Renewables drive green growth across the G20 and AI magnifies that payoff: a long-run positive effect from renewable energy (coef ≈ 0.101) is strengthened by AI (RE×AI = 0.007), whereas high emissions blunt renewable gains (RE×CO2 = −0.041); AI alone shows only transient direct effects.

Main Finding

Renewable energy (RE) has a robust, positive long-run effect on green growth across 18 G20 countries (2000–2023), and this effect is asymmetric: positive RE shocks produce stronger benefits than negative shocks produce harms. Artificial intelligence (AI) has a clear short‑run positive effect on green growth but no statistically meaningful long‑run direct effect. Crucially, AI amplifies the marginal contribution of renewable energy (positive RE×AI interaction), while high CO₂ emissions substantially erode RE’s benefits (negative RE×CO₂ interaction).

Key Points

• Long-run effect of renewable energy on green growth: coefficient ≈ 0.101 (p < 0.001).
• Asymmetry:
  • Wald tests reject symmetry (short-run χ² = 4.102, p = 0.043; long-run χ² = 5.42, p = 0.020).
  • Positive RE shocks: +0.012 (p = 0.011).
  • Negative RE shocks: short-run −0.020 (p = 0.111, not significant); long-run −0.012 (p = 0.015).
• AI effects:
  • Short-run: positive and significant (≈ 0.018, p < 0.001).
  • Long-run: small and statistically insignificant (≈ 0.001, p = 0.806).
  • Interpretation: AI’s environmental contributions are conditional and context-dependent (institutions, technology mix, energy sources).
• Interaction results (from System GMM):
  • RE × AI: +0.007 (p = 0.004) — AI strengthens RE’s marginal impact on green growth.
  • RE × CO₂: −0.041 (p < 0.001) — high emissions/intensity undermine RE effectiveness (carbon lock-in).
• Model performance & diagnostics:
  • Bounds test shows cointegration (F = 28.27, p < 0.001).
  • Error-correction terms: ARDL ECT = −0.145 (p < 0.001); NARDL ECT = −0.115 (p = 0.024) — stable adjustment to long run.
  • Adjusted R²: 0.909 (ARDL); 0.999 (NARDL).
  • GMM diagnostics: instruments valid (Sargan/Hansen), no second-order autocorrelation (Arellano‑Bond).

Data & Methods

• Sample: Panel of 18 G20 countries (Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, Mexico, Russia, Saudi Arabia, South Africa, South Korea, Turkey, UK), years 2000–2023.
• Key variables: green growth (outcome), renewable energy consumption, an AI measure/index, CO₂ emissions, FDI, plus other controls (institutional quality, trade openness, etc.; paper notes broader control set).
• Econometric strategy:
  • CS-PMG-ARDL to estimate symmetric long‑run and short‑run dynamics while accounting for cross‑sectional dependence and heterogeneity.
  • CS-PMG-NARDL to capture asymmetric (positive vs. negative) RE shocks and nonlinear adjustment.
  • Bounds test for cointegration; error-correction term for long‑run stability.
  • System GMM used to address endogeneity and estimate interaction terms (RE×AI, RE×CO₂, RE×FDI), with standard instrument validity and autocorrelation checks.
• Robustness/diagnostics: Wald tests for asymmetry, Sargan/Hansen for overidentification, Arellano‑Bond tests for autocorrelation; reported good performance.

Implications for AI Economics

Policy and research implications specific to the economics of AI and sustainability:

Policy implications - Treat AI as a conditional enabler, not an automatic green technology. Policies should: - Promote deployment of AI applications that directly increase renewable energy system efficiency (smart grids, predictive maintenance, demand response). - Incentivize energy‑efficient AI (model efficiency, green data centers powered by renewables) to avoid digital pollution and rebound effects. - Integrate AI deployment with emissions reduction strategies: decarbonize electricity systems in parallel, since high CO₂ intensity undermines AI+RE gains. - Strengthen institutions (regulatory quality, digital governance) because AI’s positive effects are conditional on governance and complementary policies.

Research implications - Model AI as a moderator/interactor in empirical work on energy and environment (include interaction terms like RE×AI). - Account for asymmetry and nonlinearity: positive and negative shocks to RE (and possibly AI investments) can have different magnitudes and policy relevance. - Measure AI more carefully: distinguish energy‑intensive AI uses (large-scale training/datacenters) from low‑energy AI applications (operational optimization), and consider lifecycle emissions. - Use panel methods that handle cross‑section dependence and endogeneity (e.g., CS‑PMG, NARDL, system GMM) when studying technology–environment links across heterogeneous countries. - Investigate heterogeneity: sectoral-level studies, country groups (advanced vs. resource-dependent), and thresholds (institutional quality, renewable penetration) to identify where AI is most effective as a green enabler. - Consider dynamic complementarities and potential trade-offs (e.g., AI-induced economic scale effects vs. energy savings).

Caveats - Results are from an unedited article in press; AI variable definition/measurement and some control details are not fully described in the abstract. Findings pertain to aggregated country-level data for G20 countries and may not generalize to non‑G20 or subnational contexts. - Future work should unpack mechanisms (which AI applications drive the interaction), quantify AI’s direct energy footprint, and test sectoral pathways.

Overall takeaway: AI can strengthen the green-growth benefits of renewable energy, but its net value depends strongly on energy-system decarbonization, institutional quality, and the types of AI deployment—implying targeted policies to align digitalization with clean-energy goals.