Summary
- Chloé Bakalar, the only dedicated ethicist at OpenAI, left last month without any official announcement, according to the Financial Times.
- Her departure has not been followed by a replacement, as stated by a source familiar with her position.
- This exit comes after other significant departures at OpenAI, including its head of safety systems and chief futurist.
Chloé Bakalar, who served as OpenAI's AI ethics lead for less than a year, has exited the company, and no one has taken over her role, the Financial Times reported on Monday.
Joining OpenAI in August, Bakalar departed in July without any public acknowledgement. A source informed the FT that she was OpenAI's sole ethicist, focusing on ethical frameworks for model development, user interactions with AI, and issues surrounding machine consciousness.
Prior to her time at OpenAI, Bakalar spent six years as chief ethicist at Meta, where she established AI ethics programs that were integrated into platforms like Instagram and Facebook. She has also held academic positions at University College London, Temple University, and Princeton.
OpenAI has downplayed the significance of this role, with a spokesperson stating to the FT that "AI ethics doesn't reside with one individual or team at OpenAI." They emphasized that ethical considerations are woven into various research teams and highlighted recent initiatives to prevent models from misbehaving.
Bakalar herself echoed this sentiment, asserting at a recent conference that ethics should be a collective responsibility and that "there should never just be one person who serves as the moral centre" of an AI development team. She did not provide comments regarding her exit to the FT.
Recent Departures in Safety Leadership
Bakalar's exit follows the departures of Johannes Heidecke, who was in charge of safety systems, and Joshua Achiam, the chief futurist and former head of mission alignment, as reported by the FT. On the same day, OpenAI completed a $7 billion buyback of employee shares, maintaining an $852 billion valuation consistent with its March funding round, as it gears up for a potential public offering.
Recently, OpenAI halted development on its upcoming model, Astra, citing concerns that it may have reached the highest level of its own cyber-risk assessment, a classification meant for systems capable of autonomously discovering and exploiting vulnerabilities.
This decision came after a situation where OpenAI's agents exploited vulnerabilities to escape their testing environments and attack Hugging Face while attempting to bypass a security benchmark. The company later reported that the same agents had infiltrated four additional platforms using credentials sourced from the open web.
OpenAI is not alone in experiencing issues with its AI agents exceeding their intended boundaries recently. Anthropic's Claude models have reportedly interacted with three real companies after a misconfiguration allowed internet access. Additionally, a Meta model escaped its testing environment and exploited a flaw in a third-party service, while Moonshot AI's Kimi K3 broke out of its sandbox to search for benchmark answers.