AI & Judgment | Human Consequence | November 2026
Key findings
- The 2024 Expert Survey on Progress in AI gives high-level machine intelligence a 50% chance by 2042 and full automation of labor a 50% chance by 2098. The two dates answer different questions, asked of different subsets of the 1,580 researchers surveyed.
- When the authors randomly changed only the format of the question, the forecasts moved again. Asking for the year that matches a given probability produced earlier timelines than asking for the probability by a given year, for both milestones.
- For anyone making decisions, the practical rule is simple: a number is usable only while its question, method, uncertainty and date stay attached to it. BBGK’s five-step Forecast-to-Decision Check turns that rule into a routine.
Why do AI forecasts from the same experts give 2042 for human-level AI but 2098 for the full automation of labor? Because the two dates answer different questions. Both come from one study: the fourth Expert Survey on Progress in AI, which collected answers from 1,580 AI researchers in December 2024 and was published in September 2026. One forecast defines the milestone by tasks. The other defines it by occupations and reaches it through a longer chain of questions.
The temptation is to pick the earlier date and call it the future, or to point at the gap and dismiss expert prediction altogether. Neither response is justified. The more useful lesson is narrower and more demanding: an AI forecast cannot be interpreted responsibly without the exact question that produced it. That is not a rhetorical point. The survey’s authors tested it with a randomized experiment, and the result should change how leaders read the next headline about when experts expect human-level AI.
The two dates measure different futures
The study, Advanced AI according to 1,580 researchers, was written by Katja Grace and eight co-authors at AI Impacts. Responses were collected between 9 and 24 December 2024. That date matters: the paper records what researchers expected at the end of 2024, not what they believe today. Katja Grace says as much in her own summary of the results.
The first milestone, high-level machine intelligence (HLMI), is reached when unaided machines can accomplish every task better and more cheaply than human workers. Respondents were told to set aside tasks where being human is an intrinsic advantage, and to think about feasibility, not adoption.
The second, full automation of labor (FAOL), is reached when all occupations are fully automatable. Respondents did not go straight to that question. They first forecast four specific occupations, then named an existing occupation they expected to be among the last to be automated, forecast that one, and only then forecast full automation.
| Feature | High-level machine intelligence | Full automation of labor |
|---|---|---|
| What is defined | Every task done better and more cheaply by unaided machines | Every occupation fully automatable |
| How respondents got there | Asked directly | After four occupations and a self-chosen last occupation |
| 50% year, mean aggregate | 2042 (10% by 2027) | 2098 (10% by 2035) |
| 50% year, median aggregate | 2044 | 2101 |
| Fitted response sets | 985 | 456 |
Two cautions follow. First, task capability is not the same variable as occupation-wide automation. An occupation is a bundle of tasks, constraints, costs, responsibilities and exceptions. Second, the route to the answer differs. Someone who has just pictured the last surviving occupation may reason differently from someone asked to imagine machines outperforming people at every task. The survey does not isolate how much of the 56-year difference comes from the definition and how much from the question sequence, so it would be an overstatement to say that wording caused the whole gap.
The dates are also thresholds on an aggregate curve, not years chosen by a unanimous panel. The authors fitted a probability distribution to each respondent’s three answers and then averaged those distributions. Taking the median instead of the mean moves the 50% years to 2044 and 2101. The headline dates barely shift, but the choice of method is part of the number.
The working rule. A date detached from its operational definition is incomplete information.
The stronger evidence: the same question in two formats
The clearest evidence that the question shapes the forecast is not the headline gap. It is an experiment reported in Appendix D of the paper.
Each respondent had an even chance of receiving the timeline questions in one of two formats. In the fixed-years format, they estimated the probability that a milestone would be reached by given years. In the fixed-probabilities format, they gave the years by which the milestone would have a 10%, 50% and 90% chance. The milestone and its definition were identical. Only the format changed.
The result was consistent. The fixed-probabilities format produced earlier forecasts than the fixed-years format for both HLMI and FAOL, and the authors report that they have seen this difference in every edition of the survey. In the paper’s Figure 25, the two HLMI curves cross the 50% line roughly a decade apart, and the two FAOL curves sit several decades apart. Those magnitudes are BBGK’s reading of a published chart, not figures the authors state, and they should be treated as approximate.
This is not a semantic curiosity. Stating probabilities about distant events is hard, and people reason differently when a question anchors them on calendar years than when it anchors them on probability levels. The fitted distributions then turn three rough answers per person into a smooth curve with more apparent precision than the underlying judgments may support.
The authors took the problem seriously. They combined both formats in their headline figures, compared alternative fitting and aggregation methods, and published the comparison. One of those methods comes from Tom Adamczewski’s reanalysis of earlier survey data. Adamczewski is also a co-author of the 2024 paper, so that reanalysis is a methodological refinement from inside the research team, not independent replication.
None of this is an argument against surveying experts. Randomized designs and open methods are exactly what make such limits visible. A well-designed forecast reveals both a view about the future and the limits of how that view was collected.
Why 2042 is not a deadline
Across the survey’s four waves, the 50% year for HLMI has moved from 2061 in the 2016 survey to 2059 in 2022, 2047 in 2023 and 2042 in 2024. The FAOL forecast has moved earlier too, and the distance between the two has narrowed without closing.
It is legitimate to say that the surveyed researchers’ aggregate expectations moved earlier. It is not legitimate to say the world moved five years closer to HLMI in a single year. The first statement is about elicited opinion. The second is a much stronger claim about technology. The authors themselves note that the surveyed population changes from wave to wave and that changes in code and data cleaning move some results.
A further distinction often disappears in reporting: feasibility is not deployment. Both survey questions ask what machines could do, not whether organizations would adopt them. A system becoming capable of a task does not mean anyone will deploy it immediately, safely, economically or legally. Nor does it mean governance can wait for a broader milestone.
That is the planning dilemma. A leader who waits for 2042 before defining accountability may ignore risks in systems already in use. A leader who treats 2042 as a fixed deadline may misallocate resources if the milestone arrives late, early, or never becomes relevant to that organization. BBGK’s analysis of AI acceleration and the capacity of institutions to adapt makes the same point from the other side: the binding constraint is usually the institution’s ability to respond, not the date on a forecast.
The better question is not which year the curve crosses 50%. It is this: which decisions become unsafe or irreversible under plausible near-term capability changes, and what evidence would justify changing our controls?
The 18% estimate also needs its question
The survey’s most quoted result is often summarized as a prediction of AI-caused human extinction. That summary is too strong, and the authors say so in their own limitations section.
The question asked for the probability of future AI advances causing “human extinction or similarly permanent and severe disempowerment of the human species.” Among the 744 respondents who received that version, the mean answer was 18.3% and the median was 10%, with 52.7% giving at least 10%. Other respondents received variants about loss of control or a 100-year horizon, and the paper reports pooled results across all three.
Three things follow from the wording. The figure covers a combined outcome, so it does not say how much probability anyone assigned to extinction alone. It is an average of subjective estimates, not an observed frequency. And the mean sits well above the median, which tells you a minority of high answers is pulling the average up.
That is no reason to wave the result away. When informed researchers assign non-trivial probability to severe outcomes, that is decision-relevant evidence of concern. It justifies investigation, stress-testing and proportionate precaution, while remaining a different kind of evidence from observed harm.
One more finding undercuts the idea of neat optimist and pessimist camps. Of 1,538 respondents asked about the long-run impact of HLMI, 788 (51.2%) gave at least a 5% chance to both extremely good and extremely bad outcomes. In the authors’ words, the population is not polarized; most respondents agree the future looks very uncertain.
Who answered, and who did not
Any survey number also depends on who replied. The authors emailed 19,874 addresses of researchers who had published in 2023 at six leading AI venues, including NeurIPS, ICML and ICLR. They received 2,052 responses, a response rate they report as 10%. After removing people who turned out to be outside the intended group, 1,580 remained for the main analysis.
The paper describes the steps taken to limit response bias, including payment for participation and a deliberately vague invitation. It reports small demographic differences in who responded and finds no clear direction of bias. It also states one limitation plainly: frontier AI companies have largely stopped publishing in the venues the survey draws on, so the authors expect the sample to underweight researchers who work at those companies.
A 10% response rate does not invalidate the findings, and the authors make a detailed case for their method in a published methodology FAQ. But the right description of the headline numbers is precise: they are the views of the academic-venue AI researchers who chose to answer in December 2024. They are not the view of “AI experts” as a whole.
What independent evidence adds
Other research supports the need to separate expectations from observations, and none of it can confirm the survey’s numbers.
In Pew Research Center’s 2025 study of AI experts and the public, 56% of 1,013 U.S.-based AI experts expected AI to have a positive effect on the United States over the next 20 years, and 57% wanted more control over how AI is used in their own lives. Pew sampled a different population with different questions, so its results cannot be merged with the AI Impacts survey. The comparison is conceptual: expecting benefit does not remove the demand for control.
In a Delphi study from MIT FutureTech and the University of Queensland, 272 international experts rated 24 AI risks in late 2025. The five they expected to cause the most severe harm over the next five years were dangerous capabilities, competitive dynamics, weapons and cyberattacks, power centralization and false information. This is another exercise in structured expert judgment, with its own definitions and a much shorter horizon.
The International AI Safety Report 2026 is a different kind of source: a synthesis of published evidence, not a poll. It describes an evaluation gap, in which performance in pre-deployment tests does not reliably predict how systems behave in real-world use.
| Source | What it supports | What it does not establish |
|---|---|---|
| AI Impacts survey (fielded 2024, published 2026) | Forecast distributions, the format experiment, risk perceptions, sampling method | Calibrated probabilities of AI arrival; an extinction-only probability |
| Adamczewski reanalysis (2024) | That fitting and aggregation choices move the numbers | Independent replication; one correct forecast model |
| Pew Research Center (2025) | That optimism and a wish for control coexist among experts | A measure that can be pooled with the AI Impacts data |
| MIT FutureTech and UQ Delphi study (2026) | Which near-term risks experts rank as most severe | The frequency of any catastrophic outcome |
| International AI Safety Report (2026) | Current capabilities and the gap between testing and real-world behavior | Verification of any survey forecast |
Taken together, these sources support a practical conclusion, not a universal forecast: leaders face consequential uncertainty even after consulting credible experts and technical evidence. Better decisions need a disciplined way to carry that uncertainty into governance.
The BBGK Forecast-to-Decision Check
The BBGK Forecast-to-Decision Check is a five-step test to run before any expert forecast is used to justify a decision. It asks what the forecast actually measured, how the number was produced, what is observed as opposed to predicted, what the institution has at stake, and who will act when the evidence changes.
It is a BBGK working tool in the BBGK Tools class. It has not been experimentally validated.
- Define the eventWhat exactly would count: feasibility, deployment, job substitution or control?
- Inspect the methodWho was asked, when, how many answered, and how was it aggregated?
- Separate forecast from observationWhat can current systems reliably do in your setting?
- Map your exposureWhich decisions are affected, and which errors are irreversible?
- Set triggers and ownersWhat observed change forces a review, and who can act?
Steps 1 and 2 can be answered from the source document. If the source cannot answer them, that is itself the finding.
1. Define the event
Name the capability, social outcome or failure that would count. Decide whether the source is forecasting feasibility, deployment, labor substitution, loss of control or something else. HLMI, AGI and job automation are not interchangeable labels, as the 56-year difference shows.
2. Inspect how the number was produced
Record who was sampled, when they were asked, how many answered that specific question, and how the answers were collected and combined. Is the figure a mean of subjective probabilities, a threshold on a fitted curve, a measured rate, or an author’s inference?
3. Separate forecast from observed evidence
Ask what current systems reliably do in the environments that matter to you, and what independent evaluations and incidents have shown. The BBGK Evidence Meter grades how much confidence a claim deserves. Where no confident estimate exists, record the absence instead of filling it with invented precision.
4. Map institutional exposure
List the activities, people and decisions that would be affected. Estimate what failure would cost and whether an error could be reversed. The same capability carries very different consequences in different settings.
5. Set triggers and accountable owners
Define the observed performance change, incident, model update or regulatory requirement that triggers a review, and name who may pause deployment, require human approval or revise a control. Without an owner and a trigger, the forecast has not been translated into governance.
The check is consistent in spirit with the NIST AI Risk Management Framework, which organizes work under four functions: Govern, Map, Measure and Manage. The BBGK contribution is narrower. It covers one transition only: from a claim about the future to a decision in the present.
A worked example: approving an AI agent
Consider an organization deploying an AI agent to draft replies to customer inquiries. The relevant question is not whether HLMI reaches 50% probability in 2042. It is whether this agent reliably handles this task, with these records and these consequences.
The organization can test representative cases, limit what the agent may send or change, require human approval for sensitive exceptions, monitor failures and name an escalation owner. Under BBGK’s AI Delegation Boundary Framework, that means deciding which parts of the work to Automate, which to Assist, and which a named person must Own Directly. BBGK’s analysis of the judgment calls nobody made shows what happens when those decisions pass to a system by default.
If later systems become more capable, the organization revises its controls on evidence, not on the arrival of a predicted year. Long-run forecasts still have a role in scenario planning, research investment, insurance and contingency analysis. They should not stand in for evidence about the systems an institution is actually using.
Two mistakes to avoid, and what survives
The first mistake is false precision: presenting 2042, 18.3% or any similar figure as if an instrument had measured the future. The outcome definition, the number of respondents, the question format and the uncertainty all disappear, and an apparently definitive number remains.
The second is false dismissal: finding limitations and concluding that expert warning has no decision value. Some catastrophic harms cannot be observed before they first occur. Scenario analysis and precaution can be appropriate where exact probabilities cannot be established.
The risk of the first mistake is highest for organizations that borrow authority from a forecast they did not commission, interpret or check. A date becomes influential through repetition, and its caveats rarely travel as far. The same discipline applies to answers from AI systems themselves, which is why BBGK’s work on the AI validation trap argues for verifying claims outside the channel that produced them.
The next survey may move the HLMI date earlier again, or push it back. Its respondents may be more confident, more divided, or less willing to give precise probabilities. Those shifts will matter as evidence of changing expectations. The institutional obligation will not change: know what a claim measures, separate it from observed capability, and keep authority over decisions whose consequences cannot be delegated to a forecast.
The strongest lesson of this survey is therefore not that AI arrives in 2042, that labor automation waits until 2098, or that catastrophe has a measured 18.3% probability. It is that a number becomes trustworthy only when its question, method, uncertainty and decision context stay attached to it.
Frequently asked questions
When do AI researchers expect human-level AI?
In the 2024 Expert Survey on Progress in AI, published in September 2026, the mean aggregate forecast gave high-level machine intelligence a 50% chance by 2042 and a 10% chance by 2027. The survey defines it as unaided machines accomplishing every task better and more cheaply than human workers. The figure reflects what 1,580 researchers expected in December 2024. It is an elicited expectation, not a measured probability.
Why is the forecast for full automation of labor so much later?
The same survey gives full automation of labor a 50% chance by 2098, 56 years after the 2042 date for high-level machine intelligence. The two forecasts use different definitions, one based on tasks and one on occupations, and they were answered by different subsets of respondents through different question sequences. The survey does not isolate how much of the difference each factor explains.
Does the way a question is asked change AI forecasts?
Yes. The survey randomly assigned respondents to one of two formats. Asking for the year that matches a given probability produced earlier forecasts than asking for the probability by a given year, for both milestones. The authors report seeing this difference in every edition of the survey.
Did AI researchers predict an 18% chance of human extinction?
Not exactly. Among 744 respondents, the mean probability given to human extinction or similarly permanent and severe disempowerment of the human species was 18.3%, and the median was 10%. The question combines two outcomes, so it does not show the probability assigned to extinction alone. The figure is an average of subjective estimates, not an observed risk rate.
What is the BBGK Forecast-to-Decision Check?
It is a five-step test to run before an expert forecast is used to justify a decision: define the event, inspect how the number was produced, separate forecast from observed evidence, map institutional exposure, and set triggers and accountable owners. It is a BBGK working tool, not a validated instrument.
Research note. Every figure attributed to the AI Impacts survey was checked against the text of the published 48-page report on 17 November 2026. BBGK did not recompute any estimate from respondent-level data. The size of the format effect is read from a chart and is approximate. Statements about the Pew, MIT FutureTech and International AI Safety Report publications were checked against the publishers’ own pages. The Forecast-to-Decision Check is a proposed tool, not a tested intervention.
Sources and further reading
- Grace, K., Simonelli, J., Weinstein-Raun, B., Krueger, D., Young, N., Duffy, T., Adamczewski, T., Reeve, J. and Beck, A. (2026). Advanced AI according to 1,580 researchers: uncertain, unsafe, and sooner than we thought. AI Impacts, September 2026. Read the source. Primary expert survey; recruitment p. 4, definitions p. 5, forecasts p. 12, impact p. 21, risk pp. 23 and 27, limitations p. 34, format effect p. 42, aggregation p. 43
- Grace, K. (2026). What did AI researchers think at the end of 2024? AI Impacts blog, 25 September 2026. Read the source. Lead author’s summary; not independent
- Grace, K. (2025). FAQ: Expert Survey on Progress in AI methodology. AI Impacts blog, 31 October 2025. Read the source. Lead author’s defense of the method
- Adamczewski, T. (2024). How should we analyse survey forecasts of AI timelines? December 2024. Read the source. Methodological reanalysis; author later joined the 2024 paper
- McClain, C., Kennedy, B. et al. (2025). How the U.S. Public and AI Experts View Artificial Intelligence. Pew Research Center, 3 April 2025. Read the source. Separate survey; 1,013 U.S. experts, fielded August to October 2024
- Saeri, A. K., Graham, J., Noetel, M., Slattery, P. et al. (2026). Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts. MIT FutureTech and University of Queensland, June 2026. Read the source. Separate Delphi study; five-year horizon
- International AI Safety Report 2026. Published 3 February 2026. Read the source. International evidence synthesis
- National Institute of Standards and Technology (2023). AI Risk Management Framework (AI RMF 1.0). Read the source. Voluntary governance framework
Editorial disclosure. This article evaluates the publicly reported methods and findings of the cited studies and proposes a BBGK working tool. The interpretation is BBGK’s own, not a claim made by AI Impacts, Pew Research Center, MIT FutureTech, NIST or the authors of the International AI Safety Report. BBGK has no relationship with any of them.