A multi-stage nonlinear relationship between artificial intelligence development and environmental pollution

Abstract

Abstract As China advances its green and low carbon transition, the influence of artificial intelligence (AI) development on environmental pollution remains a critical yet underexplored issue; nevertheless, evidence on stage dependent effects beyond the conventional inverted U Environmental Kuznets Curve (EKC) remains limited. Using provincial panel data from 2011 to 2022, this study constructs an Environmental Pollution Index (EPI) using the global entropy weight method (EWM), incorporating indicators of solid waste generation, chemical oxygen demand emissions, sulfur dioxide emissions, and carbon dioxide emissions. Kernel density estimation and Dagum Gini coefficient decomposition are employed to characterize distributional dynamics and identify the sources of regional disparities. After benchmarking seven machine learning models within a unified framework, the Generalized Additive Model (GAM) is selected as the optimal model due to its superior predictive performance and interpretability, and is then used to uncover nonlinear associations, turning points, and sign reversals. The GAM results reveal distinct multistage patterns: AI enterprises follow a segmented extended U-shaped curve, shifting from pollution aggravation to stable mitigation beyond a threshold; AI-Industry per capita GDP exhibits a W-shaped pattern; Industrial robot density declines overall with an inflection; AI patent grants display threshold effects with stronger mitigation at higher levels; R&D personal ratio in AI industry approximates a U-shaped curve, turning from positive to negative near a threshold before flattening; and AI technology innovation expenditure increases almost linearly with a stagewise shift from mitigation to aggravation. Public concern about environmental pollution follows an N-shaped curve, Environmental governance investment and Urbanization rate exhibit bimodal M-shaped patterns, and Marketization index shows a U-shaped relationship with mitigation concentrated in the middle range. Although EPI declines on average, interprovincial disparities persist, and these multistage patterns refine EKC interpretations of scale and technique effects.