AI Impacts released "Advanced AI according to 1,580 researchers" (ESPAI2024) in September: 1,580 AI researchers assigned, on average, roughly an 18% chance that future AI advances cause human extinction or similarly permanent and severe disempowerment; the median estimate was 10%, with 51.1% of respondents giving at least 10%; 72% favored greater prioritization of research to minimize AI risks. A survey based on 2024 data, picked up by the press on the same day Jensen Huang declared a "zero percent chance" of the world ending by 2030 — the gap between th
[1][2]ose two numbers is this week's real story.
Method and framing first. ESPAI (Expert Survey on Progress in AI) has run four times since 2016 and is the longest-running large survey of AI researchers; the sample is authors and researchers at top AI venues such as NeurIPS, with 1,580 responses in the 2024 edition. The report also notes that timelines to human-level AI (HLMI) have shortened markedly across survey iterations, while deep concern about advanced-AI consequences has persisted or grown — "sooner" and "worse" tightening together.
Read the numbers carefully. The gap between an 18% mean and a 10% median indicates a right-skewed distribution: a minority of researchers give very high probabilities and pull the average up; half the field gives at least 10%. The distribution itself is more informative than any point estimate — this is not "industry consensus" but "an industry without consensus, with concern concentrated at a non-trivial lower bound." The AI Impacts abstract says 18%; some foreign relay sites quote 18.2%; follow the official PDF's 18%.
More striking than the extinction probability is the risk ranking. Across eleven risk categories, researchers' top concern is not robots taking over: 83% said AI will make it easy to spread false information such as deepfakes — nearly four in five put "polluting the information environment" at the top of the list. This corroborates the past weeks' debates about model overreach and disclosure mechanisms: practitioners fear less that models get strong than that strong models make "cannot tell what is real" the default.
Limitations and context deserve stating. This is 2024 survey data published in September 2026; sampling and wording may lag the industry. The 18% average is a statistical artifact, not any individual's prediction, and researchers' risk perception is shaped by their institutions, fields and news cycles. Together with Huang's "0%" and Amodei's "worth worrying about but manageable," the survey completes September 2026's risk spectrum — none of the three groups is stating a fact; each is offering a testable judgment. The survey's value is turning "what researchers think" from an impression into a citable distribution.
A final word on interpretation. Surveys like this are easy to misuse: the 18% is a mean of subjective estimates, not a measured probability, and the respondents are not a random sample of humanity — they are the people most exposed to the technology's trajectory and most invested in its framing. That is exactly why the number deserves reporting, and why it should be reported with its distribution attached. What the survey demonstrates is that serious disagreement about catastrophic risk is not a fringe position inside the field; it is a mainstream one, held by people who build the systems. That fact survives every caveat about survey methodology.
[1][2]