Predictions on record
What did people and institutions predict about machine intelligence, and when: by what date, and with what stated probability?
Dated predictions about machine intelligence, from Turing’s paper of 1950 to statements of 2025 and 2026. Each record gives the text’s own words for the predicted event, for the horizon and for any probability. A record is dated by when the prediction was made, never by its target. The chart places each prediction with a printed target year by its source date (across) and that year (up). The table below it lists every record, with or without a year.
- The collection does not say whether any prediction came true.
- It does not rank predictors or combine estimates.
- A horizon such as “within ten years” is never turned into a year. Only years a text prints appear on the chart; every other horizon is in the table, in the text’s words.
Showing 65 of 65 records; 36 with a printed target year on the chart.
- Stated forecast. A forecast stated in a paper, essay, post or announcement by its author or publisher.
- Expert survey. A survey’s published result. The figure belongs to the survey report, with its question and respondents, not to any one respondent.
- Public bet. A bet between named parties, with its terms and stakes as the bet’s page states them.
- Forecasting platform. A forecasting platform’s community figure, read from a dated copy of the question page. The figure changes from day to day; the record holds one day’s figure.
- Reported statement. A text’s report of predictions by people it does not name. The record names the reporting text, not the people.
Every record
| Source date | Predictor | Predicted event | Horizon as stated | Probability as stated | Form |
|---|---|---|---|---|---|
| 1950-10 | Alan Turing | it will be possible, to programme computers, with a storage capacity of about 109, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identificationComputing Machinery and Intelligence | in about fifty years' time | Not stated | Stated forecast |
| 1950-10 | Alan Turing | the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradictedComputing Machinery and Intelligence | at the end of the century | Not stated | Stated forecast |
| 1955-08-31 | John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon (the proposal's signatories) | a significant advance can be made in one or more of these problemsA Proposal for the Dartmouth Summer Research Project on Artificial Intelligence | for a summer (a "2 month, 10 man study" during the summer of 1956) | Not stated | Stated forecast |
| 1958 | Herbert A. Simon and Allen Newell | within ten years a digital computer will be the world's chess champion, unless the rules bar it from competitionHeuristic Problem Solving: The Next Advance in Operations Research | within ten years | Not stated | Stated forecast |
| 1958 | Herbert A. Simon and Allen Newell | a digital computer will discover and prove an important new mathematical theoremHeuristic Problem Solving: The Next Advance in Operations Research | within ten years | Not stated | Stated forecast |
| 1958 | Herbert A. Simon and Allen Newell | a digital computer will write music that will be accepted by critics as possessing considerable aesthetic valueHeuristic Problem Solving: The Next Advance in Operations Research | within ten years | Not stated | Stated forecast |
| 1958 | Herbert A. Simon and Allen Newell | most theories in psychology will take the form of computer programs, or of qualitative statements about the characteristics of computer programsHeuristic Problem Solving: The Next Advance in Operations Research | within ten years | Not stated | Stated forecast |
| 1965 | I. J. Good | an ultraintelligent machine will be built and that it will be the last invention that man need makeSpeculations Concerning the First Ultraintelligent Machine | within the twentieth century | "more probable than not" | Stated forecast |
| 1973Wording attested 1986 | Opinion poll reported by Donald Michie, "Machines and the theory of intelligence" (Nature, 1973) | computing system exhibiting intelligence at adult human levelMachines and the theory of intelligence | Given only in Fig. 11.4, as the frequency of answers over "Estimated number of years before" the event, in bands from 10 to more than 50 years | Not stated | Expert survey |
| 1973Reported statement | Unnamed correspondents ("able and respected scientists") | an all-purpose intelligence on a human-scale knowledge baseArtificial Intelligence: A General Survey | in the nineteen-eighties | Not stated | Reported statement |
| 1973Reported statement | Unnamed correspondents ("able and respected scientists") | machine intelligence exceeding human intelligenceArtificial Intelligence: A General Survey | by the year 2000What the dataset records for 2000 → | Not stated | Reported statement |
| 1983-01 | Vernor Vinge | superhuman intelligences are createdFirst Word | between 20 and 100 years from now | Not stated | Stated forecast |
| 1983-01 | Vernor Vinge | this will changeFirst Word | During the next 20 years | Not stated | Stated forecast |
| 1989 | Hugo de Garis | it will be commonplace to say that the twenty-first century will be dominated by the machine if humanity so chooses—and maybe even if notWhat If AI Succeeds? The Rise of the Twenty-First Century Artilect | In 20 years time | Not stated | Stated forecast |
| 1993-12 | Vernor Vinge | the creation of greater than human intelligence will occurThe Coming Technological Singularity: How to Survive in the Post-Human Era | during the next thirty years | Not stated | Stated forecast |
| 1993-12 | Vernor Vinge | this event occursThe Coming Technological Singularity: How to Survive in the Post-Human Era | I'll be surprised if this event occurs before 2005 or after 2030What the dataset records for 2005 → | Not stated | Stated forecast |
| 1998Wording attested 2026-11-02 | Nick Bostrom | we will have superhuman artificial intelligenceHow Long Before Superintelligence? | within the first third of the next century | Not stated | Stated forecast |
| 1998Wording attested 2000-12-10 | Nick Bostrom | the required computational power will be reachedHow Long Before Superintelligence? | sometime between 2004 and 2008, depending on whether we assume a doubling time of 12 or 18 monthsWhat the dataset records for 2004 →What the dataset records for 2008 → | Not stated | Stated forecast |
| 1998 | Hans Moravec | the required hardware will be available in cheap machinesWhen will computer hardware match the human brain? | in the 2020s | Not stated | Stated forecast |
| 1999-12 | Hans Moravec | robot intelligence will surpass our ownRise of the Robots | well before 2050 | Not stated | Stated forecast |
| 2000-08-31 | Eliezer Yudkowsky | the Singularity (paraphrase)Staring into the Singularity | 2035 "seems like a reasonable upper bound"; "The lower bound? Thirty seconds." | Not stated | Stated forecast |
| 2001-01Wording attested 2001-08-04 | Ray Kurzweil | supercomputers will achieve one human brain capacity by 2010, and personal computers will do so by around 2020The Law of Accelerating Returns | by 2010 (supercomputers); by around 2020 (personal computers)What the dataset records for 2010 →What the dataset records for 2020 → | Not stated | Stated forecast |
| 2001-01Wording attested 2026-11-02 | Ray Kurzweil | We achieve one Human Brain capability (2 * 10^16 cps) for $1,000The Law of Accelerating Returns | around the year 2023What the dataset records for 2023 → | Not stated | Stated forecast |
| 2002Wording attested 2026-11-02 | Mitchell Kapor (predictor) and Ray Kurzweil (challenger), on Long Bets | By 2029 no computer - or "machine intelligence" - will have passed the Turing Test.Long Bet 1: By 2029 no computer - or "machine intelligence" - will have passed the Turing Test | By 2029 | Not stated | Public bet |
| 2009-12-07Wording attested 2009-12-13 | Shane Legg | roughly human level AGI (paraphrase)Tick, tock, tick, tock… BING | mode "about 2025"; expected value "a bit higher at 2028"; 90% credibility region "2018 to 2036"What the dataset records for 2018 →What the dataset records for 2025 → | mode; expected value; "90% credibility region" | Stated forecast |
| 2010-12-10Reported later · estimate of 2009-12Wording attested 2010-12-14 | Shane Legg | roughly human level AGI (paraphrase)Goodbye 2010 | log-normal distribution with a mean of 2028, and a mode of 2025What the dataset records for 2025 → | mean; mode | Stated forecast |
| 2011 | Baum, Goertzel and Goertzel, expert assessment at AGI-09 | human-level AI to be reachedHow long until human-level AI? Results from an expert assessment | within upcoming decades | Not stated | Expert survey |
| 2011Wording attested 2016-12-24 | Sandberg and Bostrom, informal poll at the FHI Winter Intelligence conference (FHI Technical Report 2011-1) | HLMIMachine Intelligence Survey | median estimates: 10% chance of HLMI 2028; 50% chance 2050; 90% chance 2150 | 10%; 50%; 90% | Expert survey |
| 2011-06-17 | Shane Legg | the development of human-level machine intelligenceQ&A with Shane Legg on risks from AI | 10%: 2018; 50%: 2028; 90%: 2050What the dataset records for 2018 → | 10%, 50% and 90% (the question's levels) | Stated forecast |
| 2011-12-31Wording attested 2012-04-06 | Shane Legg | human level AGI (paraphrase)Goodbye 2011, hello 2012 | log-normal distribution with a mean of 2028 and a mode of 2025What the dataset records for 2025 → | mean; mode (of a log-normal distribution) | Stated forecast |
| 2014 | Müller and Bostrom, survey of expert opinion | high-level machine intelligence will be developedFuture Progress in Artificial Intelligence: A Survey of Expert Opinion | 50%: around 2040-2050; 90%: by 2075 (median estimates) | "a one in two chance"; "a nine in ten chance" | Expert survey |
| 2014 | Müller and Bostrom, survey of expert opinion | systems will move on to superintelligenceFuture Progress in Artificial Intelligence: A Survey of Expert Opinion | in less than 30 years thereafter (after high-level machine intelligence); the question asked for probabilities "within (2 years / 30 years) thereafter" | Not in the abstract. The body reports, for all respondents, a median of 75% within 30 years and 10% within 2 years (means 62% and 19%), in a table (PDF p. 12), and summarises it as "a significant probability" | Expert survey |
| 2017-05-24 | Grace et al., survey of machine learning researchers (2016) | AI outperforming humans in all tasksWhen Will AI Exceed Human Performance? Evidence from AI Experts | 50% chance in 45 years | 50% | Expert survey |
| 2017-05-24 | Grace et al., survey of machine learning researchers (2016) | automating all human jobsWhen Will AI Exceed Human Performance? Evidence from AI Experts | 50% chance in 120 years | 50% | Expert survey |
| 2017-05-24 | Grace et al., survey of machine learning researchers (2016) | AI will outperform humans in many activitiesWhen Will AI Exceed Human Performance? Evidence from AI Experts | in the next ten years (examples: translating languages by 2024, writing high-school essays by 2026, driving a truck by 2027) | Not stated | Expert survey |
| 2018-01-01Wording attested 2018-01-02 | Rodney Brooks | Emergence of the generally agreed upon "next big thing" in AI beyond deep learning.My Dated Predictions | NET 2023, BY 2027 ("No Earlier Than" 2023, "By" 2027)What the dataset records for 2023 → | Not stated | Stated forecast |
| 2018-01-01Wording attested 2018-01-02 | Rodney Brooks | An AI system with an ongoing existence (no day is the repeat of another day as it currently is for all AI systems) at the level of a mouse.My Dated Predictions | NET 2030 ("No Earlier Than") | Not stated | Stated forecast |
| 2018-01-01Wording attested 2018-01-02 | Rodney Brooks | A robot that seems as intelligent, as attentive, and as faithful, as a dog.My Dated Predictions | NET 2048 ("No Earlier Than") | Not stated | Stated forecast |
| 2021-09-07 | Holden Karnofsky | transformative AIAI Timelines: Where the Arguments, and the "Experts," Stand | more than 10% within 15 years (by 2036); ~50% within 40 years (by 2060); ~2/3 this century (by 2100) | more than a 10% chance; a ~50% chance; a ~2/3 chance | Stated forecast |
| 2022-06-08 | Zhang et al., survey of AI/ML researchers (2019) | human-level machine intelligence being achievedForecasting AI Progress: Evidence from a Survey of Machine Learning Researchers | 50% likelihood by 2060 | 50% likelihood | Expert survey |
| 2022-06-16 | Joseph Carlsmith | an existential catastrophe of this kind will occurIs Power-Seeking AI an Existential Risk? | by 2070 | ~5% (April 2021); ">10%" (May 2022 update) | Stated forecast |
| 2022-08Wording attested 2022-08-04 | AI Impacts, 2022 Expert Survey on Progress in AI | HLMI2022 Expert Survey on Progress in AI | aggregate forecast time to a 50% chance: 37 years, i.e. 2059 | 50% chance | Expert survey |
| 2022-08-02Reported later · estimate of 2020-09 | Ajeya Cotra | transformative AITwo-year update on my personal AI timelines | ~15% by 2036; a median of ~2050 | ~15%; median | Stated forecast |
| 2022-08-02 | Ajeya Cotra | transformative AITwo-year update on my personal AI timelines | ~15% by 2030; ~35% by 2036; a median of ~2040; ~60% by 2050 | ~15%; ~35%; median; ~60% | Stated forecast |
| 2023-05-22 | Sam Altman, Greg Brockman and Ilya Sutskever (OpenAI) | AI systems will exceed expert skill level in most domainsGovernance of superintelligence | within the next ten years | Not stated | Stated forecast |
| 2023-07-05 | OpenAI (post by Jan Leike and Ilya Sutskever) | it could arriveIntroducing Superalignment | this decade | Not stated | Stated forecast |
| 2023-10-26 | Yoshua Bengio, Geoffrey Hinton, Andrew Yao and co-authors | generalist AI systems will outperform human abilities across many critical domainsManaging AI Risks in an Era of Rapid Progress | within this decade or the next | Not stated | Stated forecast |
| 2024-01-05 | Grace et al., survey of AI authors (2023) | unaided machines outperforming humans in every possible taskThousands of AI Authors on the Future of AI | 10% by 2027, and 50% by 2047 | 10%; 50% | Expert survey |
| 2024-01-05 | Grace et al., survey of AI authors (2023) | all human occupations becoming fully automatableThousands of AI Authors on the Future of AI | 10% by 2037, and 50% as late as 2116 | 10%; 50% | Expert survey |
| 2024-06 | Leopold Aschenbrenner | AGISituational Awareness: The Decade Ahead | by 2027 | Not stated | Stated forecast |
| 2024-09-23 | Sam Altman | we will have superintelligenceThe Intelligence Age | in a few thousand days (!); it may take longer | Not stated | Stated forecast |
| 2024-10 | Dario Amodei | powerful AIMachines of Loving Grace | as early as 2026, though there are also ways it could take much longerWhat the dataset records for 2026 → | Not stated | Stated forecast |
| 2025-01-05 | Sam Altman | we may see the first AI agents “join the workforce” and materially change the output of companiesReflections | in 2025What the dataset records for 2025 → | Not stated | Stated forecast |
| 2025-01-05 | Metaculus community prediction, question 5121 | When will the first general AI system be devised, tested, and publicly announced?When will the first general AI system be devised, tested, and publicly announced? (Metaculus question 5121, as captured on 2025-01-05) | 2031-06-14 (the community prediction as displayed; the community table's "median", between "lower 25%" 2027-05-10 and "upper 75%" 2041-08-31) | median (community table) | Forecasting platform |
| 2025-02-09Forecast or aim | Sam Altman | Anyone in 2035 should be able to marshall the intellectual capacity equivalent to everyone in 2025Three Observations | in 2035 | Not stated | Stated forecast |
| 2025-03-06 | Anthropic | powerful AI systems will emergeAnthropic's Recommendations to OSTP for the U.S. AI Action Plan | in late 2026 or early 2027 | Not stated | Stated forecast |
| 2025-04 | Dario Amodei | we could have AI systems equivalent to a “country of geniuses in a datacenter”The Urgency of Interpretability | as soon as 2026 or 2027What the dataset records for 2026 → | Not stated | Stated forecast |
| 2025-04-02 | Google DeepMind (paper by Rohin Shah and co-authors) | powerful AI systems are developedAn Approach to Technical AGI Safety and Security | by 2030 | Not stated | Stated forecast |
| 2025-04-03 | Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean (AI 2027) | the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial RevolutionAI 2027 | over the next decade | Not stated | Stated forecast |
| 2025-06-10 | Sam Altman | the arrival of systems that can figure out novel insightsThe Gentle Singularity | 2026What the dataset records for 2026 → | Not stated | Stated forecast |
| 2025-06-10 | Sam Altman | the arrival of robots that can do tasks in the real worldThe Gentle Singularity | 2027 | Not stated | Stated forecast |
| 2025-11-06Wording attested 2025-11-07 | OpenAI | AI to be capable of making very small discoveriesAI progress and recommendations | In 2026What the dataset records for 2026 → | Not stated | Stated forecast |
| 2025-11-06Wording attested 2025-11-07 | OpenAI | we will have systems that can make more significant discoveriesAI progress and recommendations | In 2028 and beyond | Not stated | Stated forecast |
| 2026-01 | Dario Amodei | powerful AIThe Adolescence of Technology | as little as 1–2 years away, although it could also be considerably further out | Not stated | Stated forecast |
| 2026-01 | Dario Amodei | AI is better than humans at essentially everythingThe Adolescence of Technology | it cannot possibly be more than a few years before | Not stated | Stated forecast |
Marks in the table
- Forecast or aim
- The sentence can be read as a forecast or as an aim, as with “should be able to”.
- Reported later
- A later text’s report of an earlier estimate. The record is dated by the later text; the earlier estimate’s date is shown beside it.
- Reported statement
- The predictors are not named; the record rests on the text that reports them.
- Wording attested from
- The quoted wording is known only from a later copy or from the page as read live. The date from which the wording is attested is shown.
Limits
- The records come from targeted searches, not from a census of predictions. A prediction missing from the table was not recorded; that says nothing about whether it was made.
- Surveys differ in their questions, definitions and respondents. Their figures are shown as each report words them and are not put on a common basis.
- Where a text gives several probabilities or dates, the record lists each one with its own horizon. On the chart, a thin line joins the printed years of one record.
- Some pages were read as they stand today. Their wording is attested only from the date they were read, which is shown.
- A printed target year can link to the dataset’s page for that year. That page lists what the dataset records for the year. It is not an outcome of any prediction.
How we searched
Predictions about machine intelligence that give a target date or a horizon, in texts dated from 1950 to 2026: forecasts in papers, essays, posts and announcements; published results of expert surveys; one public bet; one forecasting platform’s figure from a dated copy; and one report of unnamed correspondents’ predictions. A passage with no horizon was not recorded.
Methods
- Texts the dataset already holds were read first; gaps were then filled with searches of authors’ sites, survey reports, the bet’s page, laboratory publications and Internet Archive copies. Searches ran from 2026-11-02 to 2026-11-05.
- Every quotation was matched mechanically against the retrieved text, or against the dataset’s own checked passage, or read on a page image where the copy has no text layer. A fresh retrieval at admission on 2026-11-05 matched again every quotation whose source could be retrieved.
- Copies the preparing assistant could not reach (Internet Archive copies, some publisher pages) were read by a model from a different family, OpenAI’s GPT; each record says whose reading it rests on.
- A fresh-context audit by a Claude model re-read the records. Two checks by an OpenAI model followed, each answering blind before it compared its answers with the records. The maintainer decided the open questions on 2026-11-04 and 2026-11-05.
- Where a source’s length is not known, or a quotation would pass the source’s quotation limit, the passage is a labelled paraphrase with the text’s own words in quotation marks.
Gaps
- Books that could not be read in full were not used: Kurzweil’s books of 1999 and 2005, Moravec’s of 1988 and 1999, Minsky’s of 1967.
- Posts on social media, video and audio without a published transcript, and press interviews were not used.
- Of sixteen dated copies of two forecasting-platform questions that were read, one shows a readable community figure; only that one is recorded.
- Kurzweil’s essay “The Law of Accelerating Returns” is dated January 2001 on the page read; the dataset dates it 7 March 2001. No copy read settles the date.
- A search that finds nothing does not show that nothing exists.
Read and not recorded
- Simon (1965), Minsky (1967), Michie in Nature (1973), and Kurzweil’s books of 1999 and 2005. Neither model family could read the texts. They remain leads.
- Life, 20 November 1970, reporting Minsky. Only a search-index excerpt could be read. The record waits until the page itself is read, with Grudin’s later report beside it that Minsky said he was misquoted.
- Moravec, “The Age of Robots” (1993). Its section headings carry dates, but a heading frames a scenario; it is not a sentence stating an event by a date.
- Long Bet 1 as worded in 2002. An archive copy of 2002 words the bet differently and calls the parties “initiator” and “responder” without saying who predicted which side. The record keeps the wording of the page as read, with the 2002 wording in its qualification.