National AI ecosystems do not automatically deliver greener growth: across 36 countries (2017–23) an AI-ecosystem index shows no independent effect on green growth after accounting for persistence and endogeneity, whereas better government effectiveness is clearly linked to greener economic outcomes.

Main Finding

Accounting for strong persistence and endogeneity, national AI-ecosystem development (measured as a multidimensional index) has no detectable independent effect on green growth over 2017–2023, and its interaction with government effectiveness is also statistically indistinguishable from zero. By contrast, green growth is highly path-dependent and government effectiveness alone contributes positively and significantly.

Key Points

• Sample: balanced panel of 36 advanced and emerging economies, 2017–2023.
• AI-ecosystem index: multidimensional, built on five pillars — innovation, economic diffusion, skills, policy, and computing infrastructure; pillars aggregated by within-pillar principal component analysis.
• Econometric strategy: two-step System-GMM with instruments selected/validated using Wooldridge endogeneity tests robust to heteroscedasticity.
• Persistence: estimated autoregressive coefficient ≈ 0.96, implying an annual convergence speed of about 4.5% (very strong path dependence).
• Governance: government effectiveness has a positive and significant association with green growth.
• AI effects: once persistence and endogeneity are addressed, the AI ecosystem index shows no independent effect on green growth; the AI × government-effectiveness interaction term is also not significant.
• Interpretation: simple narratives that expect AI ecosystems to automatically deliver sustainability gains are not supported by these results; governance matters, but AI alone does not show measurable short-run green dividends at the national level in this period.

Data & Methods

• Data: balanced panel (36 countries × 7 years, 2017–2023). Countries include a mix of advanced and emerging economies (paper does not report subnational data).
• AI index construction:
  • Five conceptual pillars: innovation, economic diffusion, skills, policy, computing infrastructure.
  • Within-pillar principal component analysis used to aggregate indicators into pillar scores; pillar scores combined into an overall index.
• Outcome: national-level green growth (paper reports strong autoregressive dynamics; exact green growth metric not detailed here).
• Estimation approach:
  • Dynamic panel specification with lagged dependent variable to capture persistence.
  • Two-step System-GMM estimator to address endogeneity and unobserved fixed effects.
  • Instrument selection/validation informed by Wooldridge endogeneity tests robust to heteroscedasticity.
  • Models include government effectiveness and interaction term AI_index × government_effectiveness.
• Robustness: main conclusions rely on methods that explicitly address persistence and endogeneity; when these are not corrected, spurious associations could appear.

Implications for AI Economics

• Measurement and inference:
  • Short-run, national-level gains from AI ecosystems for sustainability are not evident when dynamic persistence and endogeneity are properly handled—research should control for these factors to avoid biased conclusions.
  • Disaggregating the AI index (pillars, sectors, or firm-level adoption) may be necessary to detect targeted green effects.
• Policy:
  • Strengthening government effectiveness matters for green growth; institutional quality may be a prerequisite (but not a sufficient amplifier) for AI to generate environmental benefits.
  • Policymakers should not assume AI ecosystem development alone will yield sustainability gains; complementary environmental policies, regulation, and targeted deployments are likely required.
• Future research directions:
  • Longer time horizons and post-2023 data to capture slower, cumulative effects and learning-by-doing.
  • Sectoral and firm-level analyses to identify mechanisms (e.g., energy-intensive vs. low-carbon sectors, rebound effects).
  • Causal identification of specific AI applications (optimization, monitoring, energy management) and their environmental impacts.
  • Explore heterogeneity across country types and thresholds of institutional capacity required for AI to translate into green outcomes.
• Caution: null results here do not imply AI cannot be environmentally beneficial in specific contexts; they indicate that at the national-aggregate level over 2017–2023, there is no detectable independent contribution once persistence and endogeneity are addressed.