Open Letter from Former OpenAI Researchers
Three safety researchers recently dismissed by OpenAI have issued a public letter disputing the company’s account of their firings. They expressed concerns about potential repercussions on open discussions about AI risks among colleagues.
Tomek Korbak, Jasmine Wang, and Mikita Balesni addressed their letter to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. They warned that communications surrounding their dismissals might deter employees from speaking candidly on safety issues.
The trio argued that safety researchers must challenge decisions and collaborate with external experts. They emphasized the unique nature of AI technology and OpenAI’s role in its development.
OpenAI’s Response to the Dismissals
OpenAI has denied claims that the dismissals were linked to safety concerns. The company stated that a thorough investigation revealed policy violations regarding sensitive information handling. OpenAI maintained that those decisions were unrelated to researchers raising safety concerns.
In a public statement, OpenAI reiterated its dedication to fostering research discussions and claimed these conversations are crucial to informed decision-making. The company asserted that recent actions weren’t motivated by the researchers’ advocacy for safety.
Claims and Counterclaims
The former employees have disputed the reasons given for their terminations. Korbak, Wang, and Balesni maintain they were not responsible for leaking information on AI models or improperly engaging with external organizations beyond their job requirements.
Korbak defended his communication with Model Evaluation and Threat Research, an organization assessing AI systems. He expressed concern about monitoring AI agents, suggesting his concerns led to his dismissal.
Balesni questioned OpenAI’s rationale, believing the firings prioritized corporate interests over safety measures. He stated that his interactions with third-party safety organizations adhered to company norms.
Wang contested OpenAI’s reasoning, explaining her involvement in accessing an executive’s email was authorized for recruitment purposes. She claimed efforts were made to revoke access as necessary.
Recommendations for OpenAI and AI Industry
The researchers proposed recommendations for OpenAI and similar AI companies. They emphasized the need for independent safety evaluations and maintaining collaboration with outside organizations. They also advocated for protection measures to monitor advanced AI model reasoning.
Balesni expressed concern about possible repercussions on OpenAI’s relationship with external evaluators following the firings.
OpenAI’s Final Statement
OpenAI confirmed plans to contract external safety assessors, emphasizing continued collaboration with third-party organizations. The company also reiterated the importance of model monitorability, noting alignment with key points in the former employees’ letter.
OpenAI concluded with appreciation for the researchers’ contributions while dismissing claims that their dismissal was due to raising safety concerns.
Implications for AI Development
The disagreement underscores challenges faced by AI developers in balancing proprietary information protection and independent safety oversight. Independent evaluations increasingly require collaboration with researchers within the developing companies.
Korbak, Wang, and Balesni warned of reduced effectiveness in safety evaluations if employees fear repercussions for external communication. Their letter emphasized the crucial role of internal researchers in safeguarding against AI-related issues.
Leave a Reply