Cities with more AI firms register higher renewable-energy patenting — the boost is strongest for mid-performing cities and concentrated in large, non-resource urban areas and western regions, implying tailored AI–green innovation policies rather than one-size-fits-all solutions.

Main Finding

Using a balanced panel of 180 Chinese prefecture-level cities (2010–2022) and a fixed‑effects panel quantile regression, the paper finds that local AI development (proxied by the number of AI-related enterprises) is positively and significantly associated with green technology innovation (proxied by approved renewable-energy-related patent applications) across the conditional distribution. The effect is present at all examined quantiles (10th, 25th, 50th, 75th, 90th) and is strongest at the median—implying cities with intermediate innovation capacity convert AI capabilities into green patents particularly well. Heterogeneity analysis shows larger effects in large cities, more stable effects in non-resource‑rich cities, and the strongest effects in western China. Results are robust to multiple alternative measures and endogeneity checks (placebo tests and an instrumental-variable analysis).

Key Points

• Measurement
  • AI development: number of AI-related enterprises (a market/activity proxy rather than attention-based measures).
  • Green technology innovation (GTI): approved renewable-energy-related patent applications.
• Econometric approach
  • Fixed‑effects panel quantile regression to estimate effects across the conditional distribution of GTI (10th–90th quantiles).
  • Robustness: alternative AI/GTI measures, winsorization, subgroup exclusions, placebo permutations, and IV analysis.
• Main empirical findings
  • Positive, significant AI→GTI association at all quantiles.
  • Largest marginal effect at the median (50th quantile).
• Heterogeneity
  • Larger cities: stronger positive relationship between AI and green patents.
  • Resource dependence: non-resource-rich cities show more stable AI effects than resource-rich cities.
  • Regional variation: strongest AI→GTI effect in western China.
• Mechanisms and complementarities
  • Informatization (digital infrastructure/penetration) and industrial-structure upgrading are identified as complementary channels that strengthen the AI–GTI relationship; informatization may act both as a prerequisite and a reinforcing condition for AI-driven green innovation.
• Contributions
  • City-level evidence (prefecture-level) for China, addressing spatial heterogeneity.
  • Methodological: use of panel quantile regression with fixed effects to uncover distributional heterogeneity.
  • Measurement: uses number of AI enterprises as a direct proxy for local AI commercialization.

Data & Methods

• Data
  • Balanced panel of 180 Chinese prefecture-level cities, yearly observations from 2010 to 2022.
  • Dependent variable: approved renewable-energy-related patent applications (city-year).
  • Main independent variable: count of AI-related enterprises by city-year.
  • Additional variables referenced for heterogeneity and mechanisms: city size, resource-dependence indicator, regional location (east/central/west), informatization metrics, and industrial-structure upgrading indicators. (Paper also controls for standard city-level covariates though exact control list is described in the full methods section.)
• Main method
  • Fixed‑effects panel quantile regression estimating conditional effects at multiple quantiles (10th, 25th, 50th, 75th, 90th), which captures heterogeneous impacts across cities with different GTI levels while accounting for time‑invariant city characteristics.
• Robustness and identification
  • Alternative operationalizations of AI and GTI.
  • Excluding specific city groups and winsorization to check sensitivity.
  • Placebo permutation tests to guard against spurious correlation.
  • Instrumental-variable analysis to address potential endogeneity (details in full paper).

Implications for AI Economics

• Distributional returns to AI investment: Gains from AI for green innovation are not uniform—cities at different points in the innovation distribution experience different marginal returns. Policy evaluations and cost–benefit assessments of AI should account for these distributional differences rather than report only average effects.
• Importance of local absorptive capacity and complementarities: Informatization and industrial upgrading matter. AI investments are more likely to translate into green patents where digital infrastructure, skilled labor, and appropriate industry structures exist or are being developed. This underlines the classic AI-economics insight that spillovers and complementarities determine realized returns.
• Place-based policy design: Because effects vary by city size, resource dependence, and region, national AI or green-innovation policies should be tailored to local conditions. For example, supporting AI commercialization in mid-innovation cities and western regions could yield relatively high marginal green-technology gains.
• Resource dependence and path dependency: Resource-rich cities show weaker or less stable AI→GTI links, suggesting structural path-dependence and the need for complementary policies (e.g., targeted re-skilling, finance, regulation) to redirect AI benefits toward low-carbon innovation.
• Measurement insight for empirical work: Using counts of AI enterprises as a commercialization proxy can reveal different patterns than attention- or text-based indicators; researchers should consider market-activity measures when studying AI's economic impacts.
• Research directions for AI economics: Further causal identification at firm and project levels, cross-country comparisons, long-run outcomes of AI-driven green innovation, and cost-effectiveness comparisons of AI versus other green-innovation levers would sharpen policy guidance.