10% → 50% → 90% probability Central 80% bands · all rows below
Historical forecasts remain dated. A predicted date passing does not establish that the milestone happened.
FORECAST / DEFINITION
December 2024 SURVEY · N=158
Mathematical problem solving
Match the best humans in the Putnam competition
2024 → 2055 · 10–90%50% BY 2029
December 2024 SURVEY · N=135
Security research
Find and patch a flaw in a system with 100,000+ users
2025 → 2087 · 10–90%50% BY 2031
December 2024 SURVEY · N=164
Machine-learning research
Conduct ML research and write a conference paper
2026 → 2183 · 10–90%50% BY 2036
2023 SURVEY · N=1714
Human-level machine intelligence
Historical comparison · 2023 researcher survey
2027 → 2186 · 10–90%50% BY 2047
Select a band for its source. Arrows mark portions outside the selected view; the printed dates always show the complete interval. A median marks 50% probability, not a deadline.
Exact figures & source links
Historical forecast snapshots; probabilities refer to attainment by the stated year.
The December 2024 AI Impacts survey reports HLMI thresholds of 10% in 2027 and 50% in 2042. The prior survey’s 50% threshold was 2047.
Those two thresholds do not give a complete 10–90% interval. We have not verified the newer upper endpoint, so we show the verified thresholds here rather than inventing a band.
Metaculus maintains a forecast for when a general AI system will be devised, tested, and announced. Its question-specific criteria differ from researcher-survey definitions.
Its public pages returned inconsistent snapshots during our review. No Metaculus dates are plotted until the distribution and resolution criteria can be captured together. This page is not a live feed.
The Bulletin of the Atomic Scientists set its Doomsday Clock to 85 seconds to midnight on 27 January 2026. This is a symbolic assessment involving multiple global threats, reassessed regularly.
It does not measure time remaining, or the probability of AI extinction. We keep it outside the forecast axis.
Survey bands combine fitted individual probability distributions. They reflect uncertainty and diverse beliefs, but are not direct measures of researcher agreement. Survey wording, sample selection, and aggregation affect the results.
Mathematical problem solving — source & interpretation
Match the best humans in the Putnam competition. Short label; consult the linked report for the complete task wording.
10% by 2024 · 50% by 2029 · 90% by 2055
Survey: December 2024. Report: September 2026. Responses used: 158. Checked 2026-10-07.
A specific competition task; success would not establish general intelligence. This is a historical forecast, not an assessment of whether the task has since been achieved.
The band covers the central 80% of the aggregate forecast distribution. It is not the interval containing 80% of respondents’ individual median answers. Dates are reported calendar years, without an invented day or month.
Find and patch a flaw in a system with 100,000+ users. Short label; consult the linked report for the complete task wording.
10% by 2025 · 50% by 2031 · 90% by 2087
Survey: December 2024. Report: September 2026. Responses used: 135. Checked 2026-10-07.
A capability forecast concerning flaw discovery and repair. It supplies neither a probability nor a date for a cyberattack or catastrophic outcome.
The band covers the central 80% of the aggregate forecast distribution. It is not the interval containing 80% of respondents’ individual median answers. Dates are reported calendar years, without an invented day or month.
Machine-learning research — source & interpretation
Conduct ML research and write a conference paper. Short label; consult the linked report for the complete task wording.
10% by 2026 · 50% by 2036 · 90% by 2183
Survey: December 2024. Report: September 2026. Responses used: 164. Checked 2026-10-07.
A particular research task, not complete automation of the AI-research occupation. The long tail is part of the reported forecast and must remain visible.
The band covers the central 80% of the aggregate forecast distribution. It is not the interval containing 80% of respondents’ individual median answers. Dates are reported calendar years, without an invented day or month.
Unaided machines outperforming human workers across tasks at lower cost, excluding intrinsically human advantages. Assumes no major disruption to science. The newer 2024 survey is discussed below; this row is explicitly historical.
The band covers the central 80% of the aggregate forecast distribution. It is not the interval containing 80% of respondents’ individual median answers. Dates are reported calendar years, without an invented day or month.
Capability is one question. Consequences are another.
An earlier capability forecast alone does not establish a higher probability of extinction. Understanding risk also requires examining deployment, misuse, control, and safeguards.
