The AI Impacts survey series, running since 2016, published its fourth round in September 2026 based on data collected in December 2024. The study maps how researchers view long‑term risks from artificial intelligence, including civilization‑scale harms and persistent loss of human power. The authors note that by the time of publication the data were already more than 18 months old.
Sample and methodology
Invitations were sent to 19,874 e‑mail addresses of authors who published at six major venues in 2023 (NeurIPS, ICML, ICLR, AAAI, IJCAI, JMLR). There were 2,052 responses (about a 10% response rate). After filtering out mistakenly invited addresses, 1,580 respondents remained for analysis. The authors stress that the sample is not representative of the entire AI industry; researchers at frontier AI companies are likely underrepresented because those organizations have been publishing less on the venues sampled.
When do researchers expect high‑level machine intelligence (HLMI)?
The survey defined HLMI as machines being able to perform all tasks better and more cheaply than humans. Respondents were asked to estimate the year in which there would be a 50% chance of HLMI. Median estimates in past rounds were:
- 2016: 2061
- 2022: 2059
- 2023: 2047
- 2024 (current data): 2042
In other words, over eight years the median 50% date shifted from 2061 to 2042 — a move from 45 years out to 18 years out. The biggest jump was between 2022 and 2023; the 2024 survey moved the median another five years earlier. When asked a different question — when all occupations would be automatable — the median 50% date was much later: 2098 (56 years later than the HLMI median).
The authors note that differing question wordings and definitions partly explain these gaps; nevertheless, the estimated timing matters because it affects how much time society, regulators and developers have to address safety issues.
How large do researchers estimate the worst outcomes to be?
The survey’s most cited result concerns how researchers estimate the probability of severe outcomes such as human extinction or comparably long‑lasting loss of human power. From the 2024 responses:
- The average estimate to the general question was 18.3%, with a median of 10% (the 2023 median had been 5%).
- Combining responses to three related questions (n = 1,489) yields an average estimate of 18.2%.
- 51% of respondents assigned at least a 10% probability to the most severe outcomes; 26% assigned at least a 25% probability.
The authors emphasize these figures are expert judgments about an unprecedented future risk rather than measured frequencies; the question’s broad phrasing also includes persistent loss of human power as well as outright extinction.
No consensus on the optimal pace of development
Respondents were split when asked which global development pace would make them most optimistic about the next five years:
- 34% preferred faster development,
- 29% preferred the current pace,
- 34% preferred slower development.
- Support for the “much slower” option rose from 5% to 9%.
This indicates recognizing risks does not translate to a single agreed regulatory or policy prescription within the research community.
Which specific risks worry researchers most?
Over the next thirty years respondents were most concerned about information‑space harms and societal impacts rather than a cinematic robot uprising:
- 83% regarded it a significant or extreme concern that AI would facilitate the spread of false information (for example, deepfakes).
- 77% worried that AI systems could massively manipulate public opinion.
- 69–72% expressed similar concern about dangerous groups gaining access to high‑impact tools (for example engineered pathogens), authoritarian regimes using AI for population control, and increasing economic inequality.
- 46% rated misaligned extremely powerful AI as a significant or extreme concern; 43% rated catastrophe‑causing AI similarly.
Additionally, 72% of respondents said they would support more or much more AI safety research.
Other studies provide a nuanced picture
The authors note other research offers a more nuanced image. In a separate survey of over 4,000 AI researchers, only 3% named existential risk as their primary concern in an open question — a result from a different sample and method that does not directly contradict AI Impacts. The Pew Research Center’s 2025 U.S. survey found 56% of AI experts expected AI’s net effect on the United States over the next twenty years to be positive, while only 17% of U.S. adults agreed. The Longitudinal Expert AI Panel (LEAP) found the public’s median forecasts were generally lower than experts’ — for example, experts expected generative AI to assist about 18% of U.S. work hours by 2030, whereas the public’s median estimate was 10%.
A warning about uncertainty, not a deterministic prophecy
The survey does not present a unified doomsday forecast but documents uncertainty, division, and significant concern within the research community. Many researchers do not regard civilization‑scale risks as negligible, yet there is no consensus on the right policy response. The authors reiterate that these figures are expert judgments with substantial uncertainty and should be interpreted cautiously.
The broader implication is that the recent flurry of commentary about AI risks in media and industry is rooted in several years of measurable expert concern rather than merely episodic media panic.
Tags: risk, technology, survey, artificial intelligence, extinction, deepfake, researchers, disinformation, AI Impacts, HLMI
