Higher counts of granted AI patents are associated with slower GDP growth across countries in these panel models, while investment, government spending and population remain positive growth predictors; the authors argue AI’s productivity gains may be delayed by skills, infrastructure and diffusion constraints.

Main Finding

Using panel data for 42 countries (roughly 2013–2022) and dynamic estimators (Difference and System GMM), the paper finds that AI-related patenting is negatively associated with national GDP growth. In the preferred System GMM specifications, AI patents per million people (AIPATENT) has a statistically significant negative coefficient (≈ -0.082, p < 0.01). Broad patent activity (PATENT) also shows a small negative effect in the alternative specification (≈ -0.0024, p < 0.05). By contrast, gross fixed capital formation (GFCF) and population growth are positive and significant drivers of GDP growth, while government expenditure is negatively associated with growth. Diagnostics (AR(2) and Hansen tests) do not reject the GMM instruments.

Key Points

• Sample and scope: 42 randomly selected countries, panel spanning 2013–2022 (≈ 430 observations; some estimations use N ≈ 336–378).
• Main dependent variable: annual GDP growth (%) from World Development Indicators (WDI).
• Key innovation measures:
  • AIPATENT: AI-related patent applications granted per 1 million people (CSET 2024).
  • PATENT: total patent applications per 1 million people (WDI).
• Estimation approaches: Pooled OLS, country Fixed Effects, Difference GMM, and System GMM (two-step, robust).
• Main empirical results (selected coefficients; t-statistics reported in paper):
  • AIPATENT: System GMM ≈ -0.0822 (significant at 1% in one Sys-GMM spec).
  • PATENT: alternative Sys-GMM ≈ -0.00238 (significant at 5%).
  • GFCF: positive across models (e.g., OLS 0.247*; Sys-GMM 0.565 / 0.739).
  • GOVEXP: negative across models (e.g., OLS -0.861**; Sys-GMM ≈ -0.52 to -0.55).
  • POPGRO: positive and significant in several specifications.
  • INFLA: negative and significant in GMM estimates (≈ -0.12 to -0.14).
• Interpretation offered by authors: negative AI-patent coefficient implies that AI innovations have not (yet) translated into measurable aggregate growth—likely because adoption/diffusion is early, requires complementary skills/infrastructure, and entails upfront costs that delay productivity gains.
• Other findings: unemployment was not statistically significant in the employed specifications.

Data & Methods

• Data sources:
  • GDP growth, GFCF, government expenditure, unemployment, population growth, inflation: World Development Indicators (WDI).
  • AI patent grants per million people: Center for Security and Emerging Technology (CSET, 2024).
• Time period: 2013–2022 (panel).
• Sample: 42 countries (random selection; heterogeneity across income levels implied).
• Variables treated as endogenous in GMM: lagged GDP growth and GFCF; other regressors treated as exogenous for instrumenting.
• Estimators:
  • OLS and Fixed Effects for baseline associations.
  • Difference GMM and System GMM (Arellano–Bond / Blundell–Bond framework, Roodman implementation) to address dynamic panel bias and endogeneity.
• Diagnostic checks:
  • AR(2) p-values reported and not significant, suggesting no second-order serial correlation in residuals.
  • Hansen J-test p-values acceptable, indicating instruments are not rejected.

Implications for AI Economics

• Measurement and timing: Patent counts for AI can reflect R&D intensity but may not capture adoption, commercialization, or productivity effects in the short run. The negative association suggests that patenting alone is an imperfect indicator of near-term aggregate gains.
• Diffusion and complementarities matter: AI's growth benefits likely require complementary investments—human capital (skills), digital/physical infrastructure, firm adoption, industry transformation—and these complementarities can delay positive aggregate effects.
• Heterogeneous effects likely: Aggregate (country-level) estimates can mask sectoral or firm-level productivity gains, distributional shifts, and reallocation costs (e.g., automation displacing labor in some sectors while boosting others).
• Policy priorities:
  • Promote diffusion and adoption (subsidies, procurement, regulatory clarity).
  • Invest in skills and re-skilling to capture productivity gains from AI.
  • Strengthen institutions and infrastructure so AI inventions translate into productive use.
  • Monitor distributional impacts and design social policies to manage transition costs.
• Research recommendations:
  • Move beyond patents as the sole AI measure: use direct adoption indicators (firm-level AI use, AI-capital expenditures), sectoral productivity, and labor-market adjustments.
  • Longer panels and more granular (firm/sector) data to observe the lagged effects of AI innovation on productivity and growth.
  • Causal identification strategies (natural experiments, policy shocks) to separate invention from adoption effects and to study heterogeneity by income level, institutional quality, and industrial structure.
• Caution: The negative patent–growth correlation should not be interpreted as evidence that AI is intrinsically bad for growth. Rather, it highlights adoption lags, measurement limits, and the need for complementary inputs and policies to convert AI inventions into aggregate economic gains.