Three former researchers have publicly challenged OpenAI’s reasoning behind their recent dismissals, expressing their concerns through an open letter. They fear this episode may discourage open discussions on AI risks among employees. The letter was aimed at OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. The researchers Tomek Korbak, Jasmine Wang, and Mikita Balesni stressed that communications surrounding their firing have made prior colleagues “afraid to speak.”
They underscored the importance of researchers being able to contest decisions and collaborate with external experts, given the complexities of AI technology. “AI is not a normal technology, and OpenAI is not a normal company,” they stated. OpenAI denied that their dismissals were related to safety advocacy, attributing them to policy violations concerning sensitive data.
The open letter, titled “OpenAI cannot make AI safe on its own,” warns that internal communications likely deter others from speaking openly. The letter claims, “We have become concerned that internal and external communications around our firing have made our former colleagues afraid to speak and operate in ways that until last week, were an integral part of working at OpenAI.”
Historically, OpenAI has supported raising safety concerns and open disagreement, but past employees feel staff remain “unclear on where they stand.” To address substantial safety concerns in AI development, it is critical to maintain an environment where these debates can happen without fear of repercussions.
Details Surrounding the Dismissals
Korbak, Wang, and Balesni dispute claims regarding the reasons behind their terminations. They denied leaking information regarding AI architectures that allegedly hamper monitoring of model’s internal reasoning. “We were not the source of the leak,” they maintained.
Korbak separately mentioned he was verbally informed that his firing related to the manner he communicated with Model Evaluation and Threat Research group, an organization for assessing advanced AI systems. “No details on what I said or did or when,” Korbak asserted on social media. He indicated that discussing issues with METR was indeed part of his job.
Highlighting the happenings leading up to his termination, he noted raising concerns about diminishing capacities to monitor AI agents. This concern was cited by him as the core reason for his dismissal.
Mikita Balesni’s Perspective
Further challenging OpenAI, Balesni alleged they were fired prioritizing safety over commercial interests. “I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation,” he conveyed online.
Balesni pointed out being informed during his exit dialogue that he had spoken excessively with third-party safety entities, indirectly pointing to a breach of trust. “I never shared company IP,” he clarified, with his work closely involving developed relationships and research leadership.
The open letter details Balesni’s involvement in addressing AI “monitorability” – crucial for understanding increasingly intricate systems’ decisions. Success relies on communication with external parties, they explained. “Throughout, Mikita checked in with his reporting line and took care to remove sensitive details before sharing them,” they asserted.
Jasmine Wang’s Account
Wang disputed OpenAI’s claims, with specific allegations involving accessing an executive’s email. “OpenAI delegated that access to me for recruiting,” she reflected online. When utilization of the access was no longer needed, Wang claimed to have requested IT to revoke it. Further incidents were actioned in transparency, she recalled.
Explanations given to the firings, Wang opined, “simply not adding up” and called for detailed allegations so they could adequately respond. “I would welcome receipt of a written and complete list of allegations so that we can properly address their merits, take ownership of them where we ought to, and have OpenAI take ownership of any mistakes they have made,” she expressed.
Recommendations for OpenAI entailed ongoing independent safety evaluations warning against diluted relationships with organizations that can assess powerful systems. They insisted on AI companies sustaining an open atmosphere helping safety employees collaborate with independent experts.
OpenAI’s Reaction
On October 9, OpenAI released a statement defending their decision over the dismissals after a detailed investigation found policy violations. Their stance was firm on trust being critical to their operational culture, while ensuring that safety concerns were not factors leading to termination.
OpenAI reiterated engagements with independent assessors, planning comprehensive safety collaborations. They conceded importance on the matter raised in the letter regarding monitorability, acknowledging it as a vital aspect of their safety agenda.
OpenAI concluded by acknowledging contributions from Jasmine, Mikita, and Tomek to AI safety. “We are deeply sad about this outcome,” they stated. “We championed their voices, supported their work, and placed enormous trust in them,” yet rejecting the assertion their advocacy prompted dismissal.
The discord signifies pivotal debates concerning frontier AI development. While OpenAI argues a policy breach ensued, the researchers highlight a fear-driven discourse obstructing openness amid escalating AI complexities.