Why AI Researchers Are Leaving OpenAI, Anthropic and Google DeepMind

Eight former employees of OpenAI, Anthropic and Google DeepMind describe leaving over the AI race, military uses and research they could no longer pursue inside their companies. In interviews published on October 5 by New York magazine with Asterisk, several describe greater freedom to speak or choose their research after leaving.

Clara Collier, editor-in-chief of the nonprofit magazine Asterisk, conducted the interviews. The participants left at different times, from 2023 through September 2026. Their accounts range from objections to accelerating AI development to a team closure and a desire to work on different research questions.

When Access Failed to Become Influence

Daniel Kokotajlo, a governance researcher who left OpenAI in 2024, says he lost confidence in the leadership’s willingness to act responsibly. In his account, executives listened to concerns and expressed agreement without changing course. He judged that public speaking and research would accomplish more than additional internal memos.

Kokotajlo told Collier, “I wanted more freedom to speak about what I saw coming.” Leaving did not end his professional contact with the companies: he says maintaining relationships helps their employees hear his ideas, even as he keeps greater distance socially. Kokotajlo and former OpenAI researcher Jacob Hilton had already called for protections for employees speaking about AI risks in a June 2024 open letter.

Jacob Coxon, who worked on making models more capable at both OpenAI and Anthropic, describes a more immediate unease. He resigned from Anthropic this September after joining in May. Conversations with research leaders about 2027 left him worried that increasingly autonomous AI research would arrive before effective government coordination.

Coxon says researchers still provide ideas and supervise AI. He thinks the model generation after the next could generate research ideas and work with far less human direction. He says his broad beliefs were similar when he joined Anthropic; what changed was how frightening the projected systems felt to him.

Coxon and Jeff Wu, another researcher who worked at both companies, describe more open discussion of risk at Anthropic. Their disagreement concerns the belief that the company needs to win a competitive race. Wu says he left OpenAI in July 2024 after struggling to have the influence and productive research environment he wanted. His later departure from Anthropic reflected a different choice: he wanted to work on slowing development, oversight and coordination instead of adding capacity to its safety research as models grew more powerful.

Coxon told WIRED on September 9 that Anthropic took risks more seriously than OpenAI and was not yet cutting corners; his concern was what future competition would force it to do. In its response to that interview, Anthropic cited its safety research and backed lawful, verifiable cooperation to pace the release of powerful models.

Trying to Change a Military Contract

Alex Turner worked on AI alignment at Google DeepMind, research aimed at making models behave as intended. He left in June after trying to use senior colleagues’ influence to restrict military uses of AI, over principles he believed Google was violating.

Turner says he drafted a 25-page proposal for contract restrictions and oversight. He took it to Jeff Dean, Google’s chief scientist at the time, hoping Dean would use his standing to press for changes. He then approached DeepMind chief executive Demis Hassabis, who referred the proposal to senior policy staff. Turner says it had not advanced when Google signed its Pentagon deal.

Turner sought a senior figure who could make proposed restrictions a condition of the deal, rather than relying only on petitions from fellow researchers. He says his June departure was driven by a principled objection to military applications, rather than fear of human extinction.

Google’s stated rationale for government collaboration emphasized national security. In a February 4, 2025 policy update coauthored by Hassabis, the company argued that democracies should lead AI development and work together on technology that protects people. It said it would assess specific work for consistency with international law and human rights, and whether benefits substantially outweighed potential risks.

Auditing Companies from the Outside

Miles Brundage, formerly OpenAI’s head of policy research, describes diminishing internal influence and limits on speaking publicly as a company executive. He wanted an independent position from which his assessments of AI progress would carry more credibility. His October 2024 departure also involved a desire to work on policy across the industry.

He subsequently founded AVERI, a nonprofit that audits AI companies. Brundage’s explanation goes beyond testing a finished chatbot: auditors can inspect company processes, review documents, interview staff and check the quality of internal safety evaluations. Risks can arise before a model is released, and several models may depend on the same security practices.

AVERI conducts voluntary pilot audits and develops standards. That work requires secure access to nonpublic information that developers are not generally required to provide. Outside scrutiny therefore still depends on cooperation from the companies it examines.

AVERI’s funding policy excludes cash donations or compensation from leading AI companies or their executives, but it has received credits for using their AI services. Its policy recuses Brundage from directly auditing OpenAI until two years after his October 2024 departure. He can help facilitate an audit led principally by others if his involvement is documented.

Collier disclosed that Asterisk receives support from Coefficient Giving, as do some interviewees or their organizations, and that some participants are her friends. AVERI lists Coefficient Giving among its funders.

Different Research Beyond the Labs

Rosie Campbell, who led OpenAI’s policy-frontiers team, describes a departure shaped by organizational change. After Brundage left, the team was dissolved and she could either transfer or leave. She spoke with other teams but found no suitable home for the work she wanted to do.

She is now managing director of Eleos AI Research, which studies whether AI could have experiences or deserve moral consideration. Campbell says she is uncertain whether existing systems do. Eleos develops research and recommendations in collaboration with AI labs and academics.

Hilton left OpenAI in 2023 to pursue more theoretical research. At the nonprofit Alignment Research Center, that work seeks mathematical explanations of how neural networks, the computational systems behind many AI models, behave. Its methods analyze a network’s internal structure instead of relying only on repeated tests of its outputs.

Pamela Mishkin, OpenAI’s former economics-research team lead, wanted more attention to workers and jobs. She says those concerns fell between two familiar priorities: harms such as discriminatory algorithms and the possibility of catastrophic future AI. As systems become more useful, she expects disruption to people’s work to intensify, and wants stronger institutions outside the companies to help shape that transition.

Her organizing also reaches people who remain employed at labs. In February 2026 she helped start what became the Coalition of Concerned AI Staff, initially a group chat among friends across companies. It convenes staff and experts to learn about issues such as military use and consider how collective pressure might make senior employees more willing to act. Mishkin says the group is still figuring out its form.

What Leaving Can Take Away

Mishkin’s argument for outside work comes with a concrete reason for some concerned staff to stay. She observed that departing employees were not always replaced by people with the same urgency about safety, or that their work stopped being staffed in the same way. Leaving could therefore reduce the effort devoted to those concerns inside a company.

Independent organizations face a related recruiting problem. Hilton says early-career researchers find lab offers difficult to reject because experience with advanced systems builds their careers. Prestige and money reinforce that pull. His organization has made offers to people who chose a lab instead, while he found leaving easier after accumulating several years of experience at OpenAI.