AI Impacts Survey (2023)

EMBARGOED UNTIL 0:01 AM PT, THURSDAY JANUARY 4, 2024
Survey: Median AI expert says 5% chance of human extinction
from AI
BERKELEY, CALIFORNIA: In a new survey of 2,778 AI experts, experts gave a median 5%
chance that AI would cause human extinction.
In the survey conducted by AI Impacts, a Berkeley-based think tank, in collaboration with
researchers at the University of Bonn and University of Oxford, experts offered a median
response of 5% for the chance of future AI advances causing human extinction or similarly
permanent and severe disempowerment of the human species. Mean responses indicated an
even higher risk, suggesting a nearly one-in-six (16%) chance of catastrophic outcomes — the
same odds as dying in a game of Russian roulette.
All respondents had recently published a paper in one of six top peer-reviewed AI venues,
demonstrating that this is a prevailing view among experts about the potential risks of an
emerging technology, the inner workings of which are poorly understood. One in ten
respondents put at least a 25% chance on extremely bad outcomes like human extinction, with
1% of respondents thinking there was at least a 75% chance of such outcomes.
Experts also expressed particular concern over the spread of disinformation (including
deepfakes), the potential of AI to exacerbate authoritarianism, and its ability to assist dangerous
groups in making powerful tools like engineered viruses.
The survey also found that powerful AI may arrive much sooner than many people expect.
Experts now see a 50% chance of AI models outperforming humans in every task by 2047,
assuming no major disruption to scientific activity — just over 20 years from now. Notably, this
estimated date is 13 years sooner than the date arrived at by a similar survey in 2022, in which
respondents estimated such AI would not arrive until 2060.
When asked a similar question — when all occupations would be fully automatable — experts
had a much longer time horizon, saying there was a 50% chance of this happening by 2116 —
92 years from now. This also represents an advance of 48 years on their 2022 prediction,
however: in 2022, experts thought this wouldn’t happen until 2164.
Experts expect systems will be able to do many tasks long before then, though. They thought
there was a 50% chance that AI models would be able to produce a new song indistinguishable
from a Taylor Swift song by 2027, and a 50% chance AI could write NYT best-selling fiction by
2031. They also predicted a 50% chance that AI models would be able to perform as well as the
best humans in the prestigious Putnam math competition within eight years.

-- 1 of 7 --

In response to the rapid advance of potentially dangerous AI, 70% of experts thought that AI
safety should be prioritised more than it currently is, with 36% saying it should be prioritised
“more” and a further 34% saying it should be prioritised “much more.”
Commenting on the results, Katja Grace, lead researcher at AI Impacts, said: “These results

... (truncated, 8 KB total)