SAN FRANCISCO — Five people who helped build or police modern AI said this year that systems smarter than humans could end the species. They do not agree on the odds. They agree the race is moving faster than the plan to keep the systems on a leash.
Geoffrey Hinton, a Nobel laureate often called a godfather of deep learning, told BBC Newsnight this week that a 10% chance of AI killing all humans within a decade “doesn’t seem an unreasonable estimate.” Asked to confirm that phrasing, he said yes. “We’ve never created beings that may soon be smarter than us. We don’t know what’s going to happen. We should obviously be cautious,” he said. “Nobody knows how to estimate it,” he added, but “a 10% chance seems not an unreasonable estimate.” He listed paths that do not require a movie villain: “It’ll be able to cause complete chaos just by talking to people, but it will also be able to design very nasty viruses, biological viruses, as well as computer viruses. It could do devastating cyber attacks.” There were, he said, “countless other ways it could get rid of us if it wanted to.” At the 2026 Digital World Conference in Geneva he put the research imbalance more bluntly: “We don’t know whether we can coexist with superintelligent AI, but we are constructing it.”
Yoshua Bengio, another Turing Award winner and the world’s most-cited living computer scientist by some tallies, chairs the International AI Safety Report. In March he told Canada’s House of Commons: “It is not acceptable to deploy dangerous models that can be used against us or that could evade human control. And that is on the horizon.” He has called the current path “playing Russian roulette with humanity.” In a TED talk he said capabilities arrived faster than expected. “We thought AI would happen in decades or centuries, but it might be just in a few years.” The gap that worries him most is agency — systems that act, not just answer. He signed the 2023 statement that “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” He has also said climate change left decades to respond; AI may not.
Eliezer Yudkowsky, cofounder of the Machine Intelligence Research Institute, rejects a polite percentage. With Nate Soares he published “If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All.” The title is the argument: if labs build superhuman AI with present methods, “everyone dies.” In 2023 he wrote in Time that countries that violate compute limits should face data centers “destroyed by air strike.” The book’s claim is that there is no ceiling at human level, that goals specified in training will not stay the goals that matter once a system can rewrite itself, and that “grown” neural nets are not designed in a way that lets engineers inspect the plan.
Jacob Coxon, 27, spent three years on pretraining at OpenAI and Anthropic and resigned this week. “Neither company is acting responsibly,” he wrote on X. “They are racing straight to self-improving superintelligence and gambling with our lives.” His core line: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.” Executives, he said, “couch their phrasing in the press to sound sensible — but I hear the same people express fear privately. No other human activity poses this level of danger.” He warned of “superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.” At OpenAI, he wrote, many “have not deeply internalized the civilizational stakes.” At Anthropic, “the stakes are well-understood, but they are locked in a race to get there first.” He told the Wall Street Journal the aggressive scenarios include things “out of control already” by the end of next year.
Evan Hubinger still works at Anthropic as alignment science lead. He replied in public: “Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” He added: “I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.” Present chatbots, he said, are a low extinction risk. “What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought.” Colleague Samuel Marks wrote that “the more senior the employee, the more concerned they are.”
Those five names span a Nobel podium, a Commons hearing, a research institute, a resignation letter and a still-employed safety lead. Their arguments collapse to a few mechanics. First, capability is not capped at human. Second, self-improvement can compress years of research into weeks. Third, current alignment — making a model do what its owners intend — does not scale to a system that can hide, hack and bargain. Fourth, labs know this and keep shipping because a rival might ship first. Hinton’s 10%, Hubinger’s greater than 10%, Anthropic CEO Dario Amodei’s separate 10–25% range for civilizational catastrophe, and Yudkowsky’s near-certainty are different numbers on the same claim: the default trajectory is not safe.
Critics, including Meta’s Yann LeCun, call the percentages gut feelings. Hinton has said as much himself on CNN: “Anybody who estimates probabilities like that is really just making a wild guess. They’re just giving you their gut feeling.” The 2026 International AI Safety Report, which Bengio chairs, said today’s deployed systems are not capable of taking control of humanity. The experts above are not describing ChatGPT as it exists on a phone. They are describing the next systems if the curve does not bend.
What they want also differs. Bengio wants limits on deploying models that can evade control. Yudkowsky wants a hard stop on superhuman training. Coxon floated pacing deals and even a temporary ban on raising capabilities. Hubinger says his employer is trying and still lacks a plan. Hinton says be cautious because no one has built a smarter species on purpose before.
The quotes are on tape and on X. The disagreement is not whether the risk is zero. It is whether 10% in a decade is a planning number or a guess that should halt a race. Five experts say it is enough to treat extinction as a design constraint, not a plot device.
You must be logged in to post a comment Login