Expert Survey on Progress in AI (ESPAI)
SPAR Final Report

Abstract

The Expert Survey on Progress in AI (ESPAI) has been conducted four times since 2016, polling AI researchers on when they expect key milestones, such as High-Level Machine Intelligence (HLMI) and Full Automation of Labor (FAOL), to be achieved. It also asks how they assess the risks and benefits of advanced AI, including the probability of human extinction or similarly permanent and severe disempowerment. In this final report we present three complementary analyses of the ESPAI data. First, we compare the 2024 survey responses across geographic regions (United States, China, and Europe), finding broadly similar views with modest but informative differences. Second, we track how expert predictions have shifted across the 2016, 2022, 2023, and 2024 waves, documenting a dramatic acceleration in aggregate HLMI timelines and an overall increase in safety concern. Third, we apply clustering methods to the 2024 cleaned responses to describe broad regions of a continuous belief landscape, including a notable “polarized/bimodal” group that assigns high probability to both extremely good and extremely bad outcomes.

Introduction

Understanding how AI researchers think about the trajectory and risks of their field is important for forecasting and for informing policy responses to advanced AI. The Expert Survey on Progress in AI (ESPAI), conducted by AI Impacts, is the largest recurring survey of AI researchers on these questions, with waves in 2016, 2022, 2023, and 2024. Each wave asks respondents to estimate timelines for AI milestones, to assess the probability and severity of risks from advanced AI, including human extinction or similarly permanent and severe disempowerment, and to indicate their views on the importance of safety research.

The 2024 wave received 1,580 report responses. In this report, we analyze the 2024 data alongside the three prior waves. Our contributions are:

• Regional comparisons of the 2024 respondents, disaggregated by country (US, China, Europe), examining differences in risk perception, outcome expectations, and HLMI timelines.
• Cross-year trend analysis across all four survey waves, quantifying how HLMI/FAOL timelines, task-specific predictions, value assessments, extinction/disempowerment risk estimates, and safety attitudes have evolved from 2016 to 2024.
• Cluster analysis of the 2024 respondents, using Gaussian Mixture Models to describe broad groupings in belief space and examining how these groups differ on timelines, safety concern, and other variables.

Related Work

This work builds directly on the ESPAI survey series:

• Grace et al. (2018) — 2016 wave.
• Stein-Perlman et al. (2022) — 2022 wave.
• Grace et al. (2024) — 2023 wave.

Complementary survey efforts include the AI Index Report (Maslej et al., 2024), which tracks AI progress along technical and societal dimensions, and Zhang et al. (2022), who surveyed machine learning researchers on ethics and governance of AI.

Methods

Data

We use cleaned and anonymized response data from all four ESPAI waves. For modern waves, the cleaned files retain some unfinished rows. For 2024, we use the agreed report-response denominator (N=1,580) for audience-facing sample-size prose while retaining the 1,793-row cleaned analysis file for item-level analyses and upstream comparability. Each wave asked overlapping but not identical sets of questions. Block randomization in the 2023 and 2024 waves means most respondents answered only a subset of questions, so figures and tables report item-level sample sizes where possible. Respondent location was inferred from IP geolocation at the time the survey was taken; it should not be read as citizenship, affiliation, or home institution. IPs likely to reflect travel, conferences, or VPN artifacts were excluded from regional analyses.

Timeline Aggregation

HLMI and task timelines were elicited in two framings: year-framing (respondents provide years until 10%, 50%, and 90% probability) and fixed-year/probability-framing (respondents provide probabilities at fixed future horizons). Following Grace et al. (2024), we fit a gamma CDF to each individual’s three data points, evaluate those CDFs on a shared year grid, average the probability values at each year, and read percentiles from the aggregate mean CDF. Thus, a reported gamma-CDF “median” is the 50th percentile of the aggregate CDF, not the median respondent’s direct 50% answer. The Cross-Year Trends HLMI panel uses direct HLMI blocks only; final-occupation prediction blocks are not pooled into that headline HLMI series.

The regional comparison section uses a simpler approach: the raw median of year-framing respondents’ 50%-probability estimates. This yields shorter timelines (e.g., 15 years vs. 18 years for 2024) for two reasons: (1) fixed-year/probability-framing respondents tend to give longer estimates and are excluded from the raw median, and (2) the mean CDF is pulled rightward by long-tail predictions. The two methods answer different summary questions and should not be compared as if they were the same estimand.

Clustering

We apply Gaussian Mixture Models (GMM) with full covariance to describe broad regions of respondent belief space. Because block randomization means most respondents only answered a subset of questions, we run multiple focused analyses on overlapping subsets rather than one clustering requiring all features. Model selection uses BIC with silhouette score as a secondary criterion.

Results

Researchers in the US, China, and Europe hold broadly similar views, with modest regional differences

Respondents were classified by the country in which they took the survey^[1]; the largest groups were in the US, China, and throughout Europe. Responses across location were fairly similar. On the question about the probability of AI causing human extinction or similarly permanent and severe disempowerment, researchers in China had a somewhat heavier upper tail than their American and European counterparts, though the regional distributions overlap substantially.

Respondents were also asked to evaluate whether High-Level Machine Intelligence (HLMI) would be beneficial or harmful. The three regions returned remarkably similar distributions, with about half of respondents in each assigning probability mass to good or extremely good HLMI outcomes and a quarter to a third assigning probability mass to bad or extremely bad outcomes. Within that narrow band, respondents in China assigned the highest mean probability to positive HLMI outcomes and those in Europe the lowest, while China also had higher estimates on the extinction-or-disempowerment question. This suggests value-of-HLMI assessments and extinction/disempowerment risk estimates are related but not interchangeable. Mean estimates for each outcome are shown below.

Regarding timelines, the year-framing distributions overlap heavily at shorter timelines, while the upper tail is longer for US and European respondents than for respondents in China. Among researchers with shorter-than-median timelines, estimates were similar across regions. But among those with longer timelines, Americans and Europeans predicted HLMI significantly further out than Chinese researchers. That is, the optimistic ends align, but the cautious ends diverge.

HLMI timelines have shortened in recent survey waves

Expectations for the arrival of HLMI were polled using two methods: estimating the number of years until a 10%, 50%, or 90% probability of arrival, and estimating the probability of arrival within fixed future horizons.

Under both methods, timelines actually extended slightly between 2016 and 2022, but have seen significant shortening each year from 2022 to 2024.  

The probability-framing results are shown above. The year-framing question, which asks directly how many years until each probability threshold, yields more aggressive timelines; both methods agree on shortening across recent surveys, and the year-framing analysis appears in the Cross-Year Trends section.

Expected timelines for individual tasks have fallen unevenly

The AI Impacts survey was first conducted in 2016, allowing for an analysis of how task-specific timelines have evolved. For tasks originally expected to take more than a decade, timelines have shortened significantly—particularly in math and language. However, evolution has been mixed for other domains. In many cases, expected dates have actually receded, most notably for robotics-related tasks.

1. HLMI & FAOL Timeline Trends

Aggregate-CDF 50th percentile calendar year estimates for High-Level Machine Intelligence and Full Automation of Labor. Computed via gamma CDF fitting on both year-framing and fixed-year/probability-framing respondents.

On this measure, direct HLMI estimates moved from 2067 (51.0 years after the 2016 survey) to 2042 (18.4 years after the 2024 survey), a 25-calendar-year earlier estimate. FAOL moved from 2139 to 2096, a 43-calendar-year earlier estimate.

The line is the 50th percentile of the aggregate mean CDF. The shaded band is a robust visual band from the 10th and 90th percentiles of the pointwise median CDF; the mean-CDF 90th percentile is highly sensitive to long right tails and can fall extremely far in the future.

Aggregate probability distributions across surveys

Each curve is the mean mixture CDF: gamma CDFs fitted per respondent, evaluated on a shared year grid, and averaged at each year. The horizontal axis is calendar year; the vertical axis is aggregate probability. Where each curve crosses the dashed 50% line is the 50th percentile of that survey's aggregate CDF. Curves that sit further to the left indicate sooner expectations.

The 2016 survey used somewhat different question wording and was administered to a different population, so the 2016 curve is not strictly apples-to-apples with the 2022/2023/2024 curves and should be read as indicative rather than directly comparable. The 2016 direct-HLMI estimate here is later than the original published 2061 estimate; local validation points to residual 2016 cleaning/respondent-set differences rather than the aggregation code as the likely source.

2. Task Prediction Trends

How the aggregate-CDF 50th percentile calendar year for each AI milestone has shifted. Each arrow shows where a task's prediction started (2016 survey, purple dot) and where it ended up (2024 survey, square). Green = prediction moved sooner, red = moved later. Only the 32 tasks present in all survey years are shown.

Magnitude of Shift

The same data as above, but focused on the size of the change. Green bars = sooner predictions in 2024 vs 2016 (experts expected faster progress on that task). Red bars = later predictions.

Biggest Movers Over Time

Full Table

Entries are aggregate-CDF 50th percentile calendar years. Shift is 2024 minus 2016; negative values indicate earlier expected completion in the 2024 survey. These are not raw respondent medians; they are read from pointwise-mean aggregate gamma CDFs fitted to each respondent's task-forecast inputs from whichever framing they received: either three year-framing answers for 10%, 50%, and 90% probability (ta_*) or three fixed-year/probability answers at that wave's task horizons (tb_*).

3. Value of HLMI Over Time

Respondents assign probabilities to five impact categories (summing to 100%). Bars show the mean probability assigned to each category.

"Extremely Bad" Outcome Trend

Mean probability assigned to the worst-case value-of-HLMI category, which the survey describes with examples such as human extinction, across survey waves. Note the spike in 2022 followed by a partial retreat.

Polarized Outlook: Probability on Both Extremes

Beyond the mean of each category, a persistent minority of respondents assign substantial probability to both an extremely good and an extremely bad outcome at once — a worldview in which HLMI is a high-stakes gamble rather than a clearly positive or negative development. The share doing so has stayed remarkably stable across waves.

4. Extinction/Disempowerment Risk Estimates

Probability of AI causing human extinction or similarly permanent and severe disempowerment, across survey years. Not all questions were asked in all years.

Question-framing variants

Bars compare median estimates across available extinction/disempowerment question framings. The unconditional and control-problem framings are available from 2022 onward; the within-100-years framing is available in 2023 and 2024. These are related but distinct questions, so their estimates should not be treated as one interchangeable series.

Extinction/Disempowerment Risk Thresholds Over Time

Percentage of respondents assigning at least 10% or 25% probability to AI-caused human extinction or similarly permanent and severe disempowerment, shown separately for each available question framing.

Mean vs Median Extinction/Disempowerment Risk Over Time

Mean and median probability assigned to AI causing human extinction or similarly permanent and severe disempowerment (unconditional question).

Are the year-over-year shifts statistically significant?

Each row compares two survey waves of the same question framing with a two-sided Mann–Whitney U test (Wilcoxon rank-sum). Because each wave is a different set of respondents, the samples are independent and an unpaired rank test is used; it makes no normality assumption and is robust to the heavy right-skew and the clustering of answers at round numbers. p (BH) is the Benjamini–Hochberg false-discovery-rate–adjusted p-value across the whole family of comparisons below (* marks adjusted p<0.05). At these sample sizes a small p-value is almost guaranteed for any real shift, so the more informative column is Cliff's δ, the effect size: it ranges from −1 to +1, a positive sign means the later wave gave higher estimates, and the magnitude label follows Romano et al. (negligible/small/medium/large).

A significant result means these samples differ; it does not by itself establish that the underlying expert population's beliefs moved, because the waves are not a panel and recruitment differs across years. Read significance together with the effect size and the composition caveats.

Extremely bad outcomes vs extinction/disempowerment

Across overlapping waves (2022-2024), the mean value-of-HLMI probability assigned to "extremely bad" fell from 14.1% to 9.9% after the 2022 spike, while the mean unconditional extinction/disempowerment estimate rose from 15.8% to 18.3% and the median rose from 5.0% to 10.0%. Within the same respondents, the two measures are positively correlated in each modern wave (ρ=0.41 to 0.47), but their aggregate trends are not identical. The value-of-HLMI item is a broad distribution over how good or bad HLMI's long-run effect on humanity would be, while the extinction/disempowerment item asks directly about one severe risk channel.

5. Safety Attitudes Over Time

How researcher views on AI safety have evolved across survey years.

Safety Research Prioritization

"How much should society prioritize AI safety research, relative to how much it is currently prioritized?"

6. AI Capabilities 20 Years After Each Survey

Two-year comparison for the capability-likelihood block. The 2023 wave asked about AI systems in 2043, while the 2024 wave asked about AI systems in 2044, so this should be read as a comparison of 20-year-ahead expectations, not a full four-wave trend. The plotted measure is the share rating each capability Likely or Very likely.

Across the 11 items, the average change was +3.6 percentage points. The largest increase was for Cause important real-world actions (run business, etc.) (+10.1 pp), while the largest decrease was for Frequently behave surprisingly to humans (-5.0 pp).

Change from 2023 to 2024

Bars show the 2024 percentage rating Likely or Very likely minus the 2023 percentage. Positive values mean the capability was rated more likely in 2024.

Mean score uses the 1-5 likelihood scale, from Very unlikely (1) to Very likely (5). The 2023 cleaned file stores this block as numeric 1-5 values; 2024 stores text labels, normalized here to the same scale.

7. Intelligence Explosion Feedback Loop

Distribution of responses to the repeated categorical question asking whether the AI R&D feedback-loop argument is broadly correct.

Computed from each year's cleaned ie_3 column. Response labels are shown in increasing likelihood order.

8. Sample Sizes

Cleaned files may retain unfinished rows. "Report responses" is the main respondent denominator used in audience-facing sample-size prose; it follows a configured headline count where available and otherwise uses finished responses where a completion indicator exists. Question-level analyses use their own item-level sample sizes. HLMI columns below count respondents with at least one usable answer in the year-framing or fixed-year/probability-framing HLMI block. "Task items with data" counts task questions with at least one usable year-framing response, not respondents.

A data-driven exploration of broad groupings and the safety divide among 1,793 ESPAI 2024 cleaned analysis rows.

Executive Summary

1. How Many Camps? Broad Value-Outlook Groupings

1.1 Clustering on Value Outlook (n=1,538)

The value-of-HLMI question asked respondents to assign probabilities (summing to 100%) across five outcomes: extremely good, on balance good, neutral, on balance bad, and extremely bad. It is the highest-coverage feature in this analysis. We cluster on these 5 dimensions using Gaussian Mixture Models.

Figure 1: Silhouette scores for K=2 through K=6 clusters on value outlook. K=2 has the highest silhouette, but K=3 and K=4 offer more interpretable structure.

Silhouette scores are modest (0.10-0.12), indicating the groups are not sharply separated. Treat these as broad regions of a continuous opinion landscape, not discrete camps.

1.2 Four-Camp Solution

Figure 2: Four natural camps in value outlook. Left: PCA projection. Center: mean probability profiles. Right: cluster sizes.

The most striking finding here is the Polarized/Bimodal group. These respondents assign high probability to both extremely good and extremely bad outcomes -- their P(extremely good) is comparable to the Strong Optimists, yet they simultaneously assign substantial probability to catastrophic outcomes. This is not a group of doomers or technophobes. They are researchers who believe AI will be hugely impactful, but who are genuinely torn on whether that impact will be positive or negative. The key axis for this group is magnitude of impact, not direction -- they have rejected the possibility that AI will be a modest or neutral development.

1.3 Three-Camp Solution (Simpler View)

Collapsing to 3 clusters (which has a higher silhouette score of 0.18 vs 0.07) merges the finer distinctions into a simpler high/moderate/low value-outlook framing. This loses the polarized group but provides a cleaner summary:

Figure 3: Three-camp simplification. The polarized group is absorbed into the moderate/low-value clusters.

1.4 Intuitive View: How Soon vs. How Good

The PCA axes above lack intuitive meaning. Below, we plot each respondent on two directly interpretable dimensions: their predicted HLMI arrival year (x-axis) and their net optimism score (y-axis), defined as P(good + extremely good) minus P(bad + extremely bad). This reveals where the broad value-outlook regions sit in the space of "how soon" vs. "how beneficial."

Figure 3b: Respondents (n=935) binned by predicted HLMI year and net optimism. Each bubble's size shows how many researchers from that cluster fall in the bin (labeled when ≥5). The vertical dashed line marks the median predicted year; the horizontal line separates positive net-outlook respondents from negative net-outlook respondents.

2. Higher vs Lower Safety/Severe-Risk Concern

2.1 Defining the Groups

We built a safety composite score from (normalized 0-1 and averaged):

• Alignment-problem importance rating (0-4 ordinal scale: not a real problem to among the most important problems in the field)
• Value of working on the alignment problem today compared with other AI problems (0-4 ordinal scale)
• P(extinction/disempowerment from AI) (0-100%)

Requiring at least 2 of 3 components, we obtained scores for 754 respondents and split at the median (0.500) into High (n=330) and Low (n=424) safety concern groups.

2.2 The Full Comparison

Figure 5: Comprehensive comparison of higher vs lower safety/severe-risk-concern respondents across six dimensions. Alignment-problem importance uses the 0-4 ordinal scale above; mean concern averages 11 scenario concern ratings on a 0-3 scale; the optimism-maximizing AI progress-rate answer is coded 0=much slower, 2=current speed, 4=much faster.

2.3 Statistical Tests

Selected Mann-Whitney U tests for scalar summaries from Figure 5. The value-outlook panel is summarized by P(extremely good) and P(extremely bad), the concern panel by mean concern, and the AI-capabilities panel is tested item-by-item in the next table. Effect size r is rank-biserial correlation (|r| > 0.3 = medium, |r| > 0.5 = large). Scale notes: mean concern 0-3, optimism-maximizing AI progress-rate answer 0-4, probabilities in percentage points. *** p<0.001, ** p<0.01, * p<0.05

2.4 AI Capabilities by 2044: What Do Higher Safety/Severe-Risk Concern Respondents Expect?

Safety-concerned people don't just worry more -- they have systematically different expectations for what AI will be able to do by 2044:

Scale: 0=Very unlikely, 1=Unlikely, 2=Even chance, 3=Likely, 4=Very likely. High - Low is the high-safety mean minus the low-safety mean, so positive values mean high-safety respondents rated the capability as more likely. Hypothesized direction is marked for risk-relevant capabilities where we expected High > Low; other rows are exploratory.

2.5 Intelligence Explosion Beliefs

Figure 6: Probability estimates for intelligence explosion scenarios, by safety concern level.

3. How Outlook Predicts Everything Else

The three value-outlook clusters (from Section 1) predict views across many other dimensions:

Values are medians with sample sizes in parentheses, except HLMI Year, which shows median [IQR], mean, and item-level n because several groups share the same median year. HLMI Year is an interpolated individual 50% calendar-year estimate, not the aggregate gamma-CDF median used in the cross-year section. Scale notes: alignment-problem importance 0-4, mean concern 0-3, optimism-maximizing AI progress-rate answer 0=much slower to 4=much faster, probabilities in percentage points.

Figure 8: How the four value-outlook camps compare across timelines, extinction/disempowerment probability, safety/alignment-problem views, concern levels, optimism-maximizing AI progress-rate answers, and P(extremely bad). Mean concern averages 11 scenario concern ratings on a 0-3 scale; alignment-problem importance and optimism-maximizing AI progress-rate answers use 0-4 ordinal scales.

4. The Full Correlation Structure

Figure 10: Spearman rank correlations between key variables. Stars indicate significance. Sample sizes shown in each cell. Alignment-problem importance uses a 0-4 ordinal scale from not a real problem to among the field's most important problems; value of working on the alignment problem today uses a 0-4 ordinal scale from much less valuable to much more valuable than other AI problems; mean concern uses a 0-3 scale; HLMI Year is a calendar-year estimate; probabilities are 0-100 percentages.

Key Pairwise Correlations

Discussion

Several findings stand out across the three analyses. First, HLMI timelines have shortened dramatically—the direct-HLMI aggregate-CDF 50th percentile moved from 2067 (51 years after the 2016 survey) to 2042 (18 years after the 2024 survey)—and this acceleration is consistent across both elicitation framings, though the magnitude differs. Task-level predictions have moved unevenly: mathematical and language tasks have seen the largest acceleration (consistent with rapid progress in LLMs), while robotics tasks have actually receded, suggesting experts view embodied AI as a distinct and harder challenge.

Second, safety concern has risen steadily. The share of respondents favoring more safety research increased from 49% (2016) to 71% (2024), and the median unconditional extinction/disempowerment risk estimate has risen from 5% (2022) to 10% (2024), where this question asks about human extinction or similarly permanent and severe disempowerment caused by AI. The broader value-of-HLMI "extremely bad" item moved differently: it spiked in 2022 and then partially retreated by 2024, even though it remains positively correlated with extinction/disempowerment estimates within each modern wave. Notably, the cluster analysis reveals that higher extinction/disempowerment risk estimates are correlated with shorter HLMI timelines (ρ=−0.21), suggesting that the most concerned researchers are those who believe powerful AI is arriving soon—not those with abstract, long-horizon worries.

Third, regional differences are more modest than is sometimes assumed. US and Chinese researchers hold broadly similar views on timelines and risks, with Chinese researchers assigning higher mean probability to positive HLMI outcomes while also having a heavier upper tail on the extinction/disempowerment question. European researchers assign the lowest mean probability to positive HLMI outcomes, but are not the highest on extinction/disempowerment risk, suggesting that value-of-HLMI assessments and extinction/disempowerment risk estimates are related but partially independent attitudes.

Limitations

Our analyses have several limitations. The cluster analysis silhouette scores (0.08–0.12) indicate soft, overlapping groupings rather than discrete camps; the four-group framing is an interpretive convenience. Block randomization means not all respondents answered all questions, limiting cross-cutting analyses. We do not control for demographics (career stage, subfield, institution) that may confound clustering. The methodological difference between raw medians and gamma CDF estimates, while documented, means that readers must attend to which method underlies a given figure.

Conclusion

This report presents three complementary views of expert opinion on AI progress and risk, drawing on four waves of the ESPAI survey (2016–2024). We document a pronounced acceleration in HLMI timeline expectations, an overall increase in safety concern, and a belief landscape that can be categorized into four broad, overlapping regions whose views on timelines, risk, and safety tend to differ systematically. Regional variation exists but is modest relative to the overall distribution. These findings suggest that the AI research community is converging toward shorter timelines and greater concern. We believe this shift is relevant for policymakers, funders, and safety researchers planning on longer horizons.

References

1. Grace, K., Salvatier, J., Dafoe, A., Zhang, B., & Evans, O. (2018). “When Will AI Exceed Human Performance? Evidence from AI Experts.” Journal of Artificial Intelligence Research, 62, 729–754.
2. Grace, K., Stewart, H., Sandbrink, J. B., Thomas, S., Weinstein-Raun, B., & Brauner, J. (2024). “Thousands of AI Authors on the Future of AI.” arXiv:2401.02843.
3. Maslej, N., et al. (2024). The AI Index 2024 Annual Report. Stanford Institute for Human-Centered AI.
4. Stein-Perlman, Z., Weinstein-Raun, B., & Grace, K. (2022). “2022 Expert Survey on Progress in AI.” AI Impacts.
5. Zhang, B., Anderljung, M., Kahn, L., Dreksler, N., Horowitz, M. C., & Dafoe, A. (2022). “Ethics and Governance of Artificial Intelligence: Evidence from a Survey of Machine Learning Researchers.” Journal of Artificial Intelligence Research, 71, 591–666.