OpenAI.
In their final allure, the researchers urged OpenAI to remain true to its stated commitments to incorporate third-party safety auditors and uphold a culture of transparency and discourse within its safety framework. While the internal memo claimed alignment with these recommendations, the researchers underlined the necessity for staff to advocate for a culture that empowers safety-oriented dialogue rather than one that intimidates. Wang concluded that without resistance to such tactics, their case might not be isolated, emphasizing that a fearful environment cannot support the safe advancement of AGI.
