Patent-based estimates suggest AI diffusion has already boosted short-term growth and supports higher living standards over the long run in OECD countries, but there is no clear evidence that gains favor more-educated workers; findings depend on patent proxies and identification assumptions and exclude the recent post-2017 AI surge.

Main Finding

Using a panel ARDL for 35 OECD countries (1995–2017) the authors find that AI adoption — measured chiefly with AI-related patent counts — is associated with positive short-run effects on GDP per capita growth and positive long-run improvements in living standards. These preferred results (based on patent proxies and first-differenced GMM robustness checks) are not uniform across all contexts. At the aggregate country level, they find no evidence that AI systematically benefits a particular skill group: AI neither clearly complements nor substitutes specifically for higher- or lower-skilled workers in their data.

Citation: Duarte, J., Freitas, C., & Simões, M. (2025). Economic Growth, AI Adoption and Human Capital Across the OECD: Evidence from a Panel ARDL Model. DOI: https://doi.org/10.14195/2183-203X_61_3

Key Points

• Sample and period: 35 OECD countries, annual data 1995–2017.
• Outcome: economic growth measured as Δ log real GDP per capita (PWT 10.01).
• Main explanatory variables:
  • AI adoption proxies (per million employed): patents by applicants (aiAPP), patents by inventors (aiINV), scientific publications (aiPUB) — from Parteka & Kordalska (2023).
  • General innovation proxies (innAPP, innINV, innPUB).
  • Physical capital per capita (k) and a human-capital index (h) (PWT 10.01).
• Econometric approach:
  • Panel ARDL specification to separate short-run (differences) and long-run (levels) effects and to allow for gradual adjustment (error-correction).
  • Country fixed effects included; cointegration/panel error-correction framework tested.
  • Preferred robustness checks use patent-based AI measures and first-differenced GMM.
• Main empirical results:
  • AI adoption (patent measures) contributes to short-run GDP per capita growth.
  • AI adoption is associated with higher levels of GDP per capita in the long run.
  • Results vary by context and proxy choice — some measures and specifications do not show significant effects.
  • Interaction of AI with aggregate human capital shows no robust evidence that AI disproportionately benefits either higher-educated or less-educated workers at the country-aggregate level.
• Limitations noted by authors: measurement of AI (proxies imperfect), end of sample in 2017 (rapid post-2017 AI advances), aggregation may mask sectoral/micro heterogeneity.

Data & Methods

• Data sources:
  • Real GDP per capita, capital stock, human-capital index: Penn World Table v10.01.
  • AI and general innovation proxies: counts of patents and publications (Parteka & Kordalska, 2023), normalized per million employed.
• Model (conceptual summary):
  • Start from an aggregate production function Y = A F(K,H,L). In intensive form, log y is regressed on log k, log h, log ai, and log inn with country fixed effects (long-run relationship).
  • Panel ARDL (error-correction) form: Δ log y = short-run Δ terms + γ (log y_{t-1} − long-run combination) + lagged levels for long-run coefficients. γ (error-correction coefficient) measures speed of adjustment (expected negative).
• Estimation and robustness:
  • Panel cointegration tests to justify ARDL framework.
  • Primary estimation uses panel ARDL to recover short- and long-run coefficients.
  • Preferred specification/robustness: patent-based AI proxies and first-differenced GMM estimation to address endogeneity concerns.
  • Interactions between AI and human-capital index used to probe skill-bias / complementarity.
• Diagnostics:
  • Authors check alternative AI proxies (patents vs. publications) and note heterogeneity of results across specifications.

Implications for AI Economics

• Positive but conditional macro effects: The paper provides macro-level evidence that AI adoption can raise short-run growth rates and long-run living standards in OECD countries — but benefits are context- and measure-dependent. Cross-country capacity to absorb and deploy AI (institutions, infrastructure, complementary investment) matters.
• Measurement matters: Patent counts produce the clearest signal in this study. It highlights the need for better, more granular AI adoption measures (firm-level deployment, AI capital expenditure, AI-enabled task use) for future macro studies.
• Aggregate vs. micro/sectoral effects: Finding no clear skill-bias at the aggregate country level suggests distributional and sectoral heterogeneity likely matters — micro and sectoral analyses remain essential to detect winners and losers within economies.
• Policy takeaways:
  • Complementary investments (human capital, physical capital, institutions, R&D) and policies that reduce adjustment costs are likely necessary to realize AI’s growth potential.
  • Given heterogeneity and measurement limits, policymakers should monitor outcomes, support reskilling, and prioritize data collection on AI adoption and firm-level impacts.
• Research directions:
  • Extend analysis with post-2017 data to capture recent advances (generative AI).
  • Use firm- and sector-level adoption measures to unpack distributional impacts and productivity channels.
  • Explore dynamic heterogeneity across country groups (e.g., diffusion speed, institutional complementarities) and asymmetric effects of positive vs. negative shocks to AI.

Limitations reiterated: results are based on pre-2018 data and patent/publication proxies; conclusions should be re-evaluated as new, richer measures and more recent data become available.