CoChat
Sign Up
Cited by 202Ethics and Social Impacts of AIArtificial Intelligence ApplicationsImpact of AI and Big Data on Business and Society

The DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; 1 of the 30 sources its claims cite were checked for support.

Thousands of AI Authors on the Future of AI

Open access
Grace, Katja, Sandkühler, Julia Fabienne, Stewart, Harlan, Weinstein-Raun, Benjamin, Thomas, Stephen, Stein-Perlman, Zach, Salvatier, John, Brauner, Jan, Korzekwa, Richard C.·2025·Journal of Artificial Intelligence Research
DOI 10.1613/jair.1.19087
Open this paper in CoChat — free accountOpen PDF

Abstract

In October 2023, 2,778 researchers who had published in top-tier artificial intelligence (AI) venues gave predictions on the pace, nature and impacts of AI progress. Significant steps were taken to minimize and evaluate bias. In evaluations of participation bias, we found that most groups responded at similar rates. The participants estimated that several milestones had at least a 50% chance of being feasible for AI by 2028, including constructing a payment processing site and fine-tuning an LLM. If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027 and 50% by 2047—13 years earlier than in our 2022 survey (N = 738). The chance of all occupations becoming fully automatable, however, was not expected to reach 10% until 2037, and 50% until 2116 (compared to 2164 in the 2022 survey. Most respondents expressed substantial uncertainty about long-term impacts: While 68% in 2023 thought good outcomes from high-level machine intelligence AI were more likely than bad ones, 48% of these net optimists gave at least a 5% chance of extremely bad outcomes. Conversely, 59% of net pessimists gave 5% or more to extremely good outcomes. Depending on how we asked, between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction. More than half suggested that “substantial” or “extreme” concern is warranted about AI increasing misinformation, boosting authoritarian control, worsening inequality, and other scenarios. There was broad agreement that research aimed at minimizing risks from AI systems ought to be more prioritized.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

In the largest survey of its kind, 2,778 researchers who published at top AI venues in 2022 gave forecasts about AI progress and impacts. Aggregate forecasts put a 50% chance of high-level machine intelligence by 2047 and full automation of labor by 2116, while many respondents assigned nontrivial probability to extreme bad outcomes.

Key findings

  • Aggregate forecasts gave at least a 50% chance of AI systems achieving several milestones by 2028, including autonomously constructing a payment processing site from scratch, creating a song indistinguishable from a new song by a popular musician, and autonomously downloading and fine-tuning a large language model.
  • The chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027 and 50% by 2047, 13 years earlier than the 2022 survey; full automation of labor was forecast at 10% by 2037 and 50% by 2116, compared with 2164 in 2022.
  • 68.3% thought good outcomes from superhuman AI are more likely than bad, but 48% of net optimists gave at least a 5% chance of extremely bad outcomes and 59% of net pessimists gave at least 5% to extremely good outcomes; between 37.8% and 51.4% gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction.
  • More than half of respondents said “substantial” or “extreme” concern was warranted about six AI-related scenarios, including spread of false information, authoritarian population control, and worsened inequality; 70% thought AI safety research should be prioritized more than it currently is.

Methods

  • The 2023 Expert Survey on Progress in AI was fielded in October 2023 to researchers who published in 2022 at six top AI venues: NeurIPS, ICML, ICLR, AAAI, IJCAI, and JMLR.
  • From approximately 21,800 collected names, 20,066 emails were found; after 1,607 bounces, 18,459 functioning addresses yielded 2,778 responses, a 15% response rate.
  • Questions used Likert scales, probability estimates, or future-year estimates; milestone questions were randomized between fixed-years and fixed-probabilities framings, and individual probability distributions were approximated by fitting gamma cumulative distribution functions.

Limitations

  • The paper states that forecasting is hard even for experts, that AI researchers are not necessarily skilled forecasters, and that substantial framing effects were observed, suggesting aggregate answers to any particular question may not be an accurate guide.
  • The paper states possible participation bias, though it reports no evidence of strong bias; the survey had a 15% response rate, and the 2023 sample expanded from two to six venues, which could make year-to-year comparisons less meaningful.
  • The text makes evident that most questions were answered by randomized subsets rather than all 2,778 participants, and that some question wording differences, such as HLMI versus FAOL, may have produced framing effects.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1613/jair.1.19087

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

1 of 30 cited sources checked · 1 supported · 0 flagged

  • Findingown data

    Each of the eleven potentially concerning AI scenarios was considered worthy of either substantial or extreme concern by more than 30% of respondents.

    “Each scenario was considered worthy of either substantial or extreme concern by more than 30% of respondents.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.
  • Findingcitation

    The 50% FAOL estimate of 2116 is compared to 2164 in the 2022 survey.

    “However, the chance of all human occupations becoming fully automatable was forecast to reach 10% by 2037, and 50% as late as 2116 (compared to 2164 in the 2022 survey).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.
  • Findingown data

    If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 50% by 2047.

    “If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047.”
    • Also stated in: § 3.2.1 How soon will 'High-Level Machine Intelligence' be feasible? · hedged
      “The aggregate 2023 forecast predicted a 50% chance of HLMI by 2047, down thirteen years from 2060 in the 2022 survey.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.
  • Findingown data

    There was a wide range of views about each input, implying a large degree of uncertainty.

    “There was a wide range of views about each input, implying a large degree of uncertainty.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.
  • Findingown data

    This year's survey saw a general shift towards earlier expectations for milestones.

    “While the range of views on how long it will take for milestones to be feasible can be broad, this year's survey saw a general shift towards earlier expectations.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.
  • It may state two separate claims as one. 39
  • Probably not a substantial claim. 23
  • The quote only partly states this claim. 1

  • It may state two separate claims as one.

    Premisecitation

    Forecasting is difficult in general, and subject-matter experts have been observed to perform poorly.

    “Forecasting is difficult in general, and subject-matter experts have been observed to perform poorly [Tetlock, 2005, Savage et al., 2021].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Tetlock, 2005

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Savage et al., 2021 source

  • It may state two separate claims as one.

    Findingown data

    The median probability that AI is vastly better than humans 30 years after HLMI was 50% in 2016 and 60% in 2022 and 2023.

    “AI is vastly better than humans 30 years after HLMI 50% (214) 60% ( 371) 60% ( 282)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    As in 2022, a majority of participants considered AI to pose at least a 5% chance of causing human extinction or similarly permanent and severe disempowerment of the human species, and this result was consistent across four different questions, two assigned to each participant.

    “As in 2022, a majority of participants considered AI to pose at least a 5% chance of causing human extinction or similarly permanent and severe disempowerment of the human species, and this result was consistent across four different questions, two assigned to each participant.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.
    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    The proportion of people who put at least a 10% chance on extremely bad outcomes (e.g. human extinction) has fallen from 48% in 2022 to 2023, and the mean prediction for this type of outcome is down from 14% to 9.0%.

    “The proportion of people who put at least a 10% chance on extremely bad outcomes (e.g.human extinction) has fallen from 48% in 2022 in 2023, and the mean prediction for this type of outcome is down from 14% to 9.0%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    There was disagreement on whether faster or slower progress would be preferable, though large divergence from the current speed was less popular.

    “There was disagreement on whether faster or slower progress would be preferable, though large divergence from the current speed was less popular (Table 3).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Most respondents considered it unlikely that users of AI systems in 2028 will be able to know the true reasons for the AI systems' choices, with only 20% giving it better than even odds.

    “Most respondents considered it unlikely that users of AI systems in 2028 will be able to know the true reasons for the AI systems' choices, with only 20% giving it better than even odds.(n=912)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Compared to 2016, 2023 has earlier 50% estimates but later 10% estimates for FAOL.

    “Compared to 2016, 2023 has earlier 50% estimates but later 10% estimates.(See Appendix B)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2018a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    Participants estimated that halving the drop in costs of computing would have had the greatest effect on AI progress over the last decade, while halving 'researcher effort' and 'progress in AI algorithms' would have had the least effect.

    “Participants estimated that halving the drop in costs of computing would have had the greatest effect on AI progress over the last decade, while halving 'researcher effort' and 'progress in AI algorithms' would have had the least effect.”
    • Also stated in: § E Links to Data · hedged
      “Participants estimated that halving the drop in costs of computing would have had the greatest effect on AI progress over the last decade, while halving 'researcher effort' and 'progress in AI algorithms' would have had the least effect.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    The 38% figure is comparable to, but somewhat lower than, rates of assigning at least 10% to extinction-level outcomes in answers to other questions more directly about extinction, between 41% and 51%.

    “This is comparable to, but somewhat lower than, rates of assigning at least 10% to extinction-level outcomes in answers to other questions more directly about extinction, between 41% and 51% (see Section 4.3).”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Answers to the three similar questions about human extinction were mostly consistent, with medians of 5% or 10%.

    “Answers to these questions were mostly consistent, with medians of 5% or 10%.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Even the larger of the two framing estimates (34 years) is shorter than 2022's combined forecast (37 years), demonstrating a substantial shift of predictions closer to the present.

    “However, it's notable that even the larger of these two is shorter than 2022's combined forecast (37 years), demonstrating a substantial shift of predictions closer to the present (Fig 18 in Appendix B).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    The scenarios worthy of most concern, as measured by percentage of respondents who thought a scenario constituted either a 'substantial' or 'extreme' concern, were: spread of false information e.g. deepfakes (86%), manipulation of large-scale public opinion trends (79%), AI letting dangerous groups make powerful tools (e.g. engineered viruses) (73%), authoritarian rulers using AI to control their

    “the scenarios worthy of most concern were: spread of false information e.g.deepfakes (86%), manipulation of large-scale public opinion trends (79%), AI letting dangerous groups make powerful tools (e.g.engineered viruses) (73%), authoritarian rulers using AI to control their populations (73%), and AI systems worsening economic inequality by disproportionately benefiting certain individuals (71%).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Respondents exhibited diverse views on the future impact of advanced AI, highlighting how respondents' views are more complex than can be represented by an 'optimism vs pessimism' axis.

    “Respondents exhibited diverse views on the future impact of advanced AI (Figure 10), highlighting how respondents' views are more complex than can be represented by an 'optimism vs pessimism' axis.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Participants who had thought more in the past about AI risks seem to give higher numbers, suggesting they are unlikely to give radically lower numbers with further examples of risks.

    “Participants who had thought more in the past about AI risks seem to give higher numbers, suggesting they are unlikely to give radically lower numbers with further examples of risks (see Table 4).”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingcitation

    The aggregate 2023 forecast predicted a 10% chance of HLMI by 2027, down two years from 2029 in the 2022 survey.

    “The aggregate 2023 forecast predicted a 10% chance of HLMI by 2027, down two years from 2029 in the 2022 survey.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    In general, there were a wide range of views about expected social consequences of advanced AI, and most people put some weight on both extremely good outcomes and extremely bad outcomes.

    “In general, there were a wide range of views about expected social consequences of advanced AI, and most people put some weight on both extremely good outcomes and extremely bad outcomes.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingreasoning

    On many questions different respondents give very different answers, which limits the number of them who can be close to the truth.

    “For one thing, on many questions different respondents give very different answers, which limits the number of them who can be close to the truth.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingown data

    Participants expressed a wide range of views on almost every question: some of the biggest areas of consensus are on how wide-open possibilities for the future appear to be.

    “Participants expressed a wide range of views on almost every question: some of the biggest areas of consensus are on how wide-open possibilities for the future appear to be.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    Over the fourteen months since the last survey, a similar participant pool expected human-level performance 13 to 48 years sooner on average (depending on how the question was phrased), and 21 out of 32 shorter term milestones are now expected earlier.

    “Over the fourteen months since the last survey [Grace et al., 2022], a similar participant pool expected human-level performance 13 to 48 years sooner on average (depending on how the question was phrased), and 21 out of 32 shorter term milestones are now expected earlier.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    The percentage of researchers who thought AI safety research should be prioritized more increased compared to earlier surveys, but only slightly since 2022.

    “The percentage of researchers who thought so increased compared to earlier surveys, but only slightly since 2022.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    The aggregate forecasts give at least a 50% chance of AI systems achieving several milestones by 2028, including autonomously constructing a payment processing site from scratch, creating a song indistinguishable from a new song by a popular musician, and autonomously downloading and fine-tuning a large language model.

    “The aggregate forecasts give at least a 50% chance of AI systems achieving several milestones by 2028, including autonomously constructing a payment processing site from scratch, creating a song indistinguishable from a new song by a popular musician, and autonomously downloading and fine-tuning a large language model.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The surveyed AI researchers are representative of AI researchers as a whole.

  • It may state two separate claims as one.

    Findingown data

    The aggregate 2023 forecast predicted a 50% chance of FAOL by 2116, down 48 years from 2164 in the 2022 survey.

    “The aggregate 2023 forecast predicted a 50% chance of FAOL by 2116, down 48 years from 2164 in the 2022 survey (Figure 3).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • Rests on an assumption: The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

  • It may state two separate claims as one.

    Findingown data

    In the first framing, the median prediction for 2 years post-HLMI was 20%, whereas the median prediction for 30 years post-HLMI was 80%.

    “In the first framing, the median prediction for 2 years post-HLMI was 20%, whereas the median prediction for 30 years post-HLMI was 80%.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    95% of these responses were deemed 'finished' by Qualtrics.

    “95% of these responses were deemed 'finished' by Qualtrics.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingnone given

    Small code changes and improvements to data cleaning between the 2022 and 2023 surveys shifted the aggregate slightly.

    “small code changes and improvements to data cleaning between surveys shifted the aggregate slightly.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

    Findingnone given

    There is still a lot of uncertainty.

    “However, there is still a lot of uncertainty.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    In the second framing, the median prediction for 2 years post-HLMI was 10%, whereas the median prediction for 30 years post-HLMI was 60%.

    “In the second framing, the median prediction for 2 years post-HLMI was 10%, whereas the median prediction for 30 years post-HLMI was 60%.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingown data

    In total, 196 participants gave answers with decreasing probabilities, and 7 of these users gave answers summing to 100% for fixed time questions.

    “In total, 196 participants gave answers with decreasing probabilities, and 7 of these users gave answers summing to 100% for fixed time questions.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Out of the 20,066 emails contacted, 1,607 (8%) bounced or failed, leaving 18,459 functioning email addresses.

    “Out of the 20,066 emails we contacted, 1,607 (8%) bounced or failed, leaving 18,459 functioning email addresses.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingown data

    The authors received 2,778 responses, for a response rate of 15%.

    “We received 2,778 responses, for a response rate of 15%.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Whether participants worked in academia or industry did not affect median responses much-both had a median of 5%.

    “Whether participants worked in academia or industry did not affect median responses much-both had a median of 5%.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data

    Developments since the 2022 survey have not substantially changed the proportion of participants who think AI safety should be prioritized more or much more.

    “Developments since the 2022 survey have not substantially changed the proportion of participants who think AI safety should be prioritized”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    The region respondents graduated from affected responses somewhat: the Asian median was 10%, while North American and European medians were 5%.

    “The region respondents graduated from affected responses somewhat: the Asian median was 10%, while North American and European medians were 5%.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    People who had thought more about the social impacts of smarter-than-human machines were substantially less likely to give more than 10% credence to extremely optimistic outcomes, and substantially more likely to give extremely negative outcomes more than 10% credence.

    “People who had thought more about the social impacts of smarter-than-human machines were substantially less likely to give more than than 10% credence to extremely optimistic outcomes, and substantially more likely to give extremely negative outcomes more than 10% credence.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    The question which was skipped the least often was about the long-term value of HLMI, answered by 100% of those who saw it.

    “The question which was skipped the least often was about the long-term value of HLMI (see Section 4.2), answered by 100% of those who saw it, probably because it was not possible to skip and continue the survey.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    Achieving HLMI seems to either imply having already achieved FAOL, or suggest being close.

    “Since occupations might naturally be understood either as complex tasks, composed of tasks, or closely connected with one of these, achieving HLMI seems to either imply having already achieved FAOL, or suggest being close.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Study design and conductown data

    A Welch t-test found that the difference between the two framings of the AI safety research prioritization question was not significant (t(1327) = -0.58, p = 0.564, d = -0.03), so the results were combined (n=1329).

    “A Welch t-test found that the difference between the two framings was not significant (t(1327) = -0.58,p = 0.564, d = -0.03),so the results were combined (n=1329).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Interpretationreasoning

    Insofar as HLMI and FAOL refer to the same event, the difference in predictions about the time of their arrival would seem to be a framing effect.

    “Insofar as HLMI and FAOL refer to the same event, the difference in predictions about the time of their arrival would seem to be a framing effect.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Premisecitation

    In a closely related study, Karger et al. [2023] found college graduates gave answers nearly six orders of magnitude apart when asked in different ways to estimate the size of existential risks from AI; when given example odds of low-probability events, estimates were much lower.

    “In an extreme example in a closely related study, Karger et al. [2023] found college graduates gave answers nearly six orders of magnitude apart when asked in different ways to estimate the size of existential risks from AI: When given example odds of low-probability events, estimates were much lower.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Karger et al. [2023]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    In other contexts, averages from a large set of noisy predictions can still be relatively accurate.

    “Nonetheless, in other contexts, averages from a large set of noisy predictions can still be relatively accurate [Surowiecki, 2004], so a question remains as to how informative these aggregate forecasts are.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Surowiecki, 2004

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Interpretationown data

    This may reflect a view amongst AI researchers that the general public misunderstands AI.

    “This may reflect a view amongst AI researchers that the general public misunderstands AI.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationown data

    In sum, across the three intelligence explosion questions, the median participant did not overall expect something like a rapid acceleration of progress from an 'intelligence explosion,' but did give substantial credence to it.

    “In sum, across these three questions, the median participant did not overall expect something like a rapid acceleration of progress from an 'intelligence explosion,' but did give substantial credence to it.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.

    Interpretationreasoning

    AI experts' predictions should not be seen as a reliable guide to objective truth, but can provide one important piece of the puzzle.

    “While AI experts' predictions should not be seen as a reliable guide to objective truth, they can provide one important piece of the puzzle.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.

    Study design and conductnone given

    The HLMI-framing used 40 years instead of 50 years to keep the survey consistent with previous surveys, where this discrepancy was introduced by mistake.

    “The HLMI-framing used 40 years instead of 50 years.This was done to keep the survey consistent with the previous surveys, where this discrepancy was introduced by mistake.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductcitation

    The survey and its analysis were preregistered.

    “The survey and its analysis were preregistered [Sandkühler et al., 2023].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Sandkühler et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Study design and conductown data

    The paper looked for evidence of response bias at the survey level and question level for some questions, and did not find any that would affect the results to a large extent.

    “We looked for evidence of response bias at the survey level and question level for some questions, and did not find any that would affect the results to a large extent (see Appendix D for more detail).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductnone given

    The survey and its implementation were approved by the Ethics Committee of the University of Bonn (248/23-EP).

    “The survey and its implementation were approved by the Ethics Committee of the University of Bonn (248/23-EP).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductcitation

    As in previous surveys, most questions were randomly assigned to only a subset of participants, in order to keep the number of questions for each participant low.

    “As in Grace et al. [2018a] and Grace et al. [2022], most questions were randomly assigned to only a subset of participants, in order to keep the number of questions for each participant low.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds.

    Interpretationreasoning

    AI researchers' familiarity with the relevant technology, and experience with the dynamics of its progress, make them among the best-positioned to make informative educated guesses.

    “However, AI researchers' familiarity with the relevant technology, and experience with the dynamics of its progress, make them among the best-positioned to make informative educated guesses.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The survey participants do not have any unusual skill at forecasting in general.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductnone given

    While designing new questions, the authors tested them in a series of interviews with AI researchers and students.

    “While designing new questions, we tested them in a series of interviews with AI researchers and students.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductnone given

    Any individual respondent did not see more than one of these questions.

    “Any individual respondent did not see more than one of these questions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductcitation

    Most questions were identical to those asked previously in surveys conducted in 2016 and 2022.

    “Most questions were identical to those asked previously in surveys conducted in 2016 [Grace et al., 2018a] and 2022 [Grace et al., 2022].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Grace et al., 2018a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Grace et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.

    Interpretationreasoning

    While unreliable, educated guesses are what we must all rely on, and AI researchers' guesses are informed by expertise in the relevant field.

    “While unreliable, educated guesses are what we must all rely on, and theirs are informed by expertise in the relevant field.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The survey participants do not have any unusual skill at forecasting in general.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductown data

    The number of people who answered each of the eleven questions was between 649 and 667.

    “i.e. the number of people (n) who answered each of these eleven questions was between 649 and 667”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Study design and conductown data

    For each of two questions, the authors added a new version with modified phrasing and randomly assigned participants either the version of the question asked in Grace et al. [2022] or the new version.

    “For each of two questions, we added a new version with modified phrasing and randomly assigned participants either the version of the question asked in Grace et al. [2022] or the new version.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence.

    Interpretationreasoning

    Despite these limitations, AI researchers are well-positioned to contribute to the accuracy of our collective guesses about the future.

    “Despite these limitations, AI researchers are well-positioned to contribute to the accuracy of our collective guesses about the future.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The survey participants do not have any unusual skill at forecasting in general.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductreasoning

    Because participants received randomized subsets of the questions, the number of responses is far less than 2,778 for most individual questions.

    “Because participants received randomized subsets of the questions, the number of responses is far less than 2,778 for most individual questions.”
    Quote in the paperNot a substantial claim
57 of 57

  • The paper relies on this without support.

    The surveyed AI researchers are representative of AI researchers as a whole.

    “We conducted a survey of 2,778 AI researchers who had published peer-reviewed research in the prior year in six top AI venues (NeurIPS, ICML, ICLR, AAAI, IJCAI, JMLR).”

  • The paper relies on this without support.

    The 2022 and 2023 surveys are comparable enough to attribute changes in forecasts to changes in opinion rather than differences in sample composition or survey administration.

    “The 2023 survey included around four times as many participants as the 2022 survey by expanding from two publication venues (NeurIPS and ICML) to six.”

  • The paper relies on this without support.

    The survey participants do not have any unusual skill at forecasting in general.

    “Our participants' expertise is in AI, and they do not, to our knowledge, have any unusual skill at forecasting in general.”

  • The authors state this as an assumption.

    HLMI will at some point exist.

    “Assume for the purpose of this question that HLMI will at some point exist.”

  • The authors state this as an assumption.

    Thinking more about a topic improves predictions about it.

    “while thinking more about a topic presumably improves predictions about it”
5 of 5

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentt(1327) = -0.58reported p = .564 · recomputed p = .562

    “t(1327) = -0.58,p = .564”

1 of 1

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Conflicting time horizon for the AI-performance question

    Minor

    The prose says this question asks about 2 and 20 years after HLMI, while Table 1 reports results for 2 and 30 years after HLMI.

  • Year-valued responses attributed to the wrong framing

    Minor

    The text says respondents entered calendar years in fixed-probabilities questions, but then says the correction was applied to fixed-years answers. Under the paper's definitions, fixed-years answers are probabilities, not years.

  • Conflicting counts of tasks forecast to take over ten years

    Minor

    The text first says only four of the 39 tasks fall outside the ten-year window, then identifies six tasks expected to take longer than ten years.

3 of 3

Citations

Cites

Part of this paper’s reference list could not be split into single entries.

Flagged first, then by citations

Research works printed but not linked: 4

  1. Stephen M. Omohundro. The basic ai drives. In Proceedings of the 2008 Conference on Artificial General Intelligence 2008: Proceedings of the First AGI Conference, page 483-492, NLD, 2008. IOS Press. ISBN 9781586038335.
  2. Katja Grace, John Salvatier, Allan Dafoe, Baobao Zhang, and Owain Evans. When will ai exceed human performance? evidence from ai experts. Journal of Artificial Intelligence Research, 62:729-754, 2018a.
  3. Ezra Karger, Josh Rosenberg, Zachary Jacobs, Molly Hickman, Rose Hadshar, Kayla Gamin, Taylor Smith, Tegan McCaslin, Stephen Thomas, and Philip E Tetlock. Forecasting existential risks: Evidence from a long-run forecasting tournament, 2023.
  4. Thomas R Stewart. Scientists' uncertainty and disagreement about global climate change: A psychological perspective. International Journal of Psychology, 26(5):565-573, 1991.

Cited by

Sorted by citations

Related

Metadata

Authors
Grace, K., Sandkühler, J. F., Stewart, H., Weinstein-Raun, B., Thomas, S., Stein-Perlman, Z., Salvatier, J., Brauner, J., & Korzekwa, R. C.
Affiliations
  • University of Bonn (DE)
  • University of Oxford (GB)
ORCID
Venue
Journal of Artificial Intelligence Research
Publisher
AI Access Foundation
Published
2025-10-06
Volume
84
ISSN
1076-9757 (ISSN-L)
DOI
10.1613/jair.1.19087
OpenAlex
W4414876758
Semantic Scholar
6f702f88f0d74a24f1ff06211975b30839ac2c58
Cited by
201
References
20
Citations by year
2026: 64 · 2025: 72
FWCI
35.01
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Journal of Artificial Intelligence Research (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
Open Philanthropy Project
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Impact of AI and Big Data on Business and Society
Keywords
artificial intelligence forecasting, AI risks, human extinction risk, AI misinformation, authoritarianism, inequality, AI risk mitigation, participation bias

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review