Provincial AI deployment and tougher environmental rules are linked to lower local circular-economy efficiency in China, though AI’s benefits spill across borders; combined, AI and regulation crowd each other out, implying that smarter, coordinated green governance is needed.

Main Finding

Using provincial panel data for 30 Chinese provinces (2013–2022), the authors find that urban circular-economy efficiency (UCEE) increased substantially nationwide (average index ≈0.3 → >0.7) with large regional gaps (leading eastern provinces ≈1.0; some northeastern provinces <0.1). Decomposition via the Global Malmquist–Luenberger (GML) index shows sustained efficiency gains driven mainly by technological progress. Spatial econometric estimates (spatial Durbin model) indicate that both artificial intelligence (AI) and environmental regulation (ER) significantly reduce local UCEE, while AI produces positive spatial spillovers to neighboring regions. The interaction of AI and ER generates a negative (crowding-out) effect — described as “1+1 < 2” — implying misaligned or poorly coordinated technology–policy combinations can undermine circular-economy efficiency.

Key Points

• UCEE trends and heterogeneity
  • National average UCEE rose from ~0.3 to >0.7 over 2013–2022.
  • Pronounced spatial disparities: eastern provinces near full efficiency (~1.0); some northeastern provinces remain very low (<0.1).
• Dynamics of improvement
  • GML index and its components (efficiency change and technical change) are >1, with technological progress (technical change) being the dominant driver of UCEE gains.
• Effects of AI
  • AI adoption is associated with a negative direct effect on a province’s own UCEE (possible adoption costs, learning/transition frictions).
  • AI yields positive spatial spillovers: neighboring provinces benefit from knowledge/technology diffusion.
• Effects of environmental regulation
  • ER exerts a significant negative direct effect on local UCEE (short-run compliance costs, potential constraints on firm operations).
  • Little evidence of positive spillovers from ER to other provinces.
• Interaction (AI × ER)
  • The interaction term is negative: stronger regulation in combination with higher AI does not produce additive gains and can crowd out AI’s potential benefits for UCEE.
• Theoretical framing
  • Paper frames AI and ER as technology–institution dual engines; misalignment can turn them into conflicting forces rather than complementary drivers.

Data & Methods

• Sample and period
  • Provincial panel: 30 mainland Chinese provinces (Tibet excluded), 2013–2022.
• UCEE measurement
  • Super-SBM (slack-based measure) data envelopment analysis model incorporating undesirable outputs to compute province-year UCEE indices.
  • Desirable outputs include measures such as industrial solid-waste utilization rate, domestic waste harmless-treatment rate, output value from reuse of wastes, industrial water reuse rate, hazardous-waste safe disposal rate.
  • Undesirable outputs include pollution intensity measures (e.g., "three wastes" emissions per unit GDP).
• Dynamics decomposition
  • Global Malmquist–Luenberger (GML) productivity index used to track dynamic changes and decompose overall change into efficiency change (EC) and technological change (TC).
• Spatial econometrics
  • Spatial Durbin Model (SDM) employed on the provincial panel to estimate direct and indirect (spillover) effects of AI and ER on UCEE and to test interaction effects.
  • AI and ER measured as province-level indices/indicators constructed from statistical sources (paper provides operational details).
• Data sources
  • Core data drawn from China Statistical Yearbook, China Environmental Statistical Yearbook, and other official provincial statistics; inputs include energy and water consumption per GDP, employment, fixed capital formation ratios, industrial pollution control investment rates, etc.

Implications for AI Economics

• Importance of spatial externalities
  • AI’s economic benefits frequently manifest as cross‑region spillovers. Models and policy evaluations of AI should include spatial structure and externalities rather than only local treatment effects.
• Short-run vs long-run effects
  • AI can impose short-term local costs (adoption, reallocation, compliance) even as it fosters longer-run technological progress. Empirical studies of AI’s welfare or productivity impacts need to separate transitional frictions from sustained gains (e.g., decompose via GML- or TFP-style methods).
• Interaction with institutions matters
  • Institutional design (regulatory intensity and form) conditions AI’s effectiveness. AI and regulation are not automatically complementary; misaligned regulation can crowd out AI benefits. Economic models should treat AI×policy interactions as potentially non‑linear and context dependent.
• Policy design recommendations (for economists advising policy)
  • Coordinate AI deployment and environmental regulation: align incentives so AI investments target circular-economy applications (recycling optimization, remanufacturing design) rather than only compliance/monitoring.
  • Encourage mechanisms to capture positive spillovers: data-sharing platforms, regional cooperation, and diffusion-support policies to spread AI gains to lagging areas.
  • Calibrate regulatory stringency by regional capacity: consider phased or targeted regulation, and support AI infrastructure, training, and subsidies in regions with low UCEE to avoid punitive short-run impacts.
  • Monitor thresholds: evaluate whether ER intensity crosses points where marginal costs exceed innovation benefits (the inverted-U possibility noted in literature).
• Methodological guidance for future AI-economics work
  • Use frontier/efficiency frameworks that include undesirable outputs when assessing environmental or circular outcomes.
  • Combine dynamic decomposition (e.g., GML) with spatial econometrics to disentangle technological progress from efficiency catch-up and to capture spillovers.
  • Account for regional heterogeneity — both in AI capacity and regulatory enforcement — when estimating causal impacts or designing experiments/pilots.

If you want, I can: (a) extract the paper’s exact operational definitions for the AI and ER variables, (b) draft a short policy brief based on these findings for provincial governments, or (c) produce a figure-ready summary (key tables/plots to reproduce). Which would you prefer?