Chapter 04
Experts
Published estimates: judgements, not measurements, each with the exact question it answers. Two figures that answer different questions never share an axis.
Stance
Figure 4.1 · Each point is a published estimate, grouped by the question it answers. Figures from different questions are not compared with one another, and so they never share an axis.
42 of 42 estimates
Shape: circle, concerned; diamond, middle ground; square, sceptic. Filled: survey; hollow: individual opinion. A line is a range, or the middle half of responses (p25–p75) in a survey; dotted, a range with no ceiling.
p(doom), no horizon
stated probability, 0% to 100%Declines to quantify
Catastrophe this decade
stated probability, 0% to 100%Extinction this decade
stated probability, 0% to 100%Catastrophe this century
stated probability, 0% to 100%Extinction this century
stated probability, 0% to 100%Arrival of general AI
stated yearFull automation
stated yearOpen entry: Dario Amodei
View as table
p(doom), no horizon
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| Dario Amodei · Anthropic | 25% | Concerned | Individual opinion | Sep 2025 | What is your p(doom) number? Answered as the probability that things go 'really, really badly'. | I really hate that term | THE EVENT IS NOT EXTINCTION. Amodei quantifies things going 'really, really badly', and puts 75% on them going 'really, really well'. Glossing p(doom) as 'that it destroys humanity' is Axios's wording, not Amodei's, which is why this figure cannot share a row with an extinction estimate. It is also a mixed case relative to those who decline to quantify: he rejects the term and gives the number anyway. Remarks at the Axios AI+ DC Summit, in conversation with Jim VandeHei. | [104] Amodei on AI: "There's a 25% chance that things go really, really badly" ↗ archived copy only |
| Paul Christiano · Alignment Research Center | 22% | Concerned | Individual opinion | Apr 2023 | What is the probability of an AI takeover? | Probability of an AI takeover: 22% | The 22% breaks down into 15% for a takeover by a human-built system and 7% for one that does not take over but builds a successor that does. It is not extinction and carries no calendar horizon. Christiano explicitly asks that these not be treated as calibrated predictions: they are guesses with 'half a significant figure' that fluctuate day to day. The quote joins two verbatim fragments from the same post, separated in the original. | [242] My views on "doom" ↗ verified through the LessWrong API |
| ESPAI 2023 (Grace et al.) · AI Impacts | 5% | Middle ground | Survey · n = 1321 | Oct 2023 | What probability do you put on future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species? | What probability do you put on future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species? | A DIRECT extinction question, distinct from the HLMI-conditional one even though the median also lands at 5%. Mean 16.2%, interquartile range of 19 points. It carries no time horizon. Across this question and its two variants, between 41.2% and 51.4% of respondents put more than 10%: that range is what gets confused with the 38% from the conditional question. Each subsample saw only one of the three variants. | [470] Thousands of AI Authors on the Future of AI ↗ |
| ESPAI 2023 (Grace et al.) · AI Impacts | 5% | Middle ground | Survey | Oct 2023 | Assume for the purpose of this question that HLMI will at some point exist. How positive or negative do you expect the overall impact of this to be on humanity, in the long run? Probability assigned to the worst category, 'extremely bad (e.g. human extinction)'. | Assume for the purpose of this question that HLMI will at some point exist. How positive or negative do you expect the overall impact of this to be on humanity, in the long run? | This is the famous ESPAI 5%, and it is NOT an extinction question: it is long-run impact conditional on high-level machine intelligence existing. Median 5%, mean 9%. The widely quoted 38% is something else: the share of respondents who assigned at least 10% to that category, not a probability. 57.8% treated it as a non-trivial possibility (5% or more). Between 2022 and 2023 the share at 10% or more fell from 48% to 38% and the mean from 14% to 9%. The survey had 2,778 respondents, but not everyone saw the same questions and the exact n for this one was not recorded in the research. | [470] Thousands of AI Authors on the Future of AI ↗ |
| ESPAI 2023 (Grace et al.) · AI Impacts | 10% | Middle ground | Survey · n = 661 | Oct 2023 | What probability do you put on human inability to control future advanced AI systems causing human extinction or similarly permanent and severe disempowerment of the human species? | What probability do you put on human inability to control future advanced AI systems causing human extinction or similarly permanent and severe disempowerment of the human species? | Mean 19.4%, interquartile range of 29 points. Naming the mechanism -loss of control- doubles the median relative to the same question without it (10% vs 5%), even though it logically describes a subset of the cases. It is the clearest framing effect within a single survey. | [470] Thousands of AI Authors on the Future of AI ↗ |
| Geoffrey Hinton · University of Toronto | 10%–20% | Concerned | Individual opinion | Jun 2025 | What is the probability that AI wipes out humanity? Answered with no time window and described by Hinton himself as gut feeling, not an estimate. | I often say [there's a] 10% to 20% chance [for AI] to wipe us out. But that's just gut, based on the idea that we're still making them and we're pretty ingenious. | NO HORIZON, deliberately. The «three decades» in circulation is the Guardian's wording over an audio interview, not Hinton's. That interview's primary source —BBC Radio 4's Today, 27 December 2024— is now UNRECOVERABLE: the audio 404s, there is no transcript or clip, and Hinton is not even named in the BBC's own synopses. Hence the observatory uses the June 2025 wording, in his own words. He calls the number «just gut». The brackets in the quote are CNBC's editorial insertions. | [255] There's a '10% to 20% chance' that AI will displace humans completely, says 'godfather' of the technology ↗ |
| Kai-Fu Lee · Sinovation Ventures y 01.AI | Declines to quantify | Sceptic | Individual opinion | Jun 2024 | Does loss-of-control risk exist once systems more capable than their creators appear? Answered by affirming that it exists, but without giving a figure. | 我觉得存在的，但概率不会很高。如果我们越来越依赖Reward model完全让AI自己找路径的话，发生的概率或许会增高。 · Translation: I think it exists, but the probability will not be very high. If we rely more and more on the reward model and let the AI find its own path, the probability of it happening may increase. | He answers another researcher's literal question in the same dialogue. He says the probability is not high and does not quantify it, adding the condition that would raise it: relying on the reward model and letting the system find its own path. | [37] 张亚勤 X 李开复：走向通用人工智能 ↗ |
| Yann LeCun · Meta | Declines to quantify | Sceptic | Individual opinion | Apr 2026 | How large is the existential risk from AI? Answered qualitatively, with neither the event nor the horizon operationalised. | Most "leading AI figures" think this p(doom) estimates are complete bullshit and the existential risk is essentially zero. But most of them are silent. The doomers attract a disproportionate amount of attention, of course. | 'Essentially zero' is technically an estimate, but it is recorded here WITHOUT A NUMBER on purpose: giving it decimals would lend it a precision the phrase does not have, and placing it in the same row as a 0.38% with a confidence interval would compare incomparable things. His typified argument combines two lines: timelines are inflated -in 2023 he wrote that super-human AI is nowhere near the top of the list of existential risks, largely because it does not exist yet, and that discussing how to make it safe before there is a basic design for even dog-level AI is premature- and the debate itself is skewed by media attention. Date verified by decoding the post ID (2026-04-13 12:52 UTC); the text comes from search results because X blocks direct fetching. | [631] Posteo de Yann LeCun sobre estimaciones de p(doom) ↗ |
| Melanie Mitchell · Santa Fe Institute | Declines to quantify | Sceptic | Individual opinion | Sep 2025 | What is the probability of catastrophe from AI? Answered by attacking the empirical basis of the premises rather than giving a figure. | no magical "emergence" need be invoked | Her objection is not about probabilities but about the facts used as evidence of emergent agency, and she dismantles three concrete cases: (1) the model that spoke Bengali without having been trained on Bengali -PaLM's own paper shows it was-; (2) translating without having been programmed to translate, explained by incidental bilingualism and code-switching in the corpus; and (3) the model that 'lets an executive die', a red-teaming exercise its own authors described as extremely contrived and in which the model was asked to play a character. She quotes Murray Shanahan on the science-fiction trope of the rogue AI: a suitably prompted model will begin to role-play that part. It is the strongest version of the argument that capability does not imply agency, and it comes with verifiable evidence instead of a number. Published as a reply to two Thomas Friedman columns in the New York Times. | [737] Magical Thinking on AI: A Response to Thomas Friedman's Recent AI Columns in the New York Times ↗ |
| Arvind Narayanan y Sayash Kapoor · Princeton University | Declines to quantify | Sceptic | Individual opinion | Jul 2024 | Are probabilistic forecasts of AI existential risk useful as policy inputs? Answered with a refusal to quantify and a critique of the exercise itself. | AI x-risk forecasts are far too unreliable to be useful for policy, and in fact highly misleading | The argument is structural: there are only three ways to justify a probability to a skeptic -inductive, deductive or subjective- and none works here. The inductive route needs a reference class, and for existential risk from AI there is none, because it is an event like no other. Their key warning is that it has been forgotten in this debate that probabilities carry no authority by themselves. They distinguish epistemic uncertainty -that of asteroid 2024 YR4, which is resolved by measuring- from stochastic uncertainty, and place AI risk where neither interpretation applies cleanly. They do not object to forecasting as an academic activity: they object to its use as a policy input. | [771] AI existential risk probabilities are too unreliable to inform policy ↗ |
| Arvind Narayanan y Sayash Kapoor · Princeton University | Declines to quantify | Sceptic | Individual opinion | Apr 2025 | What probability do you assign to an existential catastrophe from AI? Answered with an explicit refusal to quantify. | We have not tried to quantify probabilities, but we have tried to make predictions that can tell us whether or not AI is behaving like normal technology | The refusal to give a number IS the argument, not an omission: they propose falsifiable predictions instead of probabilities. Their thesis has four verbatim-verified pieces: (1) what is at stake is not intelligence but power, understood as the ability to modify one's environment; (2) slow diffusion protects, because a system would have to demonstrate reliable performance in less critical contexts long before being granted access to consequential decisions; (3) catastrophic misalignment is by far the most speculative of the risks they discuss; and (4) the real risk is concentration of power, which they consider a bigger problem than the possibility of AI accidents. Published by the Knight First Amendment Institute. | [773] AI as Normal Technology ↗ |
| Andrew Ng · DeepLearning.AI / Stanford University | Declines to quantify | Sceptic | Individual opinion | Jun 2023 | Is AI a risk for human extinction? Answered by asking for the mechanism rather than assigning a probability. | I'd like to have a real conversation about whether AI is a risk for human extinction. Honestly, I don't get how AI poses this risk. What are your thoughts? And, who do you think has a thoughtful perspective on how AI poses this risk that I should talk to? | He does not deny the logical possibility: he denies the argument holds, which is why he asks for the mechanism instead of debating the number. His second line is the cost of the discourse: on 31 October 2023, replying to Hinton, he wrote that overhyped fears about AI leading to human extinction are causing real harm and that he sees young students discouraged from entering AI because they do not want to contribute to human extinction. The 2015 line about overpopulation on Mars attributed to him exists only in secondary press and is not quoted here. Date verified by decoding the post ID; the text comes from search results because X blocks direct fetching. | [786] Posteo de Andrew Ng pidiendo argumentos sobre riesgo de extinción ↗ |
| Fabro Steibel · Instituto de Tecnologia e Sociedade do Rio de Janeiro | Declines to quantify | Sceptic | Individual opinion | Sep 2026 | Is a world-dominating superintelligence the most likely scenario? Answered comparatively, without a probability. | Essa visão mais extremista de superinteligência, hoje, está mais para ficção do que para realidade. Pode ser uma realidade? Pode. Mas eu não aposto que essa deva ser a principal realidade que a gente deve ter. · Translation: That more extreme vision of superintelligence, today, is closer to fiction than to reality. Could it become real? It could. But I do not bet on that being the main reality we should expect. | He does not deny the mechanism: he says the theory is not made up and rests on the evolution of computing power, and that what strikes him as unrealistic is that this scenario is the most likely one. His substantive objection is material and specific to his region: for the share of the population without internet or electricity, that AI world is not the reality that will happen. He also objects to the composition of the debate, which he describes as people of similar age and background. | [388] A humanidade vai acabar? Para especialista brasileiro, superinteligência ainda é ficção ↗ |
| Eliezer Yudkowsky · Machine Intelligence Research Institute | Declines to quantify | Concerned | Individual opinion | Mar 2023 | What is the most likely result of building a superhumanly smart AI under anything remotely like the current circumstances? Answered as a claim about the modal outcome, not a probability. | Many researchers steeped in these issues, including myself, expect that the most likely result of building a superhumanly smart AI, under anything remotely like the current circumstances, is that literally everyone on Earth will die. Not as in "maybe possibly some remote chance," but as in "that is the obvious thing that would happen." | He gives no percentage, and the absence is deliberate: it is a claim about the MODAL outcome -what is most likely- not about probability. It qualitatively implies more than 50%, but turning it into a decimal figure would invent a precision the text does not offer. Signed essay in TIME; the HTML datePublished is verified at 2023-03-29T22:01:58Z. | [1132] Pausing AI Developments Isn't Enough. We Need to Shut it All Down ↗ verified with a browser |
| Zeng Yi · Instituto de Automatización, Academia China de Ciencias | Declines to quantify | Concerned | Individual opinion | Jun 2023 | Why sign the statement on extinction risk, and how likely is that risk? Answered without giving any probability and shifting the problem to the present. | 甚至不需要达到通用人工智能的阶段就有可能对人类造成生存风险 · Translation: It does not even take reaching the stage of artificial general intelligence for it to pose an existential risk to humanity. | He signed the statement on extinction risk and still does not quantify. His thesis is absent from the Anglophone debate: existential risk does not require general intelligence, because a system that merely processes information without understanding it can cause it. His stance does not fit the site's enum cleanly: he signs with the proponents while disputing the premise of both camps. | [768] 370余专家再次联名警告AI"灭绝风险"，签名中国学者曾毅：公开信不是阻碍而是探索 ↗ |
Catastrophe this decade
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| Paul Christiano · Alignment Research Center | 46% | Concerned | Individual opinion | Apr 2023 | What is the probability that humanity irreversibly messes up its future within 10 years of building powerful AI? | Probability that humanity has somehow irreversibly messed up our future within 10 years of building powerful AI: 46% | It is the highest of Christiano's three figures and the least quoted, because the event is broader than extinction: it includes futures where humanity survives but its trajectory is ruined beyond repair. The ten-year horizon is conditional on powerful AI having been built. The quote joins two verbatim fragments from the same post, separated in the original. | [242] My views on "doom" ↗ verified through the LessWrong API |
| Paul Christiano · Alignment Research Center | 20% | Concerned | Individual opinion | Apr 2023 | What is the probability that most humans die within 10 years of building powerful AI? | Probability that most humans die within 10 years of building powerful AI (powerful enough to make human labor obsolete): 20% | The 20% breaks down into 11% from AI takeover and 9% from other paths such as war or terrorism. The ten-year horizon is CONDITIONAL on powerful AI having been built, not a calendar window: it is not comparable with a calendar-decade row without that caveat. The quote joins two verbatim fragments from the same post, separated in the original. | [242] My views on "doom" ↗ verified through the LessWrong API |
| Estudio Delphi de 272 expertos internacionales · The University of Queensland / MIT FutureTech | 21% | Middle ground | Survey · n = 272 | Jun 2026 | What is the probability of a catastrophic outcome from AI-enabled weapons and cyberattacks between end-2025 and end-2030 under a business-as-usual scenario? Catastrophic is defined as more than one million deaths, more than USD 100 billion in financial loss, or intangible damage on a civilisational scale. | catastrophic outcomes are likely correlated | Mean probability with a 95% confidence interval of 15,1%-27,5%. Mean severity 3,49 out of 5. Under a pragmatic-mitigation scenario it falls to 12% and remains among the five above 10%. IT DOES NOT MEASURE EXTINCTION and the window is five years, not a full decade: the study rates 24 risk domains and, under business as usual, 18 of the 24 exceed 10%. The authors warn the probabilities CANNOT BE SUMMED because catastrophic outcomes are correlated and the joint probability was not measured. Three Qualtrics rounds between September and November 2025; 272 experts from 37 countries, of whom 214 (79%) completed all three. Funded by the Commonwealth Bank of Australia, which reviewed the design but not data collection or analysis. Only the month of publication is documented. | [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts ↗ |
| Estudio Delphi de 272 expertos internacionales · The University of Queensland / MIT FutureTech | 21.5% | Middle ground | Survey · n = 272 | Jun 2026 | What is the probability of a catastrophic outcome from dangerous AI capabilities between end-2025 and end-2030 under a business-as-usual scenario? Catastrophic is defined as more than one million deaths, more than USD 100 billion in financial loss, or intangible damage on a civilisational scale. | catastrophic outcomes are likely correlated | Mean probability with a 95% confidence interval of 16,9%-26,4%. Mean severity 3,49 out of 5. It is the highest-probability risk of the 24 assessed; under a pragmatic-mitigation scenario it falls to 12% and remains among the five above 10%. IT DOES NOT MEASURE EXTINCTION and the window is five years, not a full decade: the study rates 24 risk domains and, under business as usual, 18 of the 24 exceed 10%. The authors warn the probabilities CANNOT BE SUMMED because catastrophic outcomes are correlated and the joint probability was not measured. Three Qualtrics rounds between September and November 2025; 272 experts from 37 countries, of whom 214 (79%) completed all three. Funded by the Commonwealth Bank of Australia, which reviewed the design but not data collection or analysis. Only the month of publication is documented. | [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts ↗ |
| Estudio Delphi de 272 expertos internacionales · The University of Queensland / MIT FutureTech | 18% | Middle ground | Survey · n = 272 | Jun 2026 | What is the probability of a catastrophic outcome from AI-driven concentration of power between end-2025 and end-2030 under a business-as-usual scenario? Catastrophic is defined as more than one million deaths, more than USD 100 billion in financial loss, or intangible damage on a civilisational scale. | catastrophic outcomes are likely correlated | Mean probability with a 95% confidence interval of 12,1%-24,8%. Mean severity 3,47 out of 5. Under a pragmatic-mitigation scenario it falls to 11% and remains among the five above 10%. IT DOES NOT MEASURE EXTINCTION and the window is five years, not a full decade: the study rates 24 risk domains and, under business as usual, 18 of the 24 exceed 10%. The authors warn the probabilities CANNOT BE SUMMED because catastrophic outcomes are correlated and the joint probability was not measured. Three Qualtrics rounds between September and November 2025; 272 experts from 37 countries, of whom 214 (79%) completed all three. Funded by the Commonwealth Bank of Australia, which reviewed the design but not data collection or analysis. Only the month of publication is documented. | [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts ↗ |
| Estudio Delphi de 272 expertos internacionales · The University of Queensland / MIT FutureTech | 16.6% | Middle ground | Survey · n = 272 | Jun 2026 | What is the probability of a catastrophic outcome from competitive dynamics between AI developers between end-2025 and end-2030 under a business-as-usual scenario? Catastrophic is defined as more than one million deaths, more than USD 100 billion in financial loss, or intangible damage on a civilisational scale. | catastrophic outcomes are likely correlated | Mean probability with a 95% confidence interval of 12,0%-21,6%. Mean severity 3,49 out of 5. It shares the study's highest mean severity with dangerous capabilities and with weapons and cyberattacks. IT DOES NOT MEASURE EXTINCTION and the window is five years, not a full decade: the study rates 24 risk domains and, under business as usual, 18 of the 24 exceed 10%. The authors warn the probabilities CANNOT BE SUMMED because catastrophic outcomes are correlated and the joint probability was not measured. Three Qualtrics rounds between September and November 2025; 272 experts from 37 countries, of whom 214 (79%) completed all three. Funded by the Commonwealth Bank of Australia, which reviewed the design but not data collection or analysis. Only the month of publication is documented. | [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts ↗ |
| Estudio Delphi de 272 expertos internacionales · The University of Queensland / MIT FutureTech | 12.8% | Middle ground | Survey · n = 272 | Jun 2026 | What is the probability of a catastrophic outcome from AI-generated or AI-amplified false information between end-2025 and end-2030 under a business-as-usual scenario? Catastrophic is defined as more than one million deaths, more than USD 100 billion in financial loss, or intangible damage on a civilisational scale. | catastrophic outcomes are likely correlated | Mean probability with a 95% confidence interval of 8,9%-18,1%. Mean severity 3,44 out of 5. It is the fifth highest-severity risk in the study. IT DOES NOT MEASURE EXTINCTION and the window is five years, not a full decade: the study rates 24 risk domains and, under business as usual, 18 of the 24 exceed 10%. The authors warn the probabilities CANNOT BE SUMMED because catastrophic outcomes are correlated and the joint probability was not measured. Three Qualtrics rounds between September and November 2025; 272 experts from 37 countries, of whom 214 (79%) completed all three. Funded by the Commonwealth Bank of Australia, which reviewed the design but not data collection or analysis. Only the month of publication is documented. | [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts ↗ |
Extinction this decade
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| Geoffrey Hinton · University of Toronto | 10% | Concerned | Individual opinion | Sep 2026 | Is there a greater than 10% chance that AI kills all humans within the next decade? Answered as a declared judgement, not as a calculation: 'nobody knows how to estimate it'. | It would be foolish to say there's a one percent chance. Nobody knows how to estimate it. A 10 percent chance seems not unreasonable to me. | FIRST EXPLICIT HORIZON IN HIS OWN WORDS. His recorded figure until now was 10-20% with no timeframe, declared 'just gut' (June 2025, cnbc-2025-hinton); the circulating 'three decades' had no recoverable source. Here he answers the same decade question as Hubinger (p-extincion-decada family), on BBC Newsnight of September 9: the presenter asked him to confirm that '10% doesn't seem an unreasonable estimate' and he said 'yes'. The number is a judgement, and he says so himself: it should not be read as a calculation with decimals. | [540] Five Leading AI Experts Warn Superintelligence Could Kill Humans and Explain Their Case ↗ |
| Evan Hubinger · Anthropic | ≥ 10% | Concerned | Individual opinion | Sep 2026 | What is the probability that AI kills all humans within the next decade? | we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade | A personal estimate and a LOWER BOUND, not a point estimate nor an official Anthropic position: when plotted it must be shown as a bound, not a point. The event is literally AI killing all humans, and the horizon is calendar-based -the next decade- not conditional on anything. It came in response to the public resignation, that same day, of Jacob Coxon, an Anthropic pre-training researcher. The date is independently verified by decoding the post identifier (2026-09-09 01:27 UTC); the text comes from secondary coverage reproducing it in full, because X blocks anonymous fetching. | [530] Publicación de Evan Hubinger sobre la probabilidad de que la IA acabe con la humanidad ↗ |
Catastrophe this century
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| Expertos de dominio del XPT · Forecasting Research Institute | 12% | Middle ground | Survey · n = 59 | Jul 2023 | What is the probability of an AI-caused catastrophe before 2100, defined as the death of more than 10% of the human population within five years? | why were superforecasters so unmoved by experts' much higher estimates of AI extinction risk, and why were experts so unmoved by the superforecasters' lower estimates? | Final group median, with a reported range of 4.0% to 18.5%. It is nearly six times that of superforecasters for the same question. Catastrophe here means the death of more than 10% of humanity within five years, not extinction. | [587] Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament ↗ |
| Superpronosticadores del XPT · Forecasting Research Institute | 2.13% | Sceptic | Survey · n = 88 | Jul 2023 | What is the probability of an AI-caused catastrophe before 2100, defined as the death of more than 10% of the human population within five years? | why were superforecasters so unmoved by experts' much higher estimates of AI extinction risk, and why were experts so unmoved by the superforecasters' lower estimates? | Final group median, with a reported range of 1.83% to 3.00%. Catastrophe is not extinction: the operational threshold is the death of more than 10% of humanity within five years, far lower than that of the extinction row. The quote is the report's own framing of the disagreement between groups. | [587] Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament ↗ |
Extinction this century
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| ESPAI 2023 (Grace et al.), subgrupo formado en Asia · AI Impacts | 10% | Middle ground | Survey | Oct 2023 | How likely is AI to cause human extinction or similarly permanent and severe disempowerment of the species? Median of the subgroup that did their undergraduate studies in Asia, within the same survey. | The region respondents graduated from affected responses somewhat: the Asian median was 10%, while North American and European medians were 5%. | NOT A SEPARATE SURVEY: it is a subgroup of the same sample, reported in appendix A.3 as a demographic comparison, with no n of its own and without saying which of the three extinction questions it uses. And it cannot be read as «Asians are more pessimistic» without its counterweight: on the question conditional on human-level AI that same group is MORE optimistic than the North American and European ones. It is published because it is the only regional breakdown that exists in the field's large surveys. | [470] Thousands of AI Authors on the Future of AI ↗ |
| ESPAI 2023 (Grace et al.) · AI Impacts | 5% | Middle ground | Survey · n = 655 | Oct 2023 | What probability do you put on future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species within the next 100 years? | What probability do you put on future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species? | The same direct question restricted to 100 years. Mean 14.4%, median 5%. Adding a century-long horizon did not move the median and lowered the mean by two points, suggesting respondents were already thinking on that timescale. The verbatim quote is that of the unbounded variant; the 'within the next 100 years' restriction is described in the paper as the difference between subsamples. | [470] Thousands of AI Authors on the Future of AI ↗ |
| Grupo escéptico de la colaboración adversarial del FRI · Forecasting Research Institute | 0.12% | Sceptic | Survey · n = 11 | Mar 2024 | Will AI cause an existential catastrophe by 2100? Defined as extinction or unrecoverable collapse: global GDP below one trillion dollars a year for a million years, or human population below one million. | Will AI cause an existential catastrophe by 2100? | Value after eight weeks of adversarial collaboration (April-May 2023) with a forum, biweekly calls and some 250 crux ideas. The group median moved from 0.10% to 0.12%: disagreement did not close despite the effort, and skeptics spent a median of 80 hours. The strongest crux they built would close only ~1.2 percentage points of a ~23-point gap. | [920] Roots of Disagreement on AI Risk: Exploring the Potential and Pitfalls of Adversarial Collaboration ↗ |
| Grupo preocupado de la colaboración adversarial del FRI · Forecasting Research Institute | 20% | Concerned | Survey · n = 11 | Mar 2024 | Will AI cause an existential catastrophe by 2100? Defined as extinction or unrecoverable collapse: global GDP below one trillion dollars a year for a million years, or human population below one million. | Will AI cause an existential catastrophe by 2100? | Post-collaboration value: the group median fell from 25% to 20%, a small move against a gap of more than two orders of magnitude with the skeptical group. It is the most direct evidence that disagreement on AI risk is mostly not empirical: no 2030 indicator would settle it. | [920] Roots of Disagreement on AI Risk: Exploring the Potential and Pitfalls of Adversarial Collaboration ↗ |
| Encuesta de Kestin y Soares en Harvard · Harvard University / Machine Intelligence Research Institute | 70% | Concerned | Survey · n = 89 | Mar 2026 | What probability do you put on future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species? (Assume that companies developing AI technology proceed largely unimpeded by external intervention.) | self-selection bias in our sample given that attendees chose to attend an event with a provocative title | DECISIVE SAMPLE BIAS: this is not a measurement of risk nor of expert opinion. Respondents were voluntary attendees of a 11 March 2026 talk on the book 'If Anyone Builds It, Everyone Dies', at an event with that title promoted to the Harvard Book Store mailing list, and one of the study's authors co-wrote that book. The 70% is the midpoint of the 60-80% band, the post-event median of the matched sample of 89; the pre-event median was in the 40-60% band, with means of 50.5% and 56.9%. The study's solid finding is not the level but the shift, and it is counterintuitive: among those with little prior familiarity (n=15), 60% moved up and none moved down; among the ten self-declared experts, none moved up and 20% moved down. | [592] Views on AI Existential Risk Before and After a Public Event at Harvard University ↗ |
| Encuesta del Summit on Existential Security · Summit on Existential Security | 25% (median; p25–p75: 14%–50%) | Concerned | Survey · n = 59 | Feb 2026 | What is the probability of human extinction or permanent human disempowerment before 2100? | What is the probability of human extinction or permanent human disempowerment before 2100? | Mean 34%, response range 5-95%, standard deviation 24.2 points. It is the same question family as the XPT and differs from superforecasters by nearly two orders of magnitude; the most plausible explanation is not new evidence but who is in the room: the authors themselves describe the sample as a self-selected group of practitioners working on existential risk. Only the month of fieldwork is documented (February 2026). | [919] Survey of AI safety leaders on x-risk, AGI timelines, and resource allocation ↗ |
| Expertos de dominio del XPT · Forecasting Research Institute | 3% | Middle ground | Survey · n = 59 | Jul 2023 | What is the probability that AI is the cause of human extinction before 2100? | why were superforecasters so unmoved by experts' much higher estimates of AI extinction risk, and why were experts so unmoved by the superforecasters' lower estimates? | Final group median, with a reported range of 0.49% to 10%. It is the same question superforecasters answered at 0.38%: the near order-of-magnitude gap did not close after four persuasion stages. This group's extinction risk from ANY cause is 6%, a figure often confused with this one. | [587] Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament ↗ |
| Superpronosticadores del XPT · Forecasting Research Institute | 0.38% | Sceptic | Survey · n = 88 | Jul 2023 | What is the probability that AI is the cause of human extinction before 2100? | why were superforecasters so unmoved by experts' much higher estimates of AI extinction risk, and why were experts so unmoved by the superforecasters' lower estimates? | Final group median, with a reported range of 0.10% to 0.75%. The tournament ran from June to October 2022 with 169 forecasters across four persuasion stages; the date is that of the first version of the report. Not to be confused with extinction risk from ANY cause, which for this same group is 1%. Conditional on AGI existing before 2070, this median rises from 0.38% to 1%. | [587] Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament ↗ |
Arrival of general AI
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| Yoshua Bengio · Mila | 2028–2043 | Concerned | Individual opinion | Jun 2023 | What is your 95% confidence interval for the time horizon of superhuman intelligence? | My current estimate places a 95% confidence interval for the time horizon of superhuman intelligence at 5 to 20 years | The absolute years 2028-2043 come from adding 5 and 20 years to the text's date (24 June 2023); the original speaks of horizons, not calendar years. This is a TIMELINE estimate, not a risk estimate: Bengio publishes no p(doom) of his own in any of the texts reviewed. In the same piece he proposes a four-statement chained survey so readers can build their own probability instead of adopting his. | [127] FAQ on Catastrophic AI Risks ↗ |
| ESPAI 2023 (Grace et al.) · AI Impacts | 2047 | Middle ground | Survey · n = 1714 | Oct 2023 | In what year is there a 50% chance of high-level machine intelligence (HLMI), achieved when unaided machines can accomplish every task better and more cheaply than human workers? Think feasibility, not adoption. | High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers. Ignore aspects of tasks for which being a human is intrinsically advantageous, e.g. being accepted as a jury member. Think feasibility, not adoption. | Aggregate forecast fitted with gamma distributions: 10% by 2027 and 50% by 2047. The 50% mark moved 13 years earlier than the 2060 of the 2022 survey, whereas between 2016 and 2022 it had shifted by barely a year. The question asks respondents to assume human scientific activity continues without major negative disruption. | [470] Thousands of AI Authors on the Future of AI ↗ |
| Huang Tiejun · Academia de Inteligencia Artificial de Beijing (BAAI) y Universidad de Pekín | 2045 | Middle ground | Individual opinion | Sep 2025 | In what year will an embodied artificial general intelligence arrive that fully surpasses humans in perception and cognition and has self-awareness? | 到2045年，感知认知全面超越人类而且具有自我意识的具身AGI有望到来，开启人机共生新篇章。 · Translation: By 2045 an embodied AGI is expected to arrive that fully surpasses humans in perception and cognition and has self-awareness, opening a new chapter of human-machine symbiosis. | In the same paragraph he sets 2030 for «an AGI surpassing human cognitive capability»: he uses the same word for two different things without distinguishing them, and 2045 is the one he calls «the true AGI era». His framing of risk is not that it is low, but that AI is the instrument for withstanding it. | [109] 智源研究院黄铁军：2015、2030、2045，AI促进可持续发展 ↗ |
| Daniel Kokotajlo · AI Futures Project | Dec 2030 | Concerned | Individual opinion | Jan 2026 | What is your all-things-considered median for the arrival of TED-AI, the superhuman AI R&D milestone that drives the AI 2027 transition? | Clarifying how our AI timelines forecasts have changed since AI 2027 | 2030.95 corresponds to December 2030. COMMON CORRECTION: it is not from November 2025; that is an earlier stage, when Kokotajlo was at 'around 2030, lots of uncertainty though'. Full series: 2027 (Dec 2022 to Jan 2025), 2028 (Feb 2025), end of 2029 (Aug 2025), 2030 (Nov 2025) and Dec 2030 (Jan 2026). The post exists precisely to correct press coverage that mistook the old mode for the new median, and it lists errors by the Guardian, the Independent, the Washington Post, Inc and the Daily Mirror. The quote is the post's title: the medians are published in a table, not in prose. The post itself notes a later Q1 2026 Timelines Update with April 2026 views, not reviewed here. | [27] Clarifying how our AI timelines forecasts have changed since AI 2027 ↗ |
| Eli Lifland · AI Futures Project | 2035 | Concerned | Individual opinion | Jan 2026 | What is your all-things-considered median for the arrival of TED-AI, the superhuman AI R&D milestone that drives the AI 2027 transition? | Clarifying how our AI timelines forecasts have changed since AI 2027 | January 2035, some four years later than his co-author Kokotajlo on exactly the same question: the spread between two people who write the same scenario together is the best reminder of how much uncertainty a median carries. Full series: 2050 (Jul 2022), 2038 (Jan 2024), 2035 (mid-2024), 2032 (Dec 2024), 2031 (Apr 2025), 2033 (Jul 2025) and 2035 (Nov 2025 and Jan 2026). The quote is the post's title: the medians are published in a table, not in prose. Neither author gives a p(doom) in this post. | [27] Clarifying how our AI timelines forecasts have changed since AI 2027 ↗ |
| Encuesta del Summit on Existential Security · Summit on Existential Security | 2033 (median; p25–p75: 2031–2036) | Concerned | Survey · n = 59 | Feb 2026 | In what year do you estimate there's a 50% chance we will have developed AGI, defined as an AI system (or collection of systems) that can fully automate the vast majority (>90%) of roles in the 2025 economy? | In what year do you estimate there's a 50% chance we will have developed AGI? | Mean 2034.3, response range 2027-2045. At the 25% threshold the median drops to 2030 (n=48). The AGI definition here is economic -automating more than 90% of roles- and is therefore closer to ESPAI's Full Automation of Labor (median 2116) than to its HLMI (median 2047), which shows the size of the sample effect: 83 years apart on similar definitions. Only the month of fieldwork is documented. | [919] Survey of AI safety leaders on x-risk, AGI timelines, and resource allocation ↗ |
| Zhang Ya-Qin · Instituto de Investigación de Industria Inteligente (AIR), Universidad de Tsinghua | 2039–2044 | Middle ground | Individual opinion | Jun 2024 | When will artificial general intelligence be achieved, understood as passing the new Turing test? | 我比较乐观，我认为15-20年内可以实现，并通过新图灵测试。 · Translation: I am fairly optimistic: I think it can be achieved within 15 to 20 years, and that it will pass the new Turing test. | The range is the 15 to 20 years counted from the date of the dialogue, using the same convention as the Bengio estimate. He states that he revised his own timeline: three years earlier he would have said fifty years. He explicitly excludes consciousness, and his declared concern is not takeover but deployment governance: he fears that if it is not addressed now, a sufficiently capable system deployed at scale brings loss-of-control risk. | [37] 张亚勤 X 李开复：走向通用人工智能 ↗ |
Full automation
| Who | Figure | Stance | Type | Date | Exact question | Quote | Note | Source |
|---|---|---|---|---|---|---|---|---|
| ESPAI 2023 (Grace et al.) · AI Impacts | 2116 | Middle ground | Survey | Oct 2023 | In what year is there a 50% chance of full automation of labor, that is, when all occupations are fully automatable? | when all occupations are fully automatable | 10% by 2037 and 50% by 2116: 69 years later than the HLMI median, even though the two definitions appear to describe almost the same thing. The authors say they cannot explain the gap. This comparison alone justifies grouping by question rather than by person: framing moves the answer by decades. | [470] Thousands of AI Authors on the Future of AI ↗ |