Advances in large language models speed professional work: each year of model progress reduces task time by about 8%, driven roughly 56% by added compute and 44% by algorithmic gains. The benefits concentrate in non-agentic analytical tasks, and if current scaling continues the authors estimate roughly a 20% boost to U.S. productivity over ten years—an extrapolation that rests on strong assumptions.

Main Finding

Access to progressively more capable LLMs materially raises professional productivity: each year of frontier model progress reduced task completion time by ≈8% (p ≈ 0.04). Across a large preregistered RCT (N>500), AI assistance raised base earnings per minute (EPM) by 81.3% ($0.56/min), improved expert-assessed quality by ~0.34 SD (0.60 points on a 7‑pt scale), and increased total earnings per minute (TEPM, including performance bonuses) by 146% ($1.06/min, p < 0.001). A tenfold increase in training compute is associated with a 6.3% reduction in task time. Decomposing frontier progress, ~56% of gains are attributable to compute scaling and ~44% to calendar-time (algorithmic/data/architecture) improvements. Gains are much larger on non‑agentic analytical tasks than on agentic, tool‑using workflows. Projecting these elasticities in an aggregate growth framework yields an estimated ~20% U.S. productivity boost over the next decade under low inference-cost assumptions.

Key Points

• Experimental design and scale

  • Large preregistered RCT with >500 screened professionals (management, data analysis, consulting).
  • Participants randomly assigned to control or one of 13 LLMs spanning multiple release dates and training‑compute scales.
  • High‑powered incentives: $15 base per task + $15 bonus for high‑quality submissions (grades ≥5/7). Expert graders assigned scores.

• Average AI uplift (pooled over models)

  • EPM +$0.56/min (81.3% increase; p = 0.001).
  • Quality +0.60 points on 7‑pt scale (~0.34 SD; p < 0.001).
  • TEPM +$1.06/min (146% increase; p < 0.001).
  • Composition: increases in speed contribute ~52.6% of TEPM gains; quality improvements ~47.4%.

• Scaling laws (economic impacts)

  • Calendar‑time scaling: each additional year of frontier model progress ≈ 8% reduction in task time.
  • Compute scaling: a 10× increase in training compute → ≈ 6.3% reduction in time.
  • Decomposition: ~56% of observed gains driven by compute scaling; ~44% driven by algorithmic/calendar‑time progress.

• Heterogeneity by task type

  • Non‑agentic (analytical/writing) tasks: much larger gains — TEPM +$1.58/min (p < 0.001).
  • Agentic (multi‑step tool use, procedural workflows): smaller and statistically weaker gains — TEPM +$0.34/min (p = 0.46); difference across task types significant (p = 0.043).
  • Interpretation: scaling so far rapidly augments analytical cognition; procedural agency/tooled workflows lag.

• Quality vs. capability scaling

  • Autonomous model outputs’ quality scales with compute.
  • Human‑assisted (human+LLM) output quality is flat across model generations — evidence of a “human cap” where users attain a satisficing quality threshold rather than exploiting maximal model capability.

• Aggregate projection

  • Using estimated elasticities in an aggregate growth framework (following Acemoglu 2024), the paper projects ~20% U.S. productivity growth over the next decade, conditional on continued scaling and low marginal inference costs.

Data & Methods

• Sample and recruitment

  • 500 participants recruited mainly via Prolific; stringent eligibility (≥$40k salary, ≥1 year experience) and competency screening (≈90% screen‑out).

  • Final sample: experienced professionals (79% ≥3 years experience; 47% ≥5 years).

• Tasks and classification

  • One task per participant drawn from profession‑specific, realistic workflows (management, consulting, data analysis).
  • Tasks categorized as agentic (require multi‑step tool interactions) or non‑agentic (analytical/writing).

• Treatments

  • Random assignment to control (no bot) or one of 13 LLMs varying in training compute and release date.
  • Custom website logged usage; monitored practice run ensured protocol compliance.

• Outcomes

  • Time taken, base Earnings Per Minute (EPM), Total Earnings Per Minute (TEPM, includes bonus), and expert‑assessed grade (0–7).
  • Primary estimands: effects of any AI vs control; elasticities of outcomes to model release month and to log(training compute). Combined specs decompose compute vs calendar‑time effects.

• Identification & analysis

  • Pre‑registered RCT (AEARCTR‑0013743); approved by Yale HRPP.
  • Regression controls included profession, task, demographics, abilities, AI familiarity, country dummies; robustness checks in appendices.
  • Statistical significance reported (p‑values); figures/tables in main text and Appendix A.

Implications for AI Economics

• Empirical link from ML scaling to economic output

  • Provides direct, experimentally grounded elasticities linking model capability (compute and calendar‑time improvements) to human productivity — useful inputs for macro/sectoral forecasting and welfare analysis.

• Heterogeneous automation potential

  • Analytical cognition appears highly susceptible to current scaling paradigms; procedural, tool‑driven tasks remain more resistant. Policy and firm strategies should distinguish between task types when forecasting displacement, re‑skilling needs, and investment in automation infrastructure (agentic toolchains, APIs, UI/UX).

• Importance of human factors and complementarity

  • The “human cap” on quality suggests realized gains depend on how humans interact with models (training, prompts, incentives, UX). Investments in user training, interface design, and workflow integration may unlock further gains beyond raw model scaling.

• Role of incentives and measurement

  • High‑powered piece‑rate style incentives amplified TEPM results; translating these experimental rates to typical salaried settings requires caution. Still, the study demonstrates that speed＋quality improvements can compound substantially under performance‑linked compensation.

• Policy and forecasting

  • A ~20% U.S. productivity uplift over a decade (conditional) is economically large and suggests rethinking near‑term growth scenarios and labor market policy (education, tax/transfer design, social insurance for transitions).
  • However, assumptions matter: inference costs, broader diffusion, regulatory constraints, and development of agentic/tooled models will affect realized aggregate impacts.

• Evaluation methodology

  • Supports “centaur evaluations” (human+AI) as economically relevant benchmarks. Benchmarks that evaluate models in isolation risk misestimating real‑world productivity impacts.

Caveats & Limitations

• Sample scope: three high‑skill professions and a screened, experienced online sample — not representative of the entire labor force.
• Interface limitations: participants used standard chatbot interfaces with limited tool access; agentic task gains may be understated relative to systems with richer tool integration or autonomous agents.
• Incentive design: large bonuses (doubling pay) may exaggerate TEPM effects relative to typical workplace compensation structures.
• Proxies for algorithmic progress: model release date is an imperfect proxy for non‑compute improvements and may confound many contemporaneous changes.
• External validity: scaling projections assume continued predictable scaling, low inference costs, and broad adoption — all uncertain.

If you want, I can (a) extract the main regression coefficients and p‑values into a single table, (b) produce a one‑page brief oriented to policymakers, or (c) simulate alternative aggregate productivity scenarios under different inference‑cost and adoption assumptions. Which would you prefer?