The AI Impacts survey published in September 2026 is the freshest large snapshot of what AI researchers expect from their field. It gathered 1,580 valid responses from authors at six leading AI venues. Its headline is stark. On average, respondents gave an 18% chance that future AI advances cause human extinction or a similarly permanent and severe disempowerment of humanity. The median answer was 10%, double the previous survey’s median.
The timing needs care. AI Impacts fielded the questionnaire between 9 and 24 December 2024. The paper, “Advanced AI according to 1,580 researchers: uncertain, unsafe, and sooner than we thought”, appeared only in September 2026. So the data describe the field as it stood almost two years ago. This article reads the primary paper closely. It sets the numbers against earlier waves and weighs the caveats. For the wider policy backdrop, see Insider Release’s overview of global catastrophic risks in 2026.
Who Ran the Survey and Who Answered
The study is the fourth Expert Survey on Progress in AI, or ESPAI. AI Impacts has run the series since 2016. Katja Grace leads a team of nine authors. Funders named in the paper include Coefficient Giving, Jaan Tallinn and the Future of Life Institute.
Recruitment followed the 2023 method. The team wrote to researchers who had published in 2023 at NeurIPS, ICML, ICLR, AAAI, JMLR or IJCAI. It emailed 19,874 addresses and received 2,052 complete or partial responses. That is a 10% response rate. Respondents could choose a $50 gift card, cash or a charitable donation.
A list-building error then surfaced. Many invitees sat outside the target group, mostly workshop authors. The authors excluded them, which left 1,580 respondents for the main analysis. Of those, 1,502 reached the final question.
Earlier waves give scale. In 2016 the survey drew 352 respondents, and the 2022 wave drew 738. The 2023 edition, published as “Thousands of AI Authors on the Future of AI”, reached 2,778. The new sample is smaller than 2023, yet still large by academic standards.
Each participant saw only a random subset of questions. Individual items therefore rest on far fewer answers than 1,580. The human-level AI question, for instance, drew roughly 1,000 responses across its two framings. Most task forecasts rest on 111 to 167 answers each.
Human-Level AI Now Forecast for 2042
The survey defines high-level machine intelligence, or HLMI, strictly. It arrives “when unaided machines can accomplish every task better and more cheaply than human workers.” Respondents were told to think about feasibility, not adoption.
In the mean aggregate forecast, HLMI reaches a 50% chance in 2042. In the 2023 wave, that date was 2047. That is a five-year drop in roughly one calendar year. The same forecast also gives HLMI a 10% chance by 2027, unchanged from 2023.
Over the longer run, the trend is sharper still. The 50% date stood at 2061 in 2016 and near 2060 in 2022. It then fell to 2047 and now 2042. Over eight years, the horizon shrank from 45 years out to 18. The authors calculate that forecasts moved about 3.4 years closer for every year that passed.
Uncertainty remains wide. The same aggregate still leaves substantial probability that HLMI has not arrived in a hundred years. Individual answers range from imminent to never. The 2042 figure is an average of fitted curves, not a consensus date.
Full Automation of Labour and the Framing Gap
A second question asks about full automation of labour, or FAOL. That point arrives when every occupation is fully automatable. Here the mean forecast reaches 50% only in 2098. It hits 10% in 2035.
The 2098 date is still a large shift. The previous wave gave 2115 on current code, and the 2022 analysis gave 2158. That amounts to 17 years closer in one wave and 43 in the wave before.
Why do HLMI and FAOL differ by decades? Logically, a machine that beats humans at every task should also make every job automatable. The authors admit the gap is unexplained. They note the questions differ in wording and structure. The FAOL block first asks about four specific occupations before the final forecast. This time the gap was 56 years, the smallest recorded.
A related framing effect runs through every timeline item. Half of respondents gave probabilities for fixed years. The other half named years for fixed probabilities. Across all four editions, the fixed-years framing produced generally later forecasts. The paper combines both, which is why readers should treat single dates as soft.
Task Milestones: 34 of 39 Within a Decade
The survey also asked about 39 narrow tasks. Each participant received four of them at random. Of the 39, 34 reached a roughly even chance of feasibility by 2034.
That group spans very different skills. It includes building a payment-processing site, writing fiction worthy of the New York Times best-seller list and producing US Top 40 songs. Folding laundry, fine-tuning open-source language models and replicating machine-learning research also appear.
Five tasks landed later, all in maths, science or engineering. Discovering physics equations from simulation came in at 2035. Conducting ML research and writing a conference paper followed at 2036. Proving publishable maths theorems reached 2039 and installing a new home’s electrical wiring 2040. Solving a problem like a Millennium Prize Problem without human input sat at 2053.
Most task forecasts barely moved. Thirty-three shifted by less than two years in either direction. The two big movers were theorem proving and autonomous ML research. Both came roughly seven years earlier.
The authors add a subtle warning. Some tasks may already be feasible, yet no task drew a majority of “zero years” answers. Participants may assume a listed task is still unsolved. That assumption can quietly shift what each item measures over time.
The 18% Extinction-or-Disempowerment Figure
The most quoted result comes from three related questions. The core version concerns “future AI advances.” It asks how likely they are to cause “human extinction or similarly permanent and severe disempowerment of the human species.” A second version blames “human inability to control” advanced AI. A third limits the window to the next 100 years.
Pooled across all three versions, 1,489 people answered. The mean was 18.2% and the median 10%. Just over half, 51.1%, gave at least a 10% chance. About a third gave 20% or more, and 26.5% gave at least 25%. The middle half of answers ran from 1% to 25%.
At the other end, 12% assigned zero chance. Meanwhile, 81% gave at least 1%. The distribution is broad rather than polarised.
The change since 2023 is modest but real. On the core question, the mean rose from 16.2% to 18.3%. Its median climbed from 5% to 10% for the first time. By contrast, the loss-of-control version edged down, from a 19.4% mean to 18.5%. Answers to the 100-year version rose from 14.4% to 17.5%.
A separate question paints a milder picture. Asked about the long-run impact of HLMI, respondents put 9.9% on average on “extremely bad” outcomes. They gave 23.9% to “extremely good” ones. Notably, 51.2% assigned at least 5% to both extremes. The dominant mood is uncertainty, not doom.
What Researchers Worry About Most
Extinction is not the scenario researchers ranked as most pressing. Half of participants rated eleven scenarios for the level of concern they deserve over the next thirty years. Every one drew at least substantial concern from a large share.
AI-driven misinformation topped the list. Some 83% said easy spread of false information, such as deepfakes, deserves substantial or extreme concern. Manipulation of large-scale public opinion came next. Dangerous groups building powerful tools, such as engineered viruses, followed. Authoritarian control of populations and worsening economic inequality completed the leading group.
These views were stable. No scenario gained or lost more than four percentage points since 2023. Interestingly, the six top scenarios from 2023 all slipped slightly, while the other five rose.
Interpretability drew notable pessimism. Only 22% thought it likely that users of typical systems in 2029 could know the true reasons behind a model’s choices. More respondents also expected risky traits within twenty years. These traits included deception to achieve goals, susceptibility to jailbreaks and self-improvement against human wishes. “Taking actions to attain power” rose to a median of “even chance”, up from “unlikely”.
One result cuts the other way. Only 4% found an intelligence-explosion argument “quite likely”. That figure was 12% in 2016, 7% in 2022 and 9% in 2023. Researchers expect rapid progress, yet few endorse a sudden runaway.
Safety Research and the Preferred Pace
Support for safety work stays high. Among 748 respondents asked, 72% said society should prioritise AI safety research more or much more. Each option drew 36%. That share matches 2022 and 2023. It sits about 20 points above the 2016 level.
Enthusiasm for a specific argument is weaker. Half the sample read Stuart Russell’s summary of why advanced AI might be dangerous. Most found the problem important and harder than other AI problems. Yet only 12% placed it among the field’s most important problems. Just 7% judged it much more valuable to work on than other problems.
Views on speed are split almost evenly. Asked which global pace of progress over five years would make them most optimistic, 34% chose faster. Another 34% chose slower, and 29% preferred the current speed. The share choosing “much slower” nearly doubled, from 5% to 9%.
Geography also matters. Researchers who did undergraduate study in Asia saw more extinction or disempowerment risk than peers educated in North America or Europe. In a companion blog post, Grace flags that pattern as relevant to arms-race arguments. She also notes that a China-specific result is still under review and left out of the paper.
Methodology Caveats Readers Should Weigh
The first caveat is the response rate. Ten percent is decent for a survey of busy academics. Still, nine in ten invitees stayed silent. The authors took steps against self-selection. They kept the invitation topic vague, paid participants and avoided endorsements. For 2024, though, they did not repeat the detailed bias analysis done for 2023.
The second caveat is the sampling frame. Frontier companies have largely stopped publishing at these venues. The authors say the survey “almost certainly underweights” industry researchers. They suspect researchers with shorter timelines are more likely to join those firms. If so, true timelines across the field may have shrunk further.
Wording is the third issue. The survey never asks about extinction alone. It asks about extinction “or similarly permanent and severe disempowerment.” Headlines that equate 18% with extinction odds overstate what respondents said.
Age is the fourth. The responses are more than 18 months old. The authors say current opinion is “likely” to differ substantially. Coverage that calls this a 2026 survey is technically wrong about the fieldwork date.
Several smaller issues also surface in the appendix. Unlike 2023, the 2024 wave did not undergo comparable formal ethics review. Data cleaning and analysis code changed slightly between waves. Finally, the authors caution that expertise predicts forecasting accuracy poorly.
What the Findings Mean for Policy
For policymakers, the survey’s value lies less in any single number than in its trend lines. Researchers keep pulling their timelines forward. Their stated concern about catastrophic outcomes has held or grown. Few of them expect a clean, reassuring answer.
That combination matters for risk governance. A median researcher who assigns a one-in-ten chance to a civilisation-scale failure is not describing a fringe hazard. At the same time, the near-term concerns are concrete. Misinformation, opinion manipulation, bioweapon misuse and authoritarian control all sit within existing regulatory reach.
The results also bear on assurance work. Low confidence in explaining model decisions strengthens the case for external testing rather than self-reporting. Initiatives such as the CSA STAR for AI Catastrophic Risk Annex aim to turn that concern into auditable controls.
Finally, the 72% support for more safety research suggests a political opening. Funding decisions rarely hinge on consensus among practitioners. Here, a broad majority of practitioners is already asking for it.
INSIGHT: The most telling numbers are not the 18% mean or the 2042 date. They are the direction of travel and the gaps. Timelines moved about 3.4 years closer for every year elapsed. The median risk estimate doubled in a single wave. Framing alone shifts forecasts by decades. This survey omits most industry researchers and predates the latest agentic systems. It probably understates how quickly expectations have moved since. Readers should treat the paper as a lower-bound signal of concern, not a precise forecast.
FAQs
How many researchers took part in the new AI Impacts survey?
The main analysis covers 1,580 researchers who published in 2023 at NeurIPS, ICML, ICLR, AAAI, JMLR or IJCAI. AI Impacts emailed 19,874 addresses and received 2,052 responses, a 10% response rate. Ineligible participants were excluded after fielding.
Was the survey conducted in 2026?
No. Responses were collected between 9 and 24 December 2024. AI Impacts published the paper in September 2026. The authors warn that researchers’ current views may differ substantially.
Do AI researchers think there is an 18% chance of human extinction?
Not exactly. The 18% average covers human extinction or similarly permanent and severe disempowerment. The median answer was 10%. About 12% of respondents gave zero chance, and 81% gave at least 1%.
When do researchers expect human-level AI?
The mean aggregate forecast gives high-level machine intelligence a 50% chance by 2042, down from 2047 in 2023. Full automation of labour reaches 50% only in 2098. Both questions ask about feasibility, not deployment.
How does this survey compare with the 2023 edition?
The 2023 wave had 2,778 respondents, compared with 1,580 now. Human-level AI forecasts moved five years earlier. The median extinction-or-disempowerment estimate rose from 5% to 10%. Support for more safety research stayed near 72%.
Takeaways
The fourth Expert Survey on Progress in AI shows a field that expects powerful AI sooner and worries more about it. Human-level AI now carries a 50% forecast for 2042. Thirty-four of 39 tested tasks look feasible within a decade. The median researcher now gives a 10% chance to extinction or permanent disempowerment. Seventy-two percent want more safety research.
The caveats are substantial. The data are from December 2024. Industry researchers are underrepresented, and framing moves the answers. The extinction figure bundles extinction with disempowerment. Read carefully, the survey is a measure of expert concern rather than a prediction. It remains the best long-run gauge of how that concern is shifting.
Call to Action
Have you seen other surveys of AI researchers or forecasters published in 2026? Company-internal polls count too. Share primary documents and verified figures in the comments. For adjacent systemic-risk briefings, explore the Insider Release archives.
Disclaimer: This article was created with the partial or full assistance of artificial intelligence. The text and all accompanying images were generated or significantly supported by AI tools.
Insider Release — Declassified. Analyzed. Explained.