AI’s environmental impact in Chinese cities follows an inverted U: early AI expansion tends to raise carbon intensity, but after a maturity threshold AI adoption lowers emissions — yet by 2020 most cities remained on the emission-increasing side. The effect operates via energy use, green innovation and industrial upgrading and spills across neighbouring cities.

Main Finding

Using panel data for 266 Chinese prefecture-level (and above) cities (2011–2020) and a city-level multidimensional AI development index, the paper finds an inverted U‑shaped relationship between AI development and urban carbon emission intensity (CEI): AI development initially raises CEI (early-stage, emission‑accelerating effect) and, after a turning point, reduces CEI (later-stage braking/mitigation effect). By 2020 most Chinese cities remain on the rising segment (i.e., the early-stage emission‑increasing phase). The nonlinear net effect operates through three channels (energy‑use intensity, green technological innovation, industrial‑structure upgrading), exhibits heterogeneity across city types, and produces nonlinear spatial spillovers to neighboring cities.

Key Points

• Empirical pattern: AI → CEI follows an inverted U (rise then fall). Early AI expansion (infrastructure, model training, equipment turnover, data centers) increases energy demand and CEI; maturity and diffusion enable efficiency, cleaner operations, and structural shifts that lower CEI.
• Mechanisms validated: mediation analyses identify three transmission channels:
  • Energy‑use intensity: early-stage AI increases energy consumption; later-stage AI reduces energy intensity via optimization, smart grids, predictive maintenance.
  • Green technological innovation: AI catalyzes green R&D and diffusion, improving technology that lowers emissions per output.
  • Industrial upgrading: AI promotes industrial intelligence and AI industrialization, shifting output toward less energy‑intensive, higher‑productivity activities.
• Spatial effects: Spatial Durbin Model results show nonlinear spillovers — AI first increases and then decreases CEI in neighboring cities, indicating regional co-movement and externalities.
• Heterogeneity: The AI–CEI relationship varies by city characteristics (economic development level, resource endowment, urbanization pattern, membership in urban agglomerations), implying place‑specific dynamics and policy needs.
• Policy orientation from the paper: promote sectoral AI deployment combined with green AI pathways, strengthen regional coordination in innovation/coopetition, and adopt place‑based policies reflecting local endowments and development stages.

Data & Methods

• Data: Panel of 266 prefecture‑level and above Chinese cities, annual observations 2011–2020.
• AI measure: A composite city‑level AI development index constructed from five dimensions — industrial base, innovation capacity, input factors, application/adoption, and infrastructure — aggregated using an entropy‑weighted TOPSIS approach to capture multidimensionality beyond single proxies (e.g., robots or patents).
• Outcome: Carbon emission intensity (CEI) at the city level (CO2 per unit of output).
• Main econometric strategy:
  • Two‑way fixed‑effects panel regressions including AI and AI^2 to identify nonlinear (inverted U) relationship, with city and year fixed effects and control variables.
  • Spatial Durbin Model (SDM) to account for spatial dependence and estimate direct and spillover effects.
  • Mediation analysis (following Wen & Ye) to test the three channels (energy intensity, green technological innovation, industrial upgrading).
• Robustness: Spatial models and mediation framework used to check mechanisms and spatial externalities (paper reports robustness checks; exact control set not listed in excerpt).

Implications for AI Economics

• Nonlinearity matters: Models and policy evaluations of AI’s environmental impact must allow for nonlinear (rise‑then‑fall) dynamics rather than assuming monotonic effects.
• Short‑run vs long‑run tradeoffs: Early AI deployment can increase emissions (data centers, training costs, equipment turnover). Policy should anticipate and explicitly manage short‑run carbon costs while enabling the diffusion and maturity that generate longer‑run mitigation benefits.
• Internalize AI infrastructure emissions: Given sizable energy demands of AI infrastructure, economic instruments or regulatory measures (e.g., carbon pricing on data centers, energy‑efficiency standards for GPUs/hardware, incentives for green power for compute) can accelerate the transition to the downward segment of the curve.
• Promote green AI and scale efficiencies: Subsidies or standards that favor energy‑efficient algorithms, specialized low‑power hardware, and co‑location with low‑carbon electricity can magnify AI’s mitigation potential.
• Place‑based and regional policy design: Heterogeneous city responses and spatial spillovers imply differentiated policies—support diffusion paths appropriate to local development stage, resource mix, and urbanization form, and coordinate regionally to capture positive spillovers and avoid pollution‑shifting.
• Measurement and evaluation: Use multidimensional AI indices (as in this paper) and spatial frameworks when assessing AI’s macro and regional environmental impacts; researchers should disaggregate sectoral effects and quantify the turning point in different contexts.
• Research agenda pointers: quantify the turning-point thresholds across sectors/regions, estimate lifecycle emissions of AI infrastructure and algorithm families, compare outcomes across electricity mixes, and evaluate policy levers that accelerate the transition from the rising to the falling segment of the AI–CEI curve.