An Anthropic researcher resigned Tuesday and said the people building artificial intelligence privately believe the technology could wipe out the human race by the end of the decade. Jacob Coxon, 27, announced his exit in an interview with the Wall Street Journal and followed it with a thread on X laying out why he left. He accused both Anthropic and OpenAI, his two former employers, of failing to act responsibly, and said the companies are racing toward systems that improve themselves while, in his words, “gambling with our lives.”
Coxon is not an outside critic. He is a mathematics graduate who spent three years inside the industry doing pretraining work, which is the stage where an AI model is fed enormous amounts of data and learns patterns from it before anyone fine tunes it for a specific task. He was on the technical staff at OpenAI from 2023 to 2026 and worked on GPT-4o, the model behind much of what people encountered in ChatGPT. He left that job earlier this year and went to Anthropic specifically because the company had a reputation for taking safety seriously. Now he has left that one too.
The thing he says he is afraid of has a name in the industry: recursive self improvement. In plain terms, it means an AI system that can rewrite and upgrade its own code, getting smarter each time without a person in the loop. Every improvement makes the next improvement faster. The concern is not that a machine wakes up angry. It is that a system smart enough to redesign itself could get away from the people supervising it long before anyone realizes what happened. Superintelligence is the term for the end result, a system better than humans at essentially everything rather than just at chess or image recognition.
What gives the resignation its weight is the language Coxon says has become normal inside these buildings. He told the Journal that his colleagues have started describing this period as crunchtime and endgame, terms that suggest the people closest to the work think the decisive stretch is already here. He also gave a timeline far shorter than the one usually debated in public, saying the industry is on track for the most aggressive scenarios and that by the end of next year things could already be out of control.
He was not the only Anthropic researcher talking Tuesday. Evan Hubinger, who leads alignment work at the company, publicly backed him. Alignment is the technical field concerned with making sure an AI system actually pursues what humans intended rather than some warped version of the instruction. Hubinger said he personally puts the odds of AI causing human extinction within the next decade at greater than 10 percent. He said he believes Anthropic is trying its best, but that the company does not yet “have a plan to solve alignment for superintelligence” and is not clearly on track to find one. He was careful to say the models available right now carry relatively low risk, and that his worry is about what comes next.
Samuel Marks, who works on oversight research at the same company and said he was speaking for himself and not for his employer, went further on the internal culture. He wrote that AI developers believe their own technology could cause human extinction or something close to it, possibly within a few years, and that the concern scales with rank. The more senior the employee, the more worried they are. His explanation for why the work continues anyway is money and fear. There are enormous commercial rewards for whoever gets there first, and a widespread belief inside each company that if they slow down, a competitor with looser standards will build the same thing worse.
Anthropic has been publicly calling for the industry to slow development since June. OpenAI released GPT-6 Astra last week and claimed the world had entered the era of artificial general intelligence, pairing that announcement with warnings of its own. On Capitol Hill, Congressman Greg Casar said on September 4 that he and Senator Bernie Sanders have introduced a bill to ban artificial superintelligence outright, arguing that a handful of companies control the data and computing power required and that their own chief executives compare the danger to nuclear weapons.
There are serious people who think the extinction timelines are overblown, and the forecasts coming out of these companies have been wrong before. But there is a version of this story that requires no prediction at all. Research on the American labor market has found entry level employment down nearly 20 percent in the sectors most exposed to AI. That is not a scenario for 2030. That is the job market right now, and it falls hardest on people trying to get their first position, which historically has been the hardest door to get through for young Black workers even when the economy was cooperating.
So there are two conversations happening at once, and only one of them is getting the headlines. Silicon Valley is arguing about whether the machines end the species before the decade closes. Meanwhile the entry level rung of the ladder is disappearing, the people who built the technology are the ones sounding the alarm about it, and nobody with the power to slow it down has shown any sign of doing so.
Coxon did not soften the reason he quit. He said the industry is not acting responsibly. Then he left.
