Experts broadly expect major AI capability gains by 2030; if AI progresses rapidly they anticipate a sizable macroeconomic re‑rating—GDP growth rising toward roughly 3.5–4% annually and labour‑force participation falling to about 55% by 2050 (equivalent to ~10 million jobs), while policy preferences split: economists favor targeted retraining whereas the public backs broader safety‑net measures.

Main Finding

A broad set of experts expect substantial AI capability gains by 2030, but unconditional macro forecasts remain close to recent historical trends. If a “rapid” AI-capabilities scenario occurs (AI surpasses humans on most cognitive and physical tasks), experts forecast large but not unprecedented economic changes: annual GDP growth rising to roughly 3.5–4% and the U.S. labor-force participation rate (LFPR) falling from ~62% today to ~55% by 2050, with roughly half of that LFPR decline (≈10 million jobs) attributable to AI. Disagreement among experts is driven more by differing beliefs about the economic effects of highly capable AI than by disagreement about the pace of capability progress. Policy preferences differ: economists favor targeted interventions (e.g., retraining) while the public is more supportive of broad programs (e.g., UBI, job guarantees).

Key Points

• Median unconditional forecasts (all groups) expect:
  • Substantial AI capability advances by 2030 (average economist assigns ~61.4% probability to moderate or rapid progress).
  • Annual GDP growth ≈ 2.5% (higher than many government/private medium- and long-run baselines of ~2.0% and 1.7%).
  • Small LFPR declines consistent with demographics (no large unconditional labor-market collapse).
• Conditional on a "rapid" AI scenario:
  • Experts forecast meaningful macro shifts: GDP growth ≈ 3.5–4% annualized; LFPR ≈ 55% by 2050.
  • Wealth concentration increases: the top 10% share could rise to ~80% by 2050 in the rapid scenario.
  • About half of the projected LFPR decline is attributed directly to AI (≈10 million jobs).
• Unconditional consensus is relatively tight, but uncertainty explodes under the rapid scenario — experts disagree substantially about possible outcomes if AI is transformative.
• Variance decomposition: most cross-forecaster disagreement about long-run outcomes stems from divergent beliefs about the economic impacts of high-capability AI (diffusion, labor substitution/complementarity, distributional effects), not mainly from disagreement about whether AI capabilities will advance.
• Policy preferences:
  • Economists strongly prefer targeted measures (e.g., AI-focused worker retraining: 71.8% support) and are less supportive of broad programs (job guarantee: 13.7%; UBI: 37.4%).
  • The general public supports both targeted and broad interventions more than economists do.
• Between-group differences (academics, AI-company employees, policy researchers, superforecasters, public) are smaller than within-group disagreement.

Data & Methods

• Survey-based elicitation (March 2026) of beliefs from five groups: academic economists (including prominent economists and those publishing on AI/economics), employees at frontier AI companies, AI-focused policy researchers (U.S. think tanks/research institutions), highly accurate forecasters (“superforecasters”), and a general public sample.
• Elicited:
  • Unconditional forecasts for key U.S. macro variables (GDP growth, total factor productivity growth, LFPR, wealth inequality) at near-term (2030) and long-term (2050) horizons.
  • Conditional forecasts under three 2030 AI-capability scenarios (slow, moderate, rapid) and each respondent’s subjective probabilities for those scenarios. Scenarios described capability levels (examples below) and respondents were reminded scenarios refer to capabilities, not adoption.
  • Marginal effects of six policy proposals on GDP and LFPR under unconditional and rapid-progress conditions, plus normative support for each policy.
• Three 2030 capability scenarios (summary):
  • Slow: AI as a capable assistant (e.g., PhD-student-level literature reviews; partial automation of short freelance software tasks; basic household robotics).
  • Moderate: AI as effective collaborator (e.g., autonomous lab advances; broad automation of mid-difficulty software tasks; robots and robo-taxis operating at human parity in many settings).
  • Rapid: AI surpasses humans on most cognitive and physical tasks (e.g., autonomous research collapsing R&D timelines, full automation of many white-collar and service occupations, high-quality creative and negotiation abilities, robotics widely effective).
• Analytical methods included conditional forecasting, counterfactual attribution (estimating share of LFPR decline due to AI), and variance decomposition to apportion sources of forecast disagreement between beliefs about capabilities vs. beliefs about economic effects. Full sampling details and appendices reported in the paper.

Implications for AI Economics

• Separate capability forecasts from diffusion/adoption modeling. The paper shows that most uncertainty about macro outcomes comes from how capable AI would be translated into economic structure — models should explicitly treat adoption, institutional constraints, market power, and reallocation frictions.
• Policy design should be conditional and tiered:
  • Prepare scalable targeted measures (retraining, transition assistance, sectoral adjustment programs) given expert preference and evidence that such programs can shorten unemployment spells and raise earnings.
  • Maintain contingency plans for broader redistribution (tax reform, UBI, job guarantees, sovereign wealth approaches) if rapid, widespread automation materializes and labor income shrinks substantially.
• Research priorities:
  • Better empirical measurement of AI diffusion and firm-level adoption; causal impacts on employment, wages, and occupational tasks.
  • Structural work on how AI affects TFP, idea production, and feedback loops (R&D acceleration → long-run growth).
  • Distributional dynamics: mechanisms driving capital/skill concentration, taxation implications, and international spillovers.
• Forecasting and policy analysis should incorporate greater tail-risk considerations: although median unconditional forecasts are modest, the rapid scenario implies large macro and distributional shifts that warrant proactive policy experimentation and monitoring.
• Communication and governance: differences in policy preferences between experts and the public underscore the need for public engagement and transparent trade-off analyses when designing large-scale redistribution or labor-market interventions.

Funding note: research supported by Open Philanthropy (Coefficient Giving).