Anthropic Researcher Quits, Warns of Racing to Superintelligence

A researcher who worked at both of the world's most closely watched AI labs has resigned with a blunt public warning, and the reaction from inside those labs may be more striking than the departure itself. Jacob Coxon, who spent around three years doing pretraining research at OpenAI and then Anthropic, announced this week that he was leaving. He said neither company was behaving responsibly.

"They are racing straight to self-improving superintelligence and gambling with our lives," Coxon wrote in posts that, by his account, reached more than 100 million people within a day. In a message to colleagues on Anthropic's internal Slack, reported by NBC News, he argued that without more caution and cooperation between the leading developers, a superintelligent system could carry a genuine risk of causing human extinction.

The response from inside

What gives the resignation weight is that it did not draw a flat denial. Evan Hubinger, Anthropic's chief safety researcher, replied on X that Coxon was correct, adding that the company does earnestly believe advanced AI could kill all humans. That is an unusual thing for a senior figure at a leading lab to say out loud, and it captures the strange position these companies occupy. They are building systems they describe, in the same breath, as economically transformative and potentially catastrophic.

Coxon's case is not that the labs are cynical. It is closer to the opposite. He suggests that people who take the risks most seriously are still choosing to push ahead, on the reasoning that if the frontier is going to be crossed, better to be the careful ones who cross it first. Whether that logic holds is the argument the whole field is now having with itself.

A pattern, not a one-off

Departures of this kind have become a recurring feature of the AI safety world, and the site has tracked the debate before, from Anthropic's experiments in having Claude help with its own alignment work to the legislative push by Senators Sanders and Casar to outlaw superintelligent systems outright. The through-line is a disagreement about pace. Nobody in this story doubts the technology is powerful. They disagree about whether the people building it can slow down without simply handing the lead to someone who will not.

For readers trying to calibrate how much of this is warranted alarm and how much is the field's characteristic drama, it is worth separating the claim from the confidence. The risk Coxon describes is real enough that serious researchers plan around it. The timeline, and the odds, remain genuinely unknown. We have looked at the evidence on that question separately, and the honest answer is that certainty is the one thing nobody in this debate can offer.