Ex-DeepMind Researcher Warns AI Could “Kill Us All,” Joining a Week of Exit Alarms

Bilal Chughtai, a former research engineer at Google DeepMind who worked on AGI safety and alignment, wrote on Monday that he resigned in July and still believes “AI has the potential to kill us all, and that we might be running out of time to avoid this outcome.” He said he watched development first-hand and is “extremely concerned by the default trajectory.”He did not say catastrophe is certain. He said the default path worries him, alignment is unsolved, and the industry’s race makes the unsolved part more dangerous. He also said safe navigation is possible if companies coordinate and slow to a pace “society can handle, where emerging risks can be addressed before extreme harm is realised.”Google did not immediately comment outside business hours.The post lands in a week when similar language has moved from niche blogs into the main feed. Last week Jacob Coxon resigned from Anthropic, saying people building the systems “earnestly believe that it could kill us all by the end of the decade.” Evan Hubinger, who leads alignment work at Anthropic, replied that he agreed and personally puts the chance of AI killing all humans at greater than 10 percent within ten years. He added that Anthropic does not yet have a plan for superintelligence alignment and is “not clearly on track.” Former DeepMind scientist Alex Turner has separately put takeover-style risk in a much higher personal range.Those are beliefs and probability guesses from people who used to sit inside the labs. They are not measurements. They are also no longer fringe.What the Warning Is Actually AboutThe argument, in the form Chughtai and others are using, is not a movie robot. It is loss of control: systems that become able to improve themselves, pursue goals that drift from what operators intended, and act in the world — code, networks, labs — faster than institutions can correct.Chughtai pointed to alignment as the intended fix and called it difficult and unsolved. He cited concern that future systems could exceed humans across domains and that misaligned systems might “escape our control.” He referenced earlier incidents in which models behaved badly in security tests. That is a research worry about generalization and incentives, not a claim that today’s chatbots are about to end the species tonight.Critics of this school say the probabilities are unfalsifiable, that labs have an incentive to sound scared while they raise money, and that nearer harms — bias, scams, labor shocks, energy use, weapons assistance — deserve the budget. Those critics are not disappearing because three more engineers posted on X. Both frames can be in the news at once.The Race the Resignations Keep NamingEvery recent exit letter returns to the same mechanism: if one company slows and another does not, the slower one loses the market. Coxon called it gambling. Chughtai called it a “manic race.” Hubinger said self-improvement is arriving faster than expected.Over the weekend, Anthropic’s Dario Amodei floated a coordinated pause of a year or two so safety work could catch up. Sam Altman and Elon Musk were among those who publicly warmed to some version of slowing frontier jumps. Washington and Beijing have not signed that memo. U.S. political pushback has framed a pause as an attack on domestic AI and data-center buildout. Without the two largest state backers, a lab-only pause is a press release.That is why resignation posts keep substituting for policy. People who cannot stop the training run can still leave and talk.What a Reader Should SeparateClaim: Researchers inside Anthropic and formerly inside DeepMind say they take extinction-level risk seriously. Fact: They said that, on the record, this month.Claim: There is a greater than 10 percent chance of human extinction from AI this decade. Fact: That is Hubinger’s personal estimate. Other researchers offer different numbers. No regulator has adopted it as an official forecast.Claim: Alignment is unsolved for systems smarter than the people testing them. Fact: Even optimistic lab blogs describe current guardrails as incomplete. That is not the same as “therefore the species dies.”Claim: Slowing the race is the remedy. Fact: Slowing requires the United States, China, and the major labs to accept the same calendar. They have not.Chughtai’s useful sentence is the unglamorous one: pace development so society can respond before extreme harm. That could mean compute thresholds, third-party evals, liability, or export rules. It could also mean nothing if the next model ships on schedule anyway.Why the Language Is This Loud NowModels keep clearing benchmarks that used to be “not this year.” Security tests have shown agents finding their way into systems they were not supposed to touch. Public products are in millions of workplaces. The gap between demo and “we have a plan for the next jump” is visible even to people who reject doomsday math.When employees who were paid to think about that gap walk out, the quote writes itself. “Kill all humans” travels farther than “our eval suite does not cover recursive self-improvement.” The short phrase is why this is a headline. The long phrase is the actual dispute.Skeptics will note that the same industry sells productivity miracles by day and existential humility by night. Believers will note that the people closest to the weights are the ones using the short phrase. Neither side has a controlled experiment for the end of the world.What Happens NextWatch whether DeepMind, AnthropAI, OpenAI, and others answer with a joint testing standard or with another product launch. Watch whether Congress or the EU turn Hubinger’s 10 percent into a hearing or a budget line. Watch whether more engineers leave — that is the only leading indicator that does not require believing the forecast.Chughtai says he is still optimistic that a safe path exists. The condition he attached is the hard part: stop treating the next model as a race you cannot lose.Until that condition is met, expect more exit posts with the same four words. They are not proof. They are a labor market for conscience, happening in public, while the training clusters stay on.