AI development raises Chinese cities' inclusive green growth by widening digital finance, reallocating factors and spurring green innovation — benefits strengthen once government S&T spending passes two smooth thresholds; spatially, AI first concentrates resources in core cities then later radiates green technology to neighbors.

Main Finding

AI development has a robust causal positive effect on Inclusive Green Growth (IGG) in Chinese cities. It raises coordinated economic, social and environmental outcomes primarily by (1) expanding digital financial coverage, (2) improving factor-allocation efficiency, and (3) stimulating green technological innovation. The effect is nonlinear—showing smooth increasing marginal returns conditional on government science & technology (S&T) fiscal support (two transition thresholds)—and spatially time-varying: AI initially induces negative “siphon” spillovers (factor agglomeration to core cities) and later positive “radiation” spillovers (green-technology diffusion) as regional integration and supporting conditions deepen. Human capital and digital inclusive finance amplify AI’s benefits; resource-based cities and areas with mature digital infrastructure capture larger sustainability dividends. Policy simulation suggests targeted AI-related S&T resource allocation can improve urban low‑carbon inclusive governance efficiency by about 160%.

Key Points

• Causal identification
  • Uses instrumental-variable augmented Double Machine Learning (IV‑DML) to estimate conditional average treatment effects (CATE) and address high-dimensional confounding and endogeneity.
  • Instrument: topographic relief (from DEM).
• Mechanisms (mediation)
  • Digital financial inclusion expansion.
  • Optimization of factor allocation (capital, labor, technology, land).
  • Promotion of green technological innovation (green R&D and patents).
• Moderation
  • Human capital and digital inclusive finance positively moderate the AI → IGG effect.
• Nonlinearity
  • Panel Smooth Transition Regression (PSTR) finds smooth, increasing marginal returns of AI on IGG as government S&T fiscal input rises, with two transition thresholds (i.e., the enabling effect strengthens after passing threshold levels of public S&T support).
• Spatial dynamics
  • Time-varying spatial Durbin model with dynamic spatial weights reveals evolution from negative siphoning spillovers (early stage) to positive radiation spillovers (later stage) in AI’s inter-city effects.
• Heterogeneity
  • Stronger effects in resource-based cities and regions with more developed digital infrastructure.
• Policy simulation
  • Targeted reallocation of AI‑related S&T resources yields large efficiency gains (≈160%) for low‑carbon, inclusive urban governance.

Data & Methods

• Sample
  • Balanced panel of 282 Chinese prefecture-level cities, 2016–2025 (2,820 observations). Winsorized continuous variables at 1st and 99th percentiles.
• Core variables and measurement
  • IGG index: constructed by fixed-base range entropy weight method integrating economic efficiency, social equity, and ecological sustainability indicators.
  • AI development index: city-level index derived from text mining (Python web-crawling) of government work reports.
  • Mediators: measures of digital financial coverage (Peking Univ Digital Inclusive Finance Index components), factor-allocation efficiency indicators, green-technology inputs/outputs (green patent data from CNRDS).
  • Threshold variable: government science & technology fiscal input.
  • Instrumental variable: topographic relief computed from Digital Elevation Model (DEM).
• Data sources
  • China Urban Statistical Yearbook; provincial and municipal statistical yearbooks; CSMAR; Peking University Digital Inclusive Finance Index; CNRDS; government work reports; DEM.
• Econometric strategy
  • IV‑DML (instrumental-variable Double Machine Learning) for causal identification and estimation of average and conditional treatment effects while controlling high-dimensional confounders with ML (random forest) and cross‑fitting.
  • Mediation and moderation models (panel framework with fixed effects) to test transmission channels and interaction effects (human capital, digital finance).
  • Panel Smooth Transition Regression (PSTR) to detect smooth, continuous nonlinear threshold effects with government S&T input as the threshold variable (identifies two transition points).
  • Time‑varying spatial Durbin model (with dynamic spatial weight matrices) to capture evolving inter-city spillovers and the shift from siphoning to radiation.
  • Robustness checks and heterogeneity analyses across city types and infrastructure maturity.

Implications for AI Economics

• Methodological benchmark
  • Demonstrates an integrated empirical pipeline for causal inference in AI economics: IV‑DML for robust causality, PSTR for realistic smooth nonlinearities, and time‑varying spatial models for dynamic externalities. This combination can serve as a template for future AI impact studies.
• Policy design
  • Public S&T investment matters: government S&T fiscal support is a binding constraint that shapes AI’s marginal returns for sustainable development. Policymakers should scale and sequence S&T support to cross thresholds that unlock stronger AI-enabled green and inclusive dividends.
  • Complementary investments are critical: raising human capital and strengthening digital inclusive finance magnifies AI’s welfare and environmental benefits—so AI policy should be paired with education, training, and digital finance expansion.
  • Regional strategy: to avoid early-stage siphoning and unequal capture of AI rents, implement regional coordination and integration policies (infrastructure, knowledge-sharing platforms, cross-city innovation networks) to accelerate the transition from agglomeration to diffusion.
  • Targeted allocation of AI-related S&T resources can yield large efficiency gains in low-carbon inclusive governance; allocate resources with heterogeneity in mind (resource-based cities and digitally mature areas derive larger returns).
• Research directions
  • Replicate the IV‑DML + PSTR + time-varying spatial framework in other countries/contexts to test external validity and institutional dependence of thresholds.
  • Decompose AI effects across sectors and occupations to quantify distributional consequences and design compensatory policies for displaced groups.
  • Further validate measurement approaches (e.g., text-mined AI indices) and explore alternative instruments to strengthen causal claims.
• Cautions
  • Results hinge on measurement choices (AI index from government reports) and validity of the instrument (topographic relief). Generalization beyond Chinese prefecture-level contexts requires care.