“Advanced AI has more than a 10% chance of killing all humans within the next decade.”

• A non-trivial minority of AI researchers assign double-digit extinction probabilities: in the 2023 AI Impacts survey of 2,778 published AI authors, between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction (the median estimate was 5%).
• Geoffrey Hinton, a Turing Award winner, publicly raised his estimate to a 10-20% chance that AI wipes out humanity, though he framed this over the next 30 years rather than 10.
• Hundreds of senior AI scientists and executives (including Hinton, Bengio, Altman, Hassabis) signed the Center for AI Safety statement declaring that 'mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war' — implying the risk is considered material, not negligible.
• Governments have treated the risk as serious enough for formal action: the 2023 Bletchley Declaration, signed by 28 countries and the EU, explicitly cites potential 'catastrophic' harm from frontier AI, and the International AI Safety Report (chaired by Yoshua Bengio) documents loss-of-control risk as an open research concern.
• Timelines for very capable systems have compressed: median expert forecasts for high-level machine intelligence shifted dramatically earlier between the 2022 and 2023 surveys, which mechanically raises near-term risk estimates for those who think capability and catastrophe are linked.

• The claim is an unfalsifiable forecast, not a verified fact; no empirical method exists to establish a specific probability of a never-observed event, and the stated figure is presented as if it were measured.
• The central tendency of expert opinion is far below 10%: the AI Impacts 2023 survey median for human extinction or similarly permanent severe disempowerment was about 5%, and that was for an open-ended/long-run horizon, not ten years.
• The Forecasting Research Institute's Existential Risk Persuasion Tournament found superforecasters put AI-caused extinction by 2100 at roughly 0.4%, and domain experts at roughly 3% — both far under 10% even over a 75-year horizon, let alone a decade.
• The decade framing is far stronger than almost any cited authority endorses. Hinton's 10-20% figure covers 30 years; most published high estimates (e.g., Bengio's ~20%) are unbounded or long-horizon. Compressing these to 10 years substantially overstates them.
• 'Killing all humans' is a much narrower and more extreme outcome than the survey wording many high estimates come from, which typically bundles extinction with 'severe disempowerment' or 'extremely bad outcomes' — so quoted percentages are not directly transferable to literal total extinction.
• No current AI system has demonstrated the autonomous capabilities (self-replication at scale, independent resource acquisition, control of physical infrastructure) that proposed extinction pathways require; the International AI Safety Report describes these as speculative and contested rather than established.
• Expert elicitation on extreme low-probability events is known to be poorly calibrated and subject to selection effects — researchers who worry about AI risk are more likely to respond to AI-risk surveys, inflating reported medians.

62/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Grace et al., 'Thousands of AI Authors on the Future of AI' (AI Impacts / arXiv)journal · 80 · Center for AI Safetyother · 65 · Forecasting Research Institute — Existential Risk Persuasion Tournamentother · 78 · The Guardiannews · 82 · UK Government (Department for Science, Innovation and Technology)gov · 90 · UK Government / International AI Safety Report (chair: Yoshua Bengio)gov · 90 · Metaculusother · 55

• Some highly relevant experts regard extinction-scale AI risk as serious rather than fringe: the Center for AI Safety statement says mitigating “risk of extinction from AI” should be a global priority, and it was signed by leading AI figures including Geoffrey Hinton and Yoshua Bengio. This supports plausibility of catastrophic-risk concern, but it does not establish a >10% probability or a one-decade timeline.
• A peer-reviewed Science article, “Managing extreme AI risks amid rapid progress,” argues that increasingly capable general-purpose AI could create extreme risks through misuse, autonomous replication, cyber/bio capabilities, deception, and loss of control. This supports the idea that advanced AI could create existential danger, though it does not quantify the chance as >10% within ten years.
• The UK government’s International Scientific Report on the Safety of Advanced AI concludes that advanced AI could pose severe risks, including malicious use and potential loss of control, and emphasizes uncertainty and rapid capability growth. This supports taking the claim seriously as a risk scenario.
• A large survey of AI researchers found non-trivial concern about catastrophic outcomes from advanced AI: the median respondent gave a 5% probability that advanced AI causes human extinction or similarly severe and permanent disempowerment. The existence of many expert estimates in this range supports the broad plausibility of serious AI-extinction concern, though the median is below the claim’s >10% threshold and the outcome is not exactly “all humans killed within a decade.”

• The claim is much stronger than the best cited quantitative evidence: the large AI-researcher survey’s median estimate was about 5% for extinction or similarly severe disempowerment, not >10%, and it was not specifically limited to the next decade.
• The UK government scientific report emphasizes deep uncertainty and disagreement among experts about both timelines and the likelihood of severe AI outcomes. That undercuts presenting “more than 10% within the next decade” as an established factual probability.
• Current AI systems, while rapidly improving, still have documented limitations in reliability, planning, robustness, and autonomous real-world execution. Stanford’s AI Index notes that AI performance remains uneven and that benchmarks can overstate real-world capability, which weakens the case for confident near-term human extinction.
• “Killing all humans” is an especially strong endpoint. Most serious AI-risk sources discuss extinction, severe disempowerment, catastrophic misuse, or loss of control, but public evidence does not demonstrate a mechanism that makes total human death within ten years more likely than not, or even clearly above 10%.
• The claim depends on a long chain of uncertain events: development of sufficiently capable AI within ten years, deployment with inadequate safeguards, emergence or use of globally lethal capabilities, failure of monitoring/governance, and inability of humans to intervene. Evidence supports concern about parts of this chain, but not a confident >10% probability for the full chain as written.

64/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Center for AI Safetyother · 78 · Sciencejournal · 92 · UK Department for Science, Innovation and Technologygov · 90 · arXiv / AI Impacts researcher surveyother · 76 · Reutersnews · 88 · Stanford Institute for Human-Centered AIedu · 86 · National Institute of Standards and Technologygov · 90

• A prominent subset of AI safety researchers and tech leaders (including Geoffrey Hinton, Yoshua Bengio, and Sam Altman) have publicly warned that advanced AI could pose an existential threat to humanity.
• In May 2023, hundreds of AI scientists and executives signed a statement by the Center for AI Safety asserting that mitigating the risk of extinction from AI should be a global priority, on par with pandemics and nuclear war.
• Surveys of AI researchers (such as the 2023 survey by Katja Grace et al.) show that a significant portion of experts assign a non-trivial probability (e.g., 5% to 10%) to extremely bad outcomes, including human extinction, caused by advanced AI over the long term.

• The claim frames a highly speculative future prediction (>10% chance) and an aggressive timeline (within the next decade) as an objective, proven fact, which is scientifically impossible for unobserved future events.
• The 10-year timeline contradicts the median consensus of AI experts. Most surveys that show a 5-10% estimated risk of extinction do so over an indefinite, long-term horizon, with most researchers predicting that Artificial General Intelligence (AGI) itself may not even be achieved within a decade.
• Many prominent AI researchers, such as Meta's Chief AI Scientist Yann LeCun and Coursera founder Andrew Ng, strongly reject existential risk narratives. They argue these claims lack empirical evidence, rely on science-fiction assumptions, and distract from immediate, tangible AI harms like algorithmic bias and misinformation.

95/100 — the lab’s own confidence in the score above, not a second reading of the claim.

The New York Timesnews · 85 · arXiv (Cornell University)edu · 80

• A large survey of 2,778 AI researchers found that half assigned at least a 5% probability to AI causing human extinction or comparably permanent and severe disempowerment; this indicates substantial expert concern, though it does not establish a greater-than-10% chance within the next decade.[2]
• The same survey series reported a median 10% estimate for extinction or similarly severe disempowerment caused by human inability to control future advanced AI, but the estimate was not specifically limited to the next ten years.[4]
• A later AI Impacts survey reported a median 10% estimate for future AI causing extinction or similarly permanent and severe disempowerment, with most respondents assigning at least 10%; this supports the claim's general direction but not its specific ten-year timeframe.[10]
• AI systems could plausibly create severe risks through misuse, loss of control, or rapid capability gains, but the cited survey evidence is based on expert judgments rather than calibrated historical frequencies.[2][4]

• The strongest available survey evidence cited here concerns AI risk over an unspecified future horizon or through 2100, not specifically the period from now through the next decade.[4][10]
• The 2023 expert-survey median estimate for AI-driven extinction by 2100 was 5%, while the median estimate for extinction or severe disempowerment from inability to control future AI was 10%; neither supports a greater-than-10% probability within the next ten years.[4]
• Survey respondents were estimating highly uncertain future events, and the estimates are not empirical measurements with demonstrated calibration; presenting them as a precise ten-year probability overstates what the evidence can establish.[9]
• Available evidence does not demonstrate that advanced AI capable of causing human extinction will be developed, deployed, or become uncontrollable within the next decade; the search results provide no reliable government or peer-reviewed forecast establishing that timeline.

86/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Scientific Americannews · 91 · IEEE Spectrumnews · 90 · AI Impactsother · 78 · The Remindernews · 55

• Some leading AI researchers (including Turing Award recipients) have publicly assigned double-digit personal probabilities to severe or extinction-level outcomes from advanced AI over coming decades and have left major labs to warn about loss of control.
• Hundreds of AI scientists and industry figures signed a 2023 statement treating extinction from AI as a risk that should be prioritized alongside pandemics and nuclear war.
• Capability jumps in foundation models and reasoning systems have shortened some experts' AGI timelines into the 2020s–2030s, and the alignment/control problem remains unsolved.
• Theoretical arguments (instrumental convergence, inner misalignment, recursive self-improvement) describe pathways by which a misaligned superintelligent system could seek resources or self-preservation in ways fatal to humans.
• Governments at the 2023 Bletchley Park AI Safety Summit formally acknowledged the potential for catastrophic harm from frontier AI.

• No existing system has autonomous long-horizon agency, reliable physical-world actuation, or the capacity to kill 8 billion people; current models are tools, not independent actors.
• Median estimates in expert surveys of AI researchers for AI-caused extinction or equivalent permanent disempowerment are typically around 5% this century, not >10% inside 10 years.
• Extinction in a decade would require near-term superintelligence plus scalable control of energy, robotics, biotech or equivalent, plus successful covert or unstoppable takeover—barriers that remain enormous.
• Many prominent researchers (including Meta's chief AI scientist) describe near-term human-extinction scenarios as extremely implausible and a distraction from measurable harms.
• Superforecaster aggregates and public prediction platforms generally put much lower probability on AI causing human extinction by the early 2030s.
• Detection, shutdown, regulation, competing labs, and military/state response would almost certainly intervene before a takeover reached global extinction scale.
• Short-timeline total-extinction claims have a long record of over-prediction in AI and do not match demonstrated scaling or embodiment progress.

68/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Center for AI Safetyother · 68 · Future of Life Instituteother · 65 · UK Governmentgov · 88 · The New York Timesnews · 90