thebestai.co encyclopedia

Reference

Forecast Register

FCST-REG-001Draftv0.1.0Reviewed 2026-10-15

A living register of 26 documented public forecasts by other parties about AI capability, adoption, AGI and ASI timing, and risk. Each record uses the field set of docs/EVIDENCE_FRAMEWORK.md section 2, plus the original wording, author, publication date, and a resolution status with the evidence for that status. The register records what was forecast and how it has fared; it does not endorse any forecast. Forecasts made by the encyclopedia itself are recorded in the articles that make them (see FCST-METHOD-001, FCST-NEAR-001, FCST-SCEN-001).

#Conventions

  • ID form. FCST-<topic>-NNN, with topics HIST (pre-2000 forecasts), TIMING (human-level, AGI and ASI timing), RISK (catastrophic and existential risk), ECON (labour and economic effects). These IDs are distinct from the encyclopedia's own forecast IDs in other articles.
  • Original wording is a short quotation (under 15 words) or a paraphrase, with the source. Where the quotation or its date was taken from a secondary source because the primary text was not opened, the record says so and carries a [VERIFICATION NEEDED] marker.
  • Confidence Level is the confidence the author stated. Where the author gave none, the field says "not stated" and the register does not supply one.
  • Assumptions and Alternative Scenarios are those stated by the author. Where none were stated the field says so; it then points to registered forecasts that disagree, as a pointer, not as the author's view.
  • Status follows four values: Active (horizon not yet passed), Confirmed, Falsified, Unresolved (the horizon passed or cannot pass, but the claim has no operational test or the evidence is contested). Status is as of 2026-10-15 and is a statement about the dated claim, not about the author.
  • Review Schedule. Annual review each October; the final evaluation date follows each record's horizon.
  • Definitions of "AGI" differ across the authors. The register does not harmonise them; each record states the author's own criterion where one was found.
  • Aggregator pages (for example Metaculus) returned HTTP 403 to automated fetches on 2026-10-15 and no aggregator forecast is recorded. Prediction-tracking sites were used only to locate primary sources, never as sources.

#Summary table

ID Author Published Horizon Status (2026-10-15)
FCST-HIST-001 McCarthy, Minsky, Rochester, Shannon 1955 summer 1956 Unresolved
FCST-HIST-002 Turing 1950 c. 2000 Falsified (as dated)
FCST-HIST-003 Simon and Newell 1958 c. 1968 Falsified (as dated)
FCST-HIST-004 Simon 1960 / 1965 c. 1980-1985 Falsified (as dated)
FCST-HIST-005 Good 1965 20th century Falsified (as dated)
FCST-TIMING-001 Vinge 1993 by 2023 Unresolved
FCST-TIMING-002 Kurzweil 1999, reaffirmed 2024 2029 Active
FCST-TIMING-003 Legg 2011, reaffirmed 2023 by 2028 Active
FCST-TIMING-004 Müller and Bostrom 2016 (survey 2012-2013) 2040-2050 median Active
FCST-TIMING-005 Grace et al. 2017 median about 45 years from survey Active
FCST-TIMING-006 Grace et al. 2024 2047 median Active
FCST-TIMING-007 Aschenbrenner 2024 2027 Active
FCST-TIMING-008 Altman 2024 "a few thousand days" Active
FCST-TIMING-009 Amodei 2024 as early as 2026 Active
FCST-TIMING-010 Altman 2025 (Jan) 2025 Unresolved
FCST-TIMING-011 Hassabis 2025 5-10 years Active
FCST-TIMING-012 Altman 2025 (Jun) 2026 Active
FCST-TIMING-013 Kokotajlo et al. (AI 2027) 2025 2027 Active
FCST-TIMING-014 Musk 2024 2025-2026 Unresolved
FCST-RISK-001 Grace et al. 2024 not time-bounded Unresolved
FCST-RISK-002 Forecasting Research Institute (XPT) 2023 by 2100 Unresolved
FCST-RISK-003 Ord 2020 this century Unresolved
FCST-ECON-001 Hinton 2016 5-10 years Unresolved
FCST-ECON-002 Frey and Osborne 2013 "decade or two" Active
FCST-ECON-003 Goldman Sachs economists 2023 10 years Active
FCST-HIST-006 Simon and Newell (theorem clause) 1958 c. 1968 Unresolved

#Records: pre-2000 forecasts

#FCST-HIST-001: Dartmouth proposal, 1955

Field Content
Forecast ID FCST-HIST-001
Author / date J. McCarthy, M. L. Minsky, N. Rochester, C. E. Shannon; proposal dated 1955-08-31 (reprinted AI Magazine 2006)
Original wording Paraphrase: a "significant advance can be made in one or more of these problems" if a carefully selected group of scientists works on them together "for a summer" (McCarthy et al., 1955/2006; Stanford-hosted copy opened 2026-10-15). The proposal conjectures that every aspect of learning or intelligence can in principle be described precisely enough for a machine to simulate it.
Forecast Statement A two-month, ten-person study in summer 1956 would produce significant advances on problems such as machine use of language, abstraction formation, and self-improvement.
Target Event Significant advance on the proposal's listed sub-problems
Operational Definition None stated. "Significant advance" is not defined.
Time Horizon Summer 1956
Assumptions Funding and a carefully selected group working together (as stated)
Confidence Level Not stated
Alternative Scenarios Not stated by the authors
Review Schedule Closed; no further review unless a source supplies an operational test
Status Unresolved
Evidence for status The claim has no operational threshold, so it cannot be scored as met or missed from the proposal alone. See HIST-002 for the workshop's history.

#FCST-HIST-002: Turing's imitation-game prediction, 1950

Field Content
Forecast ID FCST-HIST-002
Author / date A. M. Turing; Mind, October 1950
Original wording Paraphrase, quoted in the Stanford Encyclopedia of Philosophy entry on the Turing Test: "in about fifty years' time" computers with a storage capacity of about 10^9 (the entry's text shows "109", a lost superscript) could play the imitation game well enough that an average interrogator would have no more than a 70 percent chance of a right identification after five minutes of questioning. He added that by the end of the century the use of words would have altered so that one could speak of machines thinking without contradiction.
Forecast Statement By about 2000, programmed computers will play the imitation game so that an average interrogator has at most a 70 percent chance of correct identification after five minutes.
Target Event Interrogator accuracy at most 70 percent after five minutes of questioning
Operational Definition As in the statement; Turing's paper does not fix the interrogator pool or number of trials.
Time Horizon About 2000
Assumptions Computers with storage capacity about 10^9 (units as in the paper; see Turing, 1950)
Confidence Level Stated as belief ("I believe"); no probability
Alternative Scenarios Not stated in the quoted passage
Review Schedule Closed for the dated claim; revisit if independent replications of the 2025 study appear
Status Falsified (as dated)
Evidence for status A 2025 study reports "the first empirical evidence that any artificial system passes a standard three-party Turing test", with GPT-4.5 judged to be the human 73 percent of the time (Jones & Bergen, 2025, arXiv preprint). If that "first" claim is accepted, no system met the criterion around 2000. The 2025 result is a single preprint and its protocol differs from Turing's (three-party versus two-party); it is Preliminary evidence. [VERIFICATION NEEDED: independent evidence on systems tested around 2000 was not retrieved.]

#FCST-HIST-003: Simon and Newell on a chess champion, 1958

Field Content
Forecast ID FCST-HIST-003
Author / date H. A. Simon and A. Newell; Operations Research 6(1), January-February 1958, pp. 1-10 (address delivered 1957 per secondary account)
Original wording "within ten years a digital computer will be the world's chess champion" (reproduced by Quote Investigator with the qualifier "unless the rules bar it from competition", citing pp. 7-8).
Forecast Statement Within ten years of 1957/1958, a digital computer will be the world's chess champion, unless rules bar it.
Target Event A computer holds the world chess championship
Operational Definition Not stated beyond "world's chess champion".
Time Horizon About 1967-1968
Assumptions The rules do not bar computers (stated)
Confidence Level Not stated; presented as a prediction
Alternative Scenarios The "unless the rules bar it" clause
Review Schedule Closed
Status Falsified (as dated)
Evidence for status No computer held the title by about 1968. The Deep Blue system, described in Campbell et al. (2002), beat the reigning champion Kasparov in a 1997 match; secondary accounts note this was not a formal world-championship event and came about 39 years after 1958. [VERIFICATION NEEDED: 1997 match details were not extracted from the Campbell et al. paper; the match date comes from the secondary account.] The quotation is from a secondary source; the Operations Research text was not opened.

#FCST-HIST-004: Simon on machines doing any human work, 1960/1965

Field Content
Forecast ID FCST-HIST-004
Author / date H. A. Simon; chapter first in The New Science of Management Decision (1960), reprinted in The Shape of Automation for Men and Management (1965), per Quote Investigator
Original wording Paraphrase: technologically, machines will be able to do any work a person can do "within twenty years". Simon paired this with an economic argument that people would keep an edge in some kinds of work because computers were expensive (per Quote Investigator).
Forecast Statement Within twenty years of 1960 (or 1965), machines will be technologically capable of doing any work a human can do.
Target Event Machine capability to perform any human work
Operational Definition Not stated
Time Horizon About 1980 (from 1960) or 1985 (from 1965)
Assumptions Technological capability only; economics treated separately by the author
Confidence Level Not stated
Alternative Scenarios The economic argument above (labour retains advantages where machines are expensive)
Review Schedule Closed
Status Falsified (as dated)
Evidence for status The horizon passed in the 1980s without machines able to do any human work; no source retrieved in this pass states otherwise as of 2026. [VERIFICATION NEEDED: the primary texts were not opened; the wording, dates, and edition details come from the Quote Investigator trace. No dataset assessing 1980-1985 capability was retrieved.]

#FCST-HIST-005: Good on the ultraintelligent machine, 1965

Field Content
Forecast ID FCST-HIST-005
Author / date I. J. Good; "Speculations Concerning the First Ultraintelligent Machine", Advances in Computers vol. 6 (Crossref issued date 1966; commonly cited as 1965); the introduction states the first draft was completed in April 1963 and the emended version in May 1964 (per a search result describing the scanned copy)
Original wording Paraphrase: it is more probable than not that an ultraintelligent machine will be built within the twentieth century. [VERIFICATION NEEDED: the exact sentence was not located in the primary text; it is reported by a secondary quotation site and the scanned paper was not read.]
Forecast Statement An ultraintelligent machine will be built within the twentieth century (probability above 50 percent).
Target Event Construction of a machine that surpasses all intellectual activities of any person (Good's definition of ultraintelligent; definition [VERIFICATION NEEDED])
Operational Definition As above; no test stated in the retrieved material
Time Horizon Before 2001
Assumptions Not extracted
Confidence Level "More probable than not" (above 50 percent), per the secondary quotation
Alternative Scenarios Not extracted
Review Schedule Closed
Status Falsified (as dated)
Evidence for status The horizon ended in 2000 without a machine surpassing all human intellectual activity; no retrieved source claims such a machine exists as of 2026. Wording and definition unverified, so the status itself is provisional.

#FCST-HIST-006: Simon and Newell on a mathematical theorem, 1958

Field Content
Forecast ID FCST-HIST-006
Author / date H. A. Simon and A. Newell; Operations Research 6(1), 1958
Original wording Reproduced by Quote Investigator (paraphrase): within ten years a computer will "discover and prove an important new mathematical theorem" (quoted from that secondary account).
Forecast Statement Within about ten years a computer will discover and prove an important new theorem.
Target Event Computer-discovered, computer-proved important mathematical theorem
Operational Definition "Important" is undefined
Time Horizon About 1968
Assumptions Not stated
Confidence Level Not stated
Alternative Scenarios Not stated
Review Schedule Closed
Status Unresolved
Evidence for status "Important" has no operational test, so the claim cannot be scored. The quotation is from a secondary account; the primary text was not opened. [VERIFICATION NEEDED]

#Records: timing of human-level AI, AGI and ASI

#FCST-TIMING-001: Vinge, 1993

Field Content
Forecast ID FCST-TIMING-001
Author / date V. Vinge; "The Coming Technological Singularity: How to Survive in the Post-Human Era", VISION-21 Symposium (NASA Lewis Research Center and Ohio Aerospace Institute), 30-31 March 1993
Original wording "Within thirty years, we will have the technological means to create superhuman intelligence." (essay abstract, fetched 2026-10-15)
Forecast Statement By 2023 the technological means to create superhuman intelligence will exist.
Target Event Technological means to create superhuman intelligence
Operational Definition Not given in the abstract; "means to create" does not require that such an intelligence exist.
Time Horizon By 2023
Assumptions Not extracted
Confidence Level Not stated in the abstract
Alternative Scenarios Not extracted
Review Schedule Closed for the date; revisit on a source that applies an operational definition
Status Unresolved
Evidence for status The horizon has passed. Whether "means to create" and "superhuman" were satisfied depends on definitions the abstract does not supply, and the register found no source applying a test.

#FCST-TIMING-002: Kurzweil, 1999 and 2024

Field Content
Forecast ID FCST-TIMING-002
Author / date R. Kurzweil; The Age of Spiritual Machines (1999); reaffirmed in The Singularity Is Nearer (Viking / Penguin Random House, 2024-06-25) and a Science Friday interview (2024-06-14)
Original wording Paraphrase of the 2024 interview: in 1999 he predicted that by 2029 "we'd reach the Turing test and also AGI" (Science Friday's rendering), and in 2024 he kept the date. For the Singularity he said he is sticking with 2045.
Forecast Statement Human-level AI, including passing the Turing test, by 2029; the Singularity by 2045.
Target Event Human-level AI (2029); the Singularity as he defines it (2045)
Operational Definition In the interview, AGI is described as a computer able to do anything a human can; no formal test stated there.
Time Horizon 2029 (human-level AI); 2045 (Singularity)
Assumptions Not extracted (exponential growth trends are his stated basis; not verified here)
Confidence Level Not stated
Alternative Scenarios Competing registered forecasts: FCST-TIMING-004, FCST-TIMING-006
Review Schedule Annual; final evaluation 2030 for the 2029 clause
Status Active
Evidence for status Horizon not reached. A 2025 preprint reports one Turing-test protocol passed by GPT-4.5 (Jones & Bergen, 2025), which bears on the Turing-test component only; the AGI component has no operational test.

#FCST-TIMING-003: Legg, 2011 and 2023

Field Content
Forecast ID FCST-TIMING-003
Author / date S. Legg; 2011 blog post read aloud in a podcast with D. Patel, published 2023-10-26
Original wording "there's a 50% chance that we have AGI by 2028" (2023 interview). The 2011 post, as read in the interview, describes a log-normal distribution with a mean of 2028 and a mode of 2025, assuming no catastrophic interruption.
Forecast Statement A 50 percent probability of AGI by 2028.
Target Event AGI
Operational Definition Legg defines AGI in the interview as a machine that can do the cognitive things people can typically do, possibly more; a system counts once adversarial testers cannot find cognitive tasks humans do and the machine fails.
Time Horizon By 2028 (mode 2025 in the 2011 statement)
Assumptions No catastrophic event intervenes (2011 post, as read)
Confidence Level 50 percent by 2028
Alternative Scenarios He states he would not be surprised if it does not happen by then.
Review Schedule Annual; final evaluation January 2029
Status Active
Evidence for status Horizon not reached; the probabilistic forecast cannot be scored by a single outcome.

#FCST-TIMING-004: Müller and Bostrom expert survey, 2016

Field Content
Forecast ID FCST-TIMING-004
Author / date V. C. Müller and N. Bostrom; survey distributed 2012-2013 to four expert groups; published 2016 in Fundamental Issues of Artificial Intelligence (Synthese Library, Springer); preprint arXiv:2508.11681
Original wording Paraphrase of the abstract: the median estimate for a 50 percent chance of high-level machine intelligence is 2040-2050, rising to 90 percent by 2075; superintelligence is expected within less than 30 years after that; respondents gave about one in three that the outcome is "bad" or "extremely bad" for humanity.
Forecast Statement Median respondent: 50 percent chance of high-level machine intelligence by 2040-2050; 90 percent by 2075.
Target Event High-level machine intelligence (the survey's term)
Operational Definition The survey's definition of high-level machine intelligence was not extracted.
Time Horizon 2040-2050 (50 percent); 2075 (90 percent)
Assumptions Not extracted
Confidence Level Aggregate stated probabilities as above
Alternative Scenarios Individual respondents differed; spread not extracted. Selection and response-rate caveats are discussed in secondary accounts [VERIFICATION NEEDED: response rate].
Review Schedule Annual; first resolution window 2040
Status Active
Evidence for status Horizon not reached.

#FCST-TIMING-005: Grace et al. survey, 2017

Field Content
Forecast ID FCST-TIMING-005
Author / date K. Grace, J. Salvatier, A. Dafoe, B. Zhang, O. Evans (5 authors); arXiv:1705.08807, 2017-05-24
Original wording Paraphrase of the abstract: researchers believe there is a 50 percent chance of AI outperforming humans in all tasks in 45 years and of automating all human jobs in 120 years; milestones named include translating languages by 2024, writing high-school essays by 2026, driving a truck by 2027, and writing a bestselling book by 2049.
Forecast Statement 50 percent chance of AI outperforming humans in all tasks in 45 years; of full automation of all jobs in 120 years (survey elicitation year not extracted).
Target Event AI outperforming humans in all tasks; automation of all human jobs
Operational Definition Not extracted
Time Horizon 45 years and 120 years from the survey
Assumptions Not extracted
Confidence Level 50 percent (aggregate)
Alternative Scenarios The abstract notes Asian respondents expected the dates much sooner than North American respondents.
Review Schedule Annual; milestone dates 2024-2027 can be checked individually
Status Active
Evidence for status The all-task and all-jobs horizons are decades away. Individual milestones with passed dates (for example translation by 2024) were not assessed in this pass.

#FCST-TIMING-006: Grace et al. survey, 2024

Field Content
Forecast ID FCST-TIMING-006
Author / date K. Grace, H. Stewart, J. F. Sandkühler, et al. (7 authors); arXiv:2401.02843, submitted 2024-01-05; 2,778 researchers who had published at top-tier AI venues
Original wording Paraphrase of the abstract: aggregate forecasts give a 10 percent chance by 2027 and a 50 percent chance by 2047 of machines outperforming humans at every possible task unaided; for all human occupations being fully automatable, 10 percent by 2037 and 50 percent by 2116. The 2047 date is 13 years earlier than in the 2022 survey.
Forecast Statement Machines unaided outperform humans at every task: 10 percent by 2027, 50 percent by 2047.
Target Event Unaided machines outperforming humans at all tasks; full automation of all occupations
Operational Definition The survey's question wording was not extracted.
Time Horizon 2027 (10 percent); 2047 (50 percent); 2037 and 2116 for occupations
Assumptions Not extracted
Confidence Level Aggregate 10 percent and 50 percent quantiles as above (the register records the quantiles, not a single median across all questions)
Alternative Scenarios Earlier survey (2022) gave later dates; see FCST-TIMING-005 and FCST-TIMING-004
Review Schedule Annual; first quantile window 2027
Status Active
Evidence for status Horizons not reached.

#FCST-TIMING-007: Aschenbrenner, "Situational Awareness", 2024

Field Content
Forecast ID FCST-TIMING-007
Author / date L. Aschenbrenner; essay series dated June 2024 (author named on the site's About link)
Original wording "AGI by 2027 is strikingly plausible." The series frames this as a plausible trajectory from trend lines, calls AGI "plausible" and not inevitable, and acknowledges large error bars (as summarised from the fetched page).
Forecast Statement Models will be able to do the work of an AI researcher or engineer by 2027 (author's AGI criterion, per the summary of the page).
Target Event AGI as defined by AI-researcher-level work
Operational Definition AI researcher/engineer-level work, per the page summary; no detailed test extracted.
Time Horizon 2027
Assumptions Continuation of compute and algorithmic trends; the author names data limits and breakdown of trends as risks
Confidence Level Not given as a number; "plausible"
Alternative Scenarios Progress could stall if data limits or algorithmic breakthroughs prove harder
Review Schedule Annual; final evaluation January 2028
Status Active
Evidence for status Horizon not reached.

#FCST-TIMING-008: Altman, "The Intelligence Age", 2024

Field Content
Forecast ID FCST-TIMING-008
Author / date S. Altman; blog post "The Intelligence Age", 2024-09-23
Original wording Superintelligence is possible "in a few thousand days"; he acknowledges it "may take longer" but is confident it will arrive.
Forecast Statement Superintelligence will arrive within a few thousand days of September 2024.
Target Event Superintelligence
Operational Definition None stated
Time Horizon "A few thousand days" (undefined; the register does not convert it to a date)
Assumptions Not extracted
Confidence Level Stated as confident; qualitatively hedged ("may take longer")
Alternative Scenarios Longer timeline acknowledged
Review Schedule Annual; no fixed final date because the horizon is undefined
Status Active
Evidence for status No horizon defined, and no source retrieved indicates arrival.

#FCST-TIMING-009: Amodei, "Machines of Loving Grace", 2024

Field Content
Forecast ID FCST-TIMING-009
Author / date D. Amodei; essay, October 2024
Original wording Paraphrase: "powerful AI" could come "as early as 2026", though it may take much longer. The essay's definition: a model, likely similar to current LLMs, that outperforms a Nobel Prize winner across most relevant fields, uses computer interfaces, works autonomously on tasks lasting hours to weeks, can control physical tools through a computer, and could run as millions of instances at roughly 10-100 times human speed.
Forecast Statement Powerful AI, as defined in the essay, could arrive as early as 2026.
Target Event "Powerful AI" meeting the essay's listed properties
Operational Definition The property list above; no test thresholds.
Time Horizon As early as 2026
Assumptions The essay sets the timing question aside and focuses on the 5-10 years after arrival
Confidence Level "Could"; no probability
Alternative Scenarios It could take much longer (stated)
Review Schedule Annual; final evaluation January 2027
Status Active
Evidence for status The calendar horizon (end of 2026) has not passed as of 2026-10-15. The developer's Claude Opus 5.5 page (2026-09-22) makes no claim that this definition is met (MODEL-001 M-53).

#FCST-TIMING-010: Altman, "Reflections", January 2025

Field Content
Forecast ID FCST-TIMING-010
Author / date S. Altman; blog post "Reflections", early January 2025 (secondary reports give January 5 and January 6; the fetched page showed no date)
Original wording Paraphrase: "we may see the first AI agents 'join the workforce'" in 2025 and materially change company output. In the same post, he states OpenAI is confident it knows how to build AGI "as we have traditionally understood it", and is turning its aim to superintelligence.
Forecast Statement In 2025 the first AI agents may join the workforce and materially change the output of companies.
Target Event AI agents joining the workforce with material effect on company output
Operational Definition None stated. "Join the workforce" and "materially" are undefined.
Time Horizon 2025
Assumptions Not extracted
Confidence Level "May"; no probability
Alternative Scenarios Not stated
Review Schedule Closed unless an operational definition is supplied
Status Unresolved
Evidence for status Horizon passed; the claim is hedged ("may") and has no operational definition, so neither confirmation nor falsification can be established from the source. The 2025 date is from the fetched page; day-level date is from secondary reports and is uncertain.

#FCST-TIMING-011: Hassabis, March 2025

Field Content
Forecast ID FCST-TIMING-011
Author / date D. Hassabis, reported by R. Browne (CNBC; republished by NBC News), 2025-03-17
Original wording Paraphrase: over the next five to ten years AI will start moving toward what he calls artificial general intelligence; on artificial superintelligence he said "no one really knows".
Forecast Statement AGI capabilities start emerging within 5-10 years of March 2025.
Target Event Movement toward AGI
Operational Definition Not stated in the report
Time Horizon 2030-2035
Assumptions Not extracted
Confidence Level Not stated in this report
Alternative Scenarios Superintelligence timing declared unknown
Review Schedule Annual; final evaluation 2035
Status Active
Evidence for status Horizon not reached. Later reported statements by the same speaker (for example a January 2026 Davos remark reported by a secondary outlet) were not verified and are not recorded.

#FCST-TIMING-012: Altman, "The Gentle Singularity", June 2025

Field Content
Forecast ID FCST-TIMING-012
Author / date S. Altman; blog post "The Gentle Singularity", 2025-06-10 (date per a community-forum post listing and press coverage; the fetched page showed no date)
Original wording "2026 will likely see the arrival of systems that can figure out novel insights." He adds that 2027 may see robots doing real-world tasks.
Forecast Statement During 2026, AI systems will arrive that can figure out novel insights.
Target Event AI systems generating novel insights
Operational Definition None stated
Time Horizon 2026
Assumptions Not extracted
Confidence Level "Likely"; no probability
Alternative Scenarios Not stated
Review Schedule Review January 2027; the claim may be Unresolved for lack of an operational test
Status Active
Evidence for status The calendar horizon has not passed. No operational test exists, so a later resolution may also be Unresolved.

#FCST-TIMING-013: AI 2027 scenario, 2025

Field Content
Forecast ID FCST-TIMING-013
Author / date D. Kokotajlo, S. Alexander, T. Larsen, E. Lifland, R. Dean; published 2025-04-03; later notes dated July 2025, 2025-11-22, and December 2025
Original wording Paraphrase of the site: the scenario places a "superhuman coder" around March 2027, with 2027 the modal year for AGI at publication and medians somewhat later; superintelligence in the racing ending arrives around late 2027. The authors say they do not know when AGI arrives and present one scenario of many.
Forecast Statement A superhuman coder by about March 2027; superintelligence by late 2027 in the scenario's racing branch.
Target Event Superhuman coder; superintelligent AI researcher; artificial superintelligence (scenario milestones)
Operational Definition Defined in the scenario's milestone definitions (not extracted)
Time Horizon 2027
Assumptions Continued AI-accelerated AI research; compute scale-up (page summary)
Confidence Level Substantial uncertainty stated; a July 2025 update moved the median back about 1.5 years; one forecaster's median is 2030
Alternative Scenarios The authors' slower-takeoff ending and later medians
Review Schedule Annual; final evaluation January 2028
Status Active
Evidence for status Horizon not reached. The authors' own timelines moved later in 2025 notes, which the register records as author updates, not as a resolution.

#FCST-TIMING-014: Musk, April 2024

Field Content
Forecast ID FCST-TIMING-014
Author / date E. Musk, in an interview with N. Tangen of Norway's sovereign wealth fund; Reuters story dated 2024-04-08
Original wording Paraphrase: if AGI is defined as smarter than the smartest human, it is "probably next year, within two years" (as quoted in the Reuters story reproduced by syndication sites; the Reuters page itself was not opened).
Forecast Statement AI smarter than the smartest human arrives by 2025 or within two years (by 2026).
Target Event AI smarter than the smartest human
Operational Definition His stated definition of AGI; no test
Time Horizon 2025 to 2026
Assumptions Electricity and chip supply were named as constraints in the same interview
Confidence Level "Probably"
Alternative Scenarios Not stated
Review Schedule Review January 2027
Status Unresolved
Evidence for status The 2025 date passed; the 2026 horizon is still open on 2026-10-15. "Smarter than the smartest human" has no operational test. [VERIFICATION NEEDED: the Reuters original was not opened.]

#Records: risk

#FCST-RISK-001: Grace et al. survey on extinction-level outcomes, 2024

Field Content
Forecast ID FCST-RISK-001
Author / date K. Grace et al.; arXiv:2401.02843, 2024-01-05 (same survey as FCST-TIMING-006)
Original wording Paraphrase of the abstract: between 38 percent and 51 percent of respondents gave at least a 10 percent chance to advanced AI leading to outcomes as bad as human extinction; 68.3 percent thought good outcomes from superhuman AI were more likely than bad.
Forecast Statement A sizeable share of researchers assign at least 10 percent probability to extinction-level outcomes from advanced AI.
Target Event Outcome as bad as human extinction caused by advanced AI
Operational Definition Per the survey's question wording (not extracted)
Time Horizon Not time-bounded in the abstract
Assumptions Not extracted
Confidence Level Distribution of respondents' probabilities, as above
Alternative Scenarios The abstract reports majorities expecting good outcomes were more likely
Review Schedule Re-survey when a new edition is published
Status Unresolved
Evidence for status The outcome is not time-bounded and cannot be scored against a single event; a survey of beliefs is not a resolvable forecast without a horizon.

#FCST-RISK-002: Existential Risk Persuasion Tournament, 2023

Field Content
Forecast ID FCST-RISK-002
Author / date Forecasting Research Institute, "Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament", 2023-07-10 (revised 2023-08-08); tournament in 2022 with 169 forecasters (80 domain experts, 89 superforecasters)
Original wording Paraphrase of the page: the largest disagreement was over AI risk, and superforecasters were unmoved by experts' much higher estimates of AI extinction risk. Secondary summaries report median AI-extinction-by-2100 forecasts of 3 percent for experts and 0.38 percent for superforecasters. [VERIFICATION NEEDED: these two figures and the catastrophe figures come from secondary summaries; the FRI working paper was not opened.]
Forecast Statement AI-caused human extinction by 2100: expert median 3 percent, superforecaster median 0.38 percent (secondary figures).
Target Event Human extinction caused by AI
Operational Definition Per the tournament's question text (not extracted)
Time Horizon By 2100
Assumptions Not extracted
Confidence Level Group median probabilities as above
Alternative Scenarios Disagreement between the two groups persisted after discussion (per the page)
Review Schedule Annual; no resolution expected before 2100
Status Unresolved
Evidence for status The horizon is 2100; the tournament also collected short-run questions whose resolution was not assessed here.

#FCST-RISK-003: Ord, The Precipice, 2020

Field Content
Forecast ID FCST-RISK-003
Author / date T. Ord; The Precipice: Existential Risk and the Future of Humanity, Hachette Books, 2020-03-24 (Open Library record; ISBN 9780316484916)
Original wording Paraphrase per secondary sources: unaligned artificial general intelligence carries about a 1 in 10 existential risk over the next century, which he rates above any other single source; total existential risk this century about 1 in 6. He says the figures convey order of magnitude, not precise probabilities (per secondary accounts). [VERIFICATION NEEDED: the book text was not opened.]
Forecast Statement About 1 in 10 probability of existential catastrophe from unaligned AI this century.
Target Event Existential catastrophe from unaligned AI
Operational Definition The book's definition of existential catastrophe and of unaligned AI (not extracted; one critic notes it is not clearly defined)
Time Horizon Next 100 years
Assumptions Not extracted
Confidence Level About 1 in 10, rough
Alternative Scenarios Compare FCST-RISK-002 (lower superforecaster median)
Review Schedule Annual; no resolution expected before 2100
Status Unresolved
Evidence for status Century-scale horizon; not resolvable now.

#Records: labour and economic effects

#FCST-ECON-001: Hinton on radiologists, 2016

Field Content
Forecast ID FCST-ECON-001
Author / date G. Hinton; remarks at the 2016 conference "Machine Learning and the Market for Intelligence" (Toronto, per a secondary account)
Original wording Reported: people should "stop training radiologists" now because deep learning will do better than radiologists in five years, "it might be ten years". (Reported by Radiology Business and other outlets; the Radiology Business page returned HTTP 403, so the wording is from a search-result summary.) [VERIFICATION NEEDED: primary recording or transcript not opened.]
Forecast Statement Within 5-10 years of 2016, deep learning will outperform radiologists.
Target Event Deep learning outperforming radiologists
Operational Definition None stated; the task and benchmark are unspecified.
Time Horizon 2021 to 2026
Assumptions Not stated
Confidence Level "Obvious" in tone; hedged to ten years
Alternative Scenarios Langlotz (2019) asks whether AI will replace radiologists, in Radiology: Artificial Intelligence (abstract not extracted)
Review Schedule Review annually until the horizon closes (2026)
Status Unresolved
Evidence for status The performance claim has no operational test. Secondary reports say the implied employment consequence did not occur: a New York Times piece summarised by Radiology Business reported Mayo Clinic radiologist staff up 55 percent since 2016, and another outlet cited Bureau of Labor Statistics figures of about 230,000 radiology jobs in 2014 and 270,000 in 2024. Those figures are from search-result summaries and were not verified against the original sources.

#FCST-ECON-002: Frey and Osborne, 2013

Field Content
Forecast ID FCST-ECON-002
Author / date C. B. Frey and M. A. Osborne; Oxford Martin working paper 2013-09-01; published as "The future of employment: How susceptible are jobs to computerisation?", Technological Forecasting and Social Change 114, 254-280 (2017)
Original wording Paraphrase of the Oxford Martin page: about 47 percent of total US employment is at risk, estimated across 702 occupations with a Gaussian process classifier. The page does not give a time frame; the paper's own horizon wording (reported as "perhaps a decade or two") was not confirmed.
Forecast Statement About 47 percent of US employment is susceptible to computerisation.
Target Event Computerisation of occupations accounting for about 47 percent of US employment
Operational Definition An occupation-level probability of computerisation, high-risk threshold per the paper (not extracted)
Time Horizon Reported as a decade or two [VERIFICATION NEEDED: wording not confirmed from the paper]
Assumptions Technical feasibility only, per the paper's framing (not extracted)
Confidence Level Occupation-level probabilities; no aggregate confidence extracted
Alternative Scenarios Compare FCST-ECON-003, which frames its figure as exposure, not displacement
Review Schedule Annual; evaluation window 2023-2033
Status Active
Evidence for status The inferred window (2023-2033) has not closed; no direct test of the 47 percent figure was retrieved.

#FCST-ECON-003: Goldman Sachs on generative AI and GDP, 2023

Field Content
Forecast ID FCST-ECON-003
Author / date J. Briggs and D. Kodnani (Goldman Sachs Global Investment Research), 2023 (month and day not confirmed from the page; the Goldman Sachs article page returned HTTP 403)
Original wording Paraphrase per search-result summaries: generative AI could raise global GDP by about 7 percent (almost USD 7 trillion) and lift productivity growth by 1.5 percentage points over ten years; roughly 300 million full-time jobs could be exposed to automation. The authors say exposure does not equal displacement. [VERIFICATION NEEDED: figures from search-result summaries; original report not opened.]
Forecast Statement Generative AI could raise annual global GDP by about 7 percent over ten years.
Target Event Cumulative GDP uplift of about 7 percent
Operational Definition Per the report (not extracted)
Time Horizon Ten years from 2023 (reportedly eventually, with timing hard to predict)
Assumptions Not extracted
Confidence Level "Could"; no probability
Alternative Scenarios The report says most jobs are only partially exposed, so more complemented than substituted
Review Schedule Annual; evaluation 2033
Status Active
Evidence for status Horizon not reached; no measurement of the figure was retrieved.

#Notes on evidence ratings

Under docs/EVIDENCE_FRAMEWORK.md, every record above is a Forecast (section 1.6). The surveys (FCST-TIMING-004 to FCST-TIMING-006, FCST-RISK-001, FCST-RISK-002) report aggregated beliefs; the register treats the aggregate probabilities as forecasts by the respondents, not as established facts about the future. The 2025 Turing-test study cited under FCST-HIST-002 is a single preprint and is Preliminary.

#References

#Revision history

  • v0.1.0 (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.