Forecasting advanced AI is method-dependent and often discordant: experts, markets and compute-based extrapolations each have predictable blind spots. The authors map these weaknesses and offer decision frameworks that blend methods and weight reliability to guide policymakers facing deep uncertainty about AGI.

Main Finding

Recent AGI forecasts across independent methods (expert surveys, prediction markets, and compute‑centric models) have shifted materially earlier, often into the 2030s, but forecasting infrastructure remains immature and unreliable for precise timing. Forecasts are most useful as scenario‑structuring tools to guide flexible, adaptive, and robust economic and policy responses rather than as point estimates to optimize around.

Key Points

• Timeline compression: Multiple methods show a directional shift toward earlier AGI/HLMI (high‑level machine intelligence) arrival relative to a few years prior. Magnitude varies by definition (HLMI, full automation of labor, transformative AI).
• Definitional ambiguity: Much disagreement arises from different AGI definitions and targets (capability vs. deployment vs. societal transformation). Even holding definitions constant, substantial expert disagreement persists.
• Two types of value from forecasts:
  • Predictive signal from measurable inputs (compute trends, hardware investment, scaling laws, market bets).
  • Scenario value: Forecasting synthesizes diverse information to surface plausible, consequential futures for planning.
• Forecasting infrastructure is immature:
  • Lack of robust, hard‑to‑game benchmarks; many benchmarks saturate quickly or suffer training‑data contamination.
  • Insufficient independent validation and stress‑testing of influential models (compute‑centric and takeoff analyses).
  • Low‑frequency monitoring (annual reports) may be too slow under compressed timelines.
• Leading indicators: Degree of AI automating AI research is a critical early signal for rapid capability gains and potential discontinuities.
• Strategic recommendations (high level):
  • Treat forecasts as scenario tools; integrate conditional forecasting with strategic choices.
  • Build adaptive capacity with explicit reassessment triggers tied to observable indicators (e.g., systems autonomously completing multiweek software projects; >50% of AI research automated).
  • Invest in methodological diversity, independent validation, continuous capability evaluation, and higher‑frequency monitoring.
  • Frontier labs should institutionalize internal monitoring of AI‑assisted R&D; broader information sharing may require voluntary coordination or government facilitation.

Data & Methods

• Methods synthesized:
  • Expert surveys (including large recent surveys with shifting medians).
  • Prediction markets.
  • Compute‑centric models that extrapolate hardware trends, training compute, and scaling laws.
  • Scenario analysis and structured deliberation.
  • Benchmark and capability evaluations (public and private).
• Meta‑methodological approach:
  • Iterative human–AI collaboration: primary drafting by LLMs (GPT‑5.1, Gemini 3 Pro, Claude 4.5 Opus) with human direction, review, fact‑checking, and revision.
  • Extensive expert consultation and peer review across academia, industry, and policy practitioners.
• Limitations documented:
  • Sensitivity of timeline estimates to definitional choice and framing (same dataset can yield decade‑scale shifts).
  • Benchmark degradation through saturation and contamination.
  • Sparse independent, ongoing validation of compute‑centric and takeoff models.
  • Heterogeneous signals across methods; no single method provides definitive timelines.

Implications for AI Economics

• Investment timing and capital allocation
  • Earlier timelines increase the option value of near‑term investment in safety, monitoring, and governance capacity.
  • Compute and hardware investment trends are informative signals for private actors; markets may front‑run or amplify capability shifts.
  • Policymakers and funds should prepare for rapid reallocation needs (R&D, workforce programs, social safety nets).
• Endogenous technological change and modeling
  • Models of technological diffusion and economic growth should treat AI capability progress as partly endogenous (AI automating AI → nonlinear acceleration).
  • Use scenario‑based, distributional forecasts (fat tails, discontinuities) rather than single central estimates.
• Labor markets and automation economics
  • Definitions matter: “Capability” arrival (HLMI) vs. “deployment” (full automation) imply different timing for labor displacement and sectoral impacts.
  • Economic models should separate capability thresholds from adoption/deployment frictions (regulation, business incentives, complementarities).
  • Prepare for heterogeneous, rapid shocks (occupation‑ and task‑level), and model transitional dynamics (retraining, reallocations, wage effects).
• Productivity, aggregation, and inequality
  • Rapid capability gains could produce large productivity shocks with uncertain distributional consequences; models should account for uneven adoption across firms and countries.
  • Market concentration risks: frontier labs’ control over capabilities and proprietary benchmarks may amplify returns to scale and increase inequality.
• Risk management and policy design
  • Robust policy under deep uncertainty requires adaptive mechanisms (triggers, staged interventions, contingency planning) and investments that yield value across multiple scenarios.
  • Forecasts should be framed conditionally (How would particular policies change capability trajectories?) to inform cost–benefit and regulatory decisions.
• Measurement, data needs, and empirical work
  • Economists need higher‑frequency, high‑quality indicators: compute deployed for training, private benchmark performance, measures of AI‑assisted R&D, and firm‑level deployment metrics.
  • Investment in independent benchmarking and monitoring infrastructure is a public good that reduces information asymmetries and improves market and policy responses.
• Institutional and market design
  • Markets and institutions should be designed to internalize externalities from rapid AI adoption (safety, systemic risk).
  • Consider mechanisms to incentivize sharing of monitoring metrics (data trusts, regulatory safe harbors, voluntary commitments) to improve macroprudential oversight of AI‑driven economic shocks.

Short takeaway: treat AGI forecasts as scenario inputs for economic models and policy — incorporate uncertainty, conditionality, and high‑frequency indicators; prioritize investments that make economies resilient and adaptable to rapid, potentially nonlinear AI‑driven transformations.