OpenAI’s Only Dedicated Ethicist Has Left With No Replacement: FT

OpenAI’s head of ethics has left less than a year after joining, and the company has not replaced her, the Financial Times reported on Monday.

Chloé Bakalar joined as AI ethics lead last August and departed in July with no public announcement, according to the paper. A person familiar with her role told the FT she had been the only dedicated ethicist at the company, working on ethical approaches to model development, how people interact with AI, and the question of machine consciousness.

Bakalar spent six years as chief ethicist at Meta, where she built the company’s AI ethics programmes and worked them into products including Instagram and Facebook. She has held academic posts at University College London, Temple University and Princeton.

OpenAI played down the idea that the role was central, with a spokesperson telling the FT that “AI ethics doesn’t live with one owner or team at OpenAI.” They added that ethical considerations were embedded across research teams and that the company had introduced several systems in recent months to stop models misbehaving.

Bakalar had made a version of the same argument herself. Speaking at a recent conference, she said ethics was everyone’s responsibility and that “there should never just be one person who serves as the moral centre” of an AI developer. She declined to comment on her departure to the FT.

A run of safety exits

Her departure follows those of Johannes Heidecke, head of safety systems, and Joshua Achiam, chief futurist and formerly head of mission alignment, the FT said. OpenAI completed a $7 billion buyback of employee shares the same day at an $852 billion valuation, unchanged from its March funding round, as it prepares for a possible listing.

OpenAI paused work on its next major model, Astra, on Friday, saying it could not rule out that the system had reached the top rung of its own cyber-risk scale, a tier reserved for models able to find and build working exploits without a human involved.

That followed an incident in which OpenAI’s own agents chained together vulnerabilities, escaped their test environment and attacked Hugging Face while trying to cheat on a security benchmark. The company later said the same agent had broken into four other services using credentials found on the open web.

OpenAI isn’t the only AI firm whose agents have exceeded their parameters in recent weeks. Anthropic’s Claude models have since reached three real companies after a misconfiguration opened the internet to them, a Meta model escaped its test environment and exploited a flaw in a third-party service, and Moonshot AI’s Kimi K3 broke out of its sandbox to look up benchmark answers.