AI Forecasting

In the AI Story

The historical record of AI forecasting is a mixed one. Expert surveys in the 1960s predicted AGI in a decade; surveys in the 1980s, after the second AI winter, predicted it in centuries; surveys in the 2010s converged on a median of thirty to fifty years. As of 2025 expert median estimates have compressed to under a decade for several benchmarks previously considered out of reach. The compression is not itself evidence of imminent arrival; it is evidence that expert estimates move in response to recent progress more than they should under a Bayesian analysis of long-run base rates. The famous 1973 Lighthill report to the UK government predicted that general-purpose AI was a dead end; the 2022 US National Security Commission on AI concluded that general-purpose systems within the decade were credible enough to justify national-scale investment. Neither report was obviously wrong at the time it was written; both illustrate how sensitively "expert consensus" depends on the immediate pre-report state of the field.

Clarke's First Law — "When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong" — names the specific failure mode that has dogged AI forecasting. Confident declarations of impossibility have been repeatedly falsified (Go, protein structure, long-form conversation, commercially useful code generation, image-from-text synthesis). Confident declarations of imminent success have also been falsified, often by longer timelines than predicted. The asymmetric pattern Clarke identifies — that the impossibility claim is more wrong more often than the possibility claim — is observable in the AI record with some qualifications.

The methodologies now in use include: scaling-law extrapolation (given compute-loss relationships, project when a given loss yields a given capability); benchmark time-series (track how quickly specific benchmarks are saturated and extrapolate); expert elicitation (structured surveys of researchers, with calibration checks); public prediction markets (Metaculus, Manifold, Polymarket, Kalshi where permitted); and trend analysis of AI R&D throughput itself. Each method has characteristic failure modes: scaling laws assume architectural constancy; benchmarks saturate in ways that distort extrapolation; experts anchor on recent progress; markets have thin liquidity and fat-tail distortions. Combining methods improves on any single method but does not eliminate the fundamental uncertainty.

The decision-relevant question is not "when will AGI arrive" but "what actions are robust to a wide range of arrival times." Governance-design, safety research, compute policy, and educational investment all look different under the extreme timelines; actions that are valuable across most of the probability distribution are the operationally important ones to identify. The forecasting literature has shifted in the last two years toward this framing: less prediction of specific dates, more identification of actions whose value is preserved under a plausible range of arrival scenarios.

Origin

Structured AI forecasting begins with the 1973 Lighthill report and the contemporaneous Dreyfus critique; it professionalizes in the 2010s with Stuart Armstrong and Kaj Sotala's How We're Predicting AI (2012), Katja Grace's AI Impacts program (2015 onward), and Open Philanthropy's internal forecasting work (2016 onward). The forecasting-market era begins with Metaculus's AI timelines questions (2017 onward). The 2022 expert surveys by Grace et al. and the 2023 survey by Stein-Perlman et al. are the most-cited recent datasets.

Key Ideas

Clarke's asymmetry holds. Impossibility claims about AI capabilities have been falsified more often than timeline claims; the base rate of "this will never work" surviving a decade is low.

Expert consensus is recency-weighted. Median estimates move with recent progress more than Bayesian rationality permits; this is a known bias, not a sign of genuine update.

Scaling laws are the firmest extrapolation. Compute-loss relationships have held across five orders of magnitude and remain the single most reliable forecasting tool; their capability-loss translation is less reliable.

Robustness beats precision. The decision-relevant output of forecasting is a set of actions that are valuable across scenarios, not a specific date.