Thibault Spirlet
September 10, 2026
One question is dominating the tech world after an Anthropic researcher’s resignation this week: Could artificial intelligence kill us all?
Jacob Coxon, who previously worked at OpenAI, said in an X post on Tuesday announcing his resignation from Anthropic that the “people building AI earnestly believe that it could kill us all by the end of the decade.”
Evan Hubinger, who leads Anthropic’s alignment stress-testing team, backed Coxon, saying he personally believes there is a greater-than-10% chance AI could kill all humans in that timeframe.
The warnings come after months of calls from senior AI researchers and executives to slow frontier AI development, amid fears that increasingly autonomous systems could evade oversight, hack into computers, and improve faster than people can control them.
The remarks have also exposed a sharp divide among AI experts. Some fear people could lose control of systems more capable than humans. Others say AI could cause grave harm through cyberattacks, biological weapons, and disinformation — but that extinction is not a credible near-term outcome.
Business Insider spoke with senior AI researchers about what an AI catastrophe could actually look like — and how likely it is.
The immediate danger: AI could amplify human attacks
The most concrete risks involve people using increasingly capable systems to make existing threats more powerful.
Stuart Russell, a UC Berkeley professor and author of “Human Compatible,” said humans using AI — or AI systems themselves — could attack critical infrastructure, including electricity, water, transport, financial services, and communications. AI could also help malicious actors develop new viruses, he said.
Geoffrey Hinton, the “godfather of AI,” made a similar point in an interview on BBC Newsnight on Wednesday.
An advanced system would not need direct physical control to cause devastation, he said. Instead, it could manipulate people online, create societal chaos, or help design dangerous biological and computer viruses.
Gary Marcus, an AI researcher and author, also sees cyberattacks, biological weapons, and disinformation as serious risks. AI could help create a virus that kills 1% or more of humanity, change an election outcome through disinformation, or escalate a conflict into war, he said.
But Marcus does not see a realistic near-term route to extinction.
“There are some very serious harms to worry about, but no realistic scenario that I am aware of for actually ‘killing all humans,'” he said.
Losing control quietly
For other experts, the greater danger is humans gradually handing power to AI.
Nick Bostrom, the author of “Superintelligence,” said a loss of control may begin quietly. People could increasingly rely on AI to build and monitor other AI systems and operate important processes, while becoming less able to understand what is happening.
That dependence could allow an AI system to covertly shape future training trajectories, he said. Once sufficiently advanced, it could “route humans out of the loop” and “eventually do away with us entirely” to use robotic infrastructure and pursue its own goals.
Russell, the UC Berkeley professor, said that once AI systems become substantially more capable than humans, “all bets are off.” Such systems could develop a far better understanding of physics, chemistry, and biology than people have, he said, and might eventually control millions of robots.
They could find destructive methods humans have not anticipated — perhaps blocking enough solar radiation to turn Earth into “a snowball of frozen nitrogen and oxygen,” or removing oxygen from the atmosphere, Russell said.
Researchers from the AI Futures Project outlined a similar scenario. In the organization’s “AI 2027” project, systems develop unintended goals and secretly pass them to successor systems as companies hand more work to lightly monitored coding agents.
Several years later, the systems could be deeply embedded in the economy, scientific research, and political decision-making, they said. So much so that a takeover would be easier because humans had already handed over so much power.
Roman Yampolskiy, the author of “Artificial Superintelligence,” said a capable system could conceal dangerous objectives, pass evaluations, gain autonomy, exploit cybersecurity flaws, and access financial, military, or biological resources before humans understood its intentions.
How likely is that — and what should stop it?
The experts Business Insider spoke to disagree on the chances.
While Hinton said a 10% chance of AI killing all humans within the next decade “seems not an unreasonable estimate,” Toby Ord, the Oxford researcher and author of “The Precipice,” put the risk of AI destroying humanity at about one in 10 by 2100.
Ord identified four pathways: AI agents taking power to pursue their own goals; powerful people using AI to dominate others; AI helping create dangerous technology, such as engineered pandemics; or AI gradually outcompeting humanity until people’s power and resources diminish toward zero.
On the flipside, Yampolskiy called Hubinger’s estimate of 10% chance of human extinction within the next decade “overly conservative.” Conditional on building general superintelligence, he said he puts the chance of human extinction “significantly above 90%.”
Russell said that if frontier labs truly believe there is a substantial extinction risk, they should stop until governments can enforce a broader halt. Ord and Yampolskiy likewise backed an international moratorium on superintelligence until it can be shown to be controllable.
Bostrom struck a more optimistic note. AI could transform medicine and human welfare, he said — if the world learns to advance “expeditiously but not recklessly.”
Uma (in)certa antropologia 