Forecasting AI Progress

Forecasting AI progress is the practice of making explicit, checkable statements about future AI capabilities, resources, or impacts, and of judging those statements by how well their stated probabilities match later outcomes. Calibration is the property that events assigned probability $p$ occur with relative frequency close to $p$; resolution (or sharpness) is the ability to separate events that occur from those that do not.

AI forecasts inform capital allocation, safety research priorities, and policy. They are also unusually hard: the field changes quickly, the target concepts (for example "AGI") are contested, there are few historical reference cases, and the forecasters are often participants with a stake in the outcome. This article covers the methods and the scoring of forecasts. The companion article AGI-TIME-001 addresses AGI timelines specifically; this article takes the wider view, including benchmark, compute, and economic forecasts, and sets out the standards that the other Volume XII articles apply to themselves.

The central methodological claim is modest: a forecast is only useful to the extent that it can be scored. A statement such as "AI will transform the economy" cannot be wrong, and therefore carries no information. The encyclopedia's forecast record format (docs/EVIDENCE_FRAMEWORK.md, section 2) forces each forecast to name a measurable event, a horizon, assumptions, a confidence level, and a review date.

Probabilistic forecast verification is older than AI. Brier (1950) proposed a squared-error score for weather forecasts expressed as probabilities (Brier, 1950). Tetlock's long-running study of expert political judgment compared many experts' forecasts against outcomes and found that experts' forecasts were often little better than simple baselines in the domains studied (Tetlock, 2005). [VERIFICATION NEEDED: characterisation of Tetlock (2005) findings not checked against the book.] From 2011 a US government-sponsored forecasting tournament led to the Good Judgment Project; Mellers et al. (2014) report psychological interventions (training, teaming, and tracking of top performers) that improved accuracy in that tournament, and Tetlock and Gardner (2015) popularised the resulting "superforecaster" findings.

Forecasting of AI specifically has a history of overconfident timelines (see AGI-TIME-001) and, more recently, of formalised methods: expert surveys (Grace et al., 2024), compute-centric and biological-anchor models (Cotra, 2020; Davidson, 2023), and empirical trend extrapolation of compute and capability measures (Sevilla et al., 2022; Kwa et al., 2025). Since 2024, benchmark projects have also begun to evaluate language models as forecasters (Karger et al., 2024).

#4.1 Proper scoring rules

Let a forecaster assign probability $f_t \in [0,1]$ to event $t$, with outcome $o_t \in \{0,1\}$. The Brier score over $N$ binary forecasts is

where lower is better (Brier, 1950). Always forecasting $0.5$ yields $\mathrm{BS}=0.25$, a useful no-information baseline. The Brier score is a proper scoring rule: expected score is minimised by reporting one's true belief, so there is no incentive to shade probabilities. The logarithmic score $-\frac{1}{N}\sum_t \left[o_t\ln f_t + (1-o_t)\ln(1-f_t)\right]$ is also proper and penalises confident errors more severely.

The Brier score decomposes (Murphy decomposition) into reliability (calibration), resolution, and an irreducible uncertainty term determined by the base rate. [VERIFICATION NEEDED: Murphy (1973) decomposition citation not checked against a source.] A forecaster can be well calibrated yet uninformative (always stating the base rate); calibration alone is therefore not a sufficient quality criterion.

#4.2 Calibration diagnostics

Group forecasts into probability bins and compare the mean stated probability with the observed frequency in each bin (a reliability diagram). With small samples, bins are noisy: with only 20 forecasts at "70%", an observed frequency anywhere in a wide band is statistically compatible with good calibration. For AI milestones, where each question resolves once and years apart, sample sizes are usually far too small for individual forecasters to be assessed, which is a structural limitation of the field.

#4.3 Operational definitions and resolution criteria

A well-posed question specifies (i) the measured quantity, (ii) the resolution source, (iii) the date or window, and (iv) what happens under ambiguity. Many disputed AI forecasts diverge mainly because the target differs. Grace et al. (2024), for example, elicited both a "high-level machine intelligence" milestone (unaided machines outperforming humans at every task) and a separate "full automation of occupations" milestone; the two received very different aggregate dates (see section 8.2).

#4.4 Aggregation

Combining forecasts from several people typically improves accuracy relative to the average individual. Aggregation choices (mean of probabilities, median, geometric mean of odds, extremising) matter, and published surveys state which they use. The Grace et al. (2024) survey reports an aggregate forecast distribution across 2,778 respondents; individual medians differ widely.

#5.1 Inside and outside views

An inside view builds a model from the specifics of the case; an outside view uses the distribution of outcomes in a reference class. Good practice is to state the reference class used and to show how far the inside-view estimate departs from it. For AI, reference-class choice is itself contested (for example, other general-purpose technologies versus other research-driven "moonshot" projects).

#6.1 Expert surveys

Grace et al. (2024) surveyed 2,778 researchers who had published at top AI venues (survey conducted in 2023; arXiv first posted January 2024; published in the Journal of Artificial Intelligence Research, 2025 per the arXiv record). The abstract reports aggregate forecasts of a 10% chance by 2027 and a 50% chance by 2047 that unaided machines outperform humans at every task, and 10% by 2037 and 50% by 2116 for full automation of all human occupations. The 50% date for the first milestone was 13 years earlier than in the authors' 2022 survey, and the 50% date for the second moved from 2164 to 2116 (Grace et al., 2024). These are survey aggregates, not a consensus of the field, and they pre-date developments after the survey window.

#6.2 Superforecasters and experts on AI risk

The Existential Risk Persuasion Tournament (conducted in 2022) compared 80 domain experts with 89 superforecasters, 169 forecasters in all (Karger et al., 2023). The tournament's authors' announcement reports median probabilities of AI-caused human extinction by 2100 of 3% for domain experts and 0.38% for superforecasters, the largest gap between the groups on any topic studied (Karger et al., 2023). [VERIFICATION NEEDED: the medians were taken from the authors' public announcement and secondary coverage, not read from the full working-paper tables, and the paper's definitions of the catastrophe thresholds were not checked.] The example shows that groups with different track records can disagree by roughly an order of magnitude on the same question, and that the disagreement did not close with discussion in the tournament; it does not show which group is correct.

#6.3 Compute and capability trends

Sevilla et al. (2022) estimate that training compute for notable systems doubled roughly every 20 months before 2010 and approximately every 6 months in the deep-learning era (data through the paper's 2022 date). Epoch AI (Sevilla et al., 2024) projected, as of August 2024, that training runs of about $2\times10^{29}$ FLOP would likely be feasible by the end of the decade, identifying power as the constraint most likely to bind first. Kwa et al. (2025), measuring the length of software tasks that models complete at 50% reliability, report a doubling time of approximately seven months since 2019, with a possible acceleration in 2024, and a 50% time horizon of around 50 minutes for Claude 3.7 Sonnet as of the original March 2025 posting. These are empirical trends on specific metrics; extrapolating them is a forecast with its own assumptions (see FCST-NEAR-001).

#6.4 Models as forecasters

ForecastBench contains 1,000 automatically generated and regularly updated forecasting questions; in the evaluation on a random subset of 200, expert forecasters outperformed the top-performing language model with $p<0.001$ (Karger et al., 2024, version of 28 February 2025). Because the benchmark is dynamic, the current standing should be checked on the project's leaderboard rather than inferred from this article.

Strengths. Scoring discipline exposes overconfidence, forces definitions, and makes disagreement concrete. Trend-based forecasts are cheap to check.

Limitations.

• Resolution lag. Questions about 2040 cannot be scored for decades, so forecaster skill must be inferred from short-horizon proxies whose transfer to long horizons is unproven.
• Definition drift. Terms such as AGI change meaning over time; resolution criteria must be fixed in writing at forecast time.
• Non-stationarity. Scaling trends have held over many orders of magnitude on some metrics, but nothing guarantees continuation; a single algorithmic or economic discontinuity can invalidate fitted trends.
• Reflexivity. Forecasts affect investment and policy, and thereby the outcome.
• Survey limits. Respondents are not forecasting specialists; question wording and ordering can shift answers. [VERIFICATION NEEDED: the size of framing effects in the Grace et al. surveys is not quantified here.]

Open questions. How well do short-horizon forecasting skills transfer to technological change over decades? Which reference classes are appropriate for transformative AI? How should model-based forecasts (with many uncertain parameters) be combined with judgemental aggregates?

As of 2026-10-15, the evidence base consists of a small number of large expert surveys, a handful of model-based reports, empirical trend datasets maintained by organisations such as Epoch AI and METR, and tournaments of limited size. The best-supported methodological conclusions are the general ones from forecasting research (proper scoring, explicit definitions, aggregation) rather than any particular AI timeline. There is no agreed forecast of AGI arrival; stated dates depend strongly on the definition and the method (see AGI-TIME-001). Biological-anchor results illustrate the sensitivity: a widely cited summary of Cotra's report states probabilities of greater than 10% for transformative AI by 2036, roughly 50% by 2055, and roughly 80% by 2100, while also noting that the report itself suggests it may be too aggressive for the next few years and too conservative beyond about fifty years (Karnofsky, n.d.). [VERIFICATION NEEDED: figures taken from a secondary summary, not from the report; the report's later revisions and the author's subsequent updates were not checked.]

#8.1 Standards used in this encyclopedia

1. Every forecast has a record with ID, statement, operational definition, horizon, assumptions, confidence, alternatives, and review date.
2. Confidence is stated as a probability or a bracket; where a bracket reflects editorial judgment rather than a computed aggregate, the record says so.
3. Dated numbers are kept in Living References (FCST-REG-001, MODEL-001, TIME-001).
4. Scenarios are labelled as scenarios and carry no probability unless a cited source supplies one.

#8.2 Worked example: why definitions matter

Two milestones in Grace et al. (2024) share a survey and respondents yet have 50% dates of 2047 and 2116, a gap of 69 years, reflecting the difference between machines being able to perform every task better than humans and every occupation being fully automated. A reader quoting only "2047" without the definition misstates the survey.

Near term (0-3 years). Forecasting infrastructure is likely to improve: more dynamic benchmarks, more public track records, and more resolvable questions tied to measured quantities. Expected developments, not firm predictions, include continued use of language models as forecasting assistants and further evaluation of them against human baselines.

Medium term (3-10 years). Long-horizon questions begun now will start to resolve, allowing the first direct tests of whether survey medians and model-based dates were well calibrated.

Long term (10+ years). Not meaningfully forecastable at the method level; the relevant point is that records opened now will be the evidence for evaluating method choices later.

#Forecast record

• Foundational Concepts:
  • AGI-TIME-001 AGI Forecasting & Timeline Methodology: AGI-specific timeline methods; this article generalises and sets scoring standards.
  • AGI-001 Artificial General Intelligence: The most commonly forecast target.
  • CAP-BENCH-001 AI Evaluation, Benchmarking & Contamination: Measurement problems that carry into benchmark-based forecasts.
• Related Methods & Technologies:
  • ECON-COMP-001 The Economics of Compute & Scaling: Compute-cost and scaling inputs to model-based forecasts.
  • TECH-SCALE-001 Scaling Laws: Empirical regularities used for extrapolation.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: A hypothesis that, if true, would break trend extrapolation.
  • FCST-NEAR-001 Near-Term Outlook (0-3 years): Living record applying these methods.
  • FCST-SCEN-001 Medium- and Long-Term Scenarios: Scenario method and its limits.
  • FCST-SIGNALS-001 Signals, Indicators and Inflection Points: Indicators for updating forecasts.
  • FCST-UNCERT-001 The Uncertainties That Matter Most: Where forecast disagreement originates.
  • FCST-REG-001 Forecast Register: Living register of forecast records.