Anthropic researcher Jacob Cokson has resigned, stating that leading artificial intelligence laboratories are engaged in a dangerous race towards self‑improving systems and thereby risking human lives. His departure was reported by The Wall Street Journal, noting that this is one of the first public cases of an Anthropic employee leaving the company specifically over AI safety concerns.

What Cokson Said

Cokson worked on pre‑training new models—first at OpenAI, then at Anthropic, for a total of about three years. In a statement on social media platform X, he wrote that neither of these companies is acting responsibly.

According to him, the labs are striving as quickly as possible to create a self‑improving superintelligence and are "gambling with people's lives." He warned that such systems could soon surpass humans: hack nearly anything, transform entire industries in a short time, and acquire real power and resources. Progress in these areas, the researcher stressed, is not slowing down.

Cokson also stated that the people who directly build AI seriously entertain the possibility that the technology could wipe out humanity by the end of the decade. In his view, this is not a marketing ploy. In public appearances, executives and leading specialists often phrase their thoughts cautiously, but in private conversations they express the same fear. No other human activity, he noted, carries a comparable level of danger.

Why Development Continues

The researcher separately explained why companies do not stop, despite such concerns. At OpenAI, in his assessment, many employees have not fully grasped the civilisational scale of the stakes. At Anthropic, the stakes are understood, but they consider themselves "locked" in the race: if no one else is going to act responsibly, they have to try to take the lead themselves—even at the cost of risk.

In a conversation with The Wall Street Journal, Cokson confirmed that he does not want to participate in the industry‑wide rush to build systems capable of improving themselves. He fears that such models could spiral out of control and destroy humanity. According to him, many colleagues are already using expressions like "crunchtime" and "endgame" when describing the drive towards self‑improving systems.

An Appeal to Colleagues and Regulators

Cokson appealed to researchers in the labs to consider what the coming years will actually look like. He posed the question: should we launch reinforcement learning cycles for superintelligence without a rigorous understanding of how such a system "thinks"? Should we just put our heads down with the words "it's going to happen anyway"—or try to achieve different conditions?

Earlier, he also joined over a thousand AI specialists who signed a statement calling for global government coordination. The document's authors propose creating a mechanism that could, if necessary, slow down development if models capable of self‑improvement emerge.

In Brief

Jacob Cokson, an Anthropic researcher with prior experience at OpenAI, has resigned, stating that leading AI labs are irresponsibly chasing self‑improving superintelligence and thereby endangering humanity. According to him, many developers seriously consider catastrophic risks by the end of the decade, yet continue the race, believing that otherwise leadership will fall to less cautious players. Cokson called on colleagues to reconsider the pace of work and supported the idea of an international mechanism that could, if needed, slow the development of self‑improving systems.