Private R&D and tax breaks correlate with more AI patenting across G8 countries and Turkey, while public R&D spending and stronger GDP growth are associated with fewer AI patents, implying tax incentives and private investment may be more closely linked to AI innovation than direct public R&D.

Main Finding

Using a balanced panel of the G8 countries plus Türkiye (2010–2020, 9 countries × 11 years = 99 observations), the paper finds that private-sector R&D spending and R&D tax incentives are strongly and positively associated with AI patent output, while government R&D spending and (surprisingly) GDP growth are negatively associated with AI patents. Estimated elasticities (Driscoll & Kraay robust RE estimates): - Private R&D (PRIRD): +4.28% in AI patents per 1% point increase in private R&D/GDP (p < 0.01).
- R&D tax incentives (TAXINC): +4.53% per 1% point increase in tax-incentive/GDP (p < 0.01).
- Government R&D (GOVRD): −1.99% per 1% point increase in public R&D/GDP (p < 0.1).
- GDP growth (GROWTH): −0.11% per 1 percentage-point increase in annual growth (p < 0.05).
Model explains about 54% of variation in log AI patents (R2 ≈ 0.54).

Key Points

• Dependent variable: log of AI patent counts (Our World in Data). Independent variables: government R&D (%GDP), private R&D (%GDP), R&D tax incentives (%GDP), and annual GDP growth (%).
• Model selection: random effects (RE) chosen after F, LM and Hausman tests.
• Diagnostics revealed serial correlation, heteroskedasticity and cross-sectional dependence → Driscoll & Kraay standard errors applied to RE estimates.
• Main substantive conclusions:
  • Private R&D and R&D tax incentives are effective for generating measurable AI innovation (patents).
  • Public R&D spending, as measured and aggregated here, appears negatively associated with patent counts—interpreted by the author as evidence of inefficiency in public R&D allocation.
  • The negative coefficient on GDP growth is small but significant; the author notes it as indicating that growth in this period/cross-section is not driving AI patenting positively in the sample.
• R2 and significance levels indicate a reasonably well‑fitting model but with notable caveats (see Limitations below).

Data & Methods

• Sample: G8 countries + Türkiye, annual data 2010–2020 (t = 11, i = 9; 99 observations).
• Data sources: AI patents (Our World in Data); GOVRD and PRIRD and TAXINC (OECD); GDP growth (World Bank).
• Model (log-linear specification): ln(AI_patents_it) = α_it + β1 GOVRD_it + β2 PRIRD_it + β3 GROWTH_it + β4 TAXINC_it + ε_it
• Econometric approach:
  • Considered OLS, Fixed Effects (FE), and Random Effects (RE); diagnostic tests led to RE.
  • Tested for autocorrelation (Bhargava et al. Durbin–Watson; Baltagi & Wu LBI), heteroskedasticity (Breusch–Pagan; Cook–Weisberg), and cross-sectional dependence (Frees; Pesaran). All tests signaled problems.
  • Because violations were present, the author reported RE estimates with Driscoll & Kraay robust standard errors to handle serial correlation, heteroskedasticity, and cross-sectional dependence.
• Summary statistics: mean (log AI patents) ≈ 14.06; average GOVRD ≈ 2.07% GDP; PRIRD ≈ 1.38% GDP; TAXINC mean ≈ 0.12% GDP.

Implications for AI Economics

Policy implications and economic interpretations: - Private R&D and tax incentives are effective levers for increasing measured AI innovation (patents). Policies that lower private cost of R&D—tax credits, tax reductions, subsidies, low-interest loans—appear justified for promoting AI patenting. - The positive and economically large elasticity on TAXINC suggests that tax‑based incentives are an important complement to direct funding; well‑designed tax incentives can stimulate private AI R&D. - The negative association for government R&D suggests possible inefficiencies in public R&D spending or misalignment between public research priorities and patentable AI outputs. This supports policy moves toward performance-/project-based public support, stronger evaluation, and better targeting toward commercialization and private sector complementarities. - The small negative effect of GDP growth on AI patents cautions against assuming aggregate growth automatically increases AI patenting; structural factors (institutional quality, absorptive capacity, human capital, VC financing, sectoral composition) matter. Research and measurement implications: - Patents are an imperfect proxy for AI progress (quality, commercial deployment, open‑source contributions, algorithms not patented). Findings therefore speak to patentable innovation, not all AI capability. - Potential endogeneity (reverse causality: countries with more AI activity may attract private R&D/tax incentives) and omitted variable bias (institutions, human capital, VC, regulation) are important. Future work should use IV, dynamic panel methods (e.g., system GMM), disaggregate public R&D, or exploit policy changes as quasi‑experiments. - Cross‑country heterogeneity matters: pooling G8 + Türkiye may mask different regimes of public vs. private roles in innovation. Country‑level case studies or interaction terms (e.g., GOVRD × institutional quality) could clarify mechanisms.

Suggested next steps for researchers and policymakers: - Use instrumental variables or natural experiments to address endogeneity (e.g., policy changes in tax incentives). - Disaggregate R&D by sector and by type (basic vs applied) and measure patent quality (citations) and non‑patent indicators (publications, open‑source contributions, firm adoption). - Evaluate public R&D programme design to improve efficiency—link funding to milestones, commercialization outcomes, and private co‑funding. - Combine tax incentive design with measures to strengthen absorptive capacity (skilled labor, university–industry links, IP regimes) to maximize returns on incentives for AI.

References and data sources noted in the paper: OECD (R&D and tax incentives), Our World in Data (AI patents), World Bank (growth); econometric tools: Driscoll & Kraay (1998) robust SEs.