Spatial Effects of Artificial Intelligence Innovation on Regional Carbon Intensity

Abstract

This study investigates the spatial effects of artificial intelligence (AI) innovation on carbon intensity using provincial panel data from 30 Chinese provinces over 2010–2023. Employing the Spatial Durbin Model (SDM), we find that a 1% increase in AI patent count reduces local carbon intensity by 0.034% (direct effect, p < 0.01) but increases carbon intensity in neighboring regions by 0.069% (indirect effect, p < 0.05). Heterogeneity analysis shows that AI innovation reduces local carbon intensity by 0.069% in non-western regions (p < 0.01) but has no significant effect in the western region. In regions with above-median R&D intensity, both direct and indirect effects become negative (−0.094% and −0.069%, respectively), indicating that AI innovation reduces carbon intensity locally and in neighboring areas. Mechanism tests confirm that industrial structure upgrading mediates this relationship, with AI innovation increasing the industrial structure hierarchy coefficient by 0.004 (p < 0.05). These findings provide quantitative evidence that AI innovation has opposing local and spillover effects on carbon intensity, and that high R&D intensity can reverse negative spillovers into positive ones. The results offer empirically grounded policy recommendations for China’s dual-carbon targets and sustainable development.

1. Introduction

In recent years, the rapid advancement of artificial intelligence (AI) has brought revolutionary changes to social production and daily life. The depth of AI research continues to progress, with computer vision, large language models, and generative AI becoming current research hotspots. Concurrently, the scope of AI technology application has broadened, permeating a wide array of fields including agriculture, industry, and service sectors. Compared to other scientific and technological advancements, one of the most prominent characteristics of AI is its substantial capacity for innovation. The emergence of diverse AI applications is rapidly reshaping societal understanding, altering not only social life but also revolutionizing production methods across numerous industries. In the context of the new development stage, AI is recognized as a significant new growth driver for China’s industrial development [1]. Innovation-driven and low-carbon green development constitute essential characteristics and fundamental requirements of high-quality growth. To provide quantitative context, China’s carbon intensity (CO₂ per unit of GDP) fell from 1.29 kg/USD in 2010 to 0.70 kg/USD in 2023 (EDGAR database), yet it remains among the highest globally. Meanwhile, China’s AI patent applications grew from fewer than 2000 in 2010 to over 60,000 in 2023 (China National Intellectual Property Administration), making it the world leader in AI innovation. However, existing studies have not systematically linked these trends spatially. Current estimates suggest that industrial robot density reduces carbon intensity by 0.1–0.3% per 1% increase [2,3], but these findings are confined to manufacturing and ignore spillovers across regions. No study has quantified the direct and indirect effects of a broad AI innovation measure (patents across all sectors) on carbon intensity using spatial econometrics. This paper fills this gap by providing the first empirical estimates of AI innovation’s local and spillover effects on carbon intensity, based on 14 years of balanced provincial data.

Despite the growing scholarly attention to AI’s economic and environmental impacts, existing research exhibits three notable gaps. First, most studies proxy AI development by industrial robot density, which fails to capture AI’s broader innovation capacity across agriculture, services, and emerging sectors. Second, the spatial spillover effects of AI innovation on carbon emissions remain underexplored, particularly whether AI innovation in one region reduces or increases carbon intensity in neighboring regions—a question critical for understanding pollution transfer versus green technology diffusion. Third, while recent evaluation frameworks have emphasized the importance of assessing AI-enhanced systems in terms of sustainability outcomes [4] and the role of digitalization in transforming policy evaluations [5], empirical evidence on how AI innovation spatially affects carbon intensity in large developing economies is still lacking.

However, China, as a whole, is still at the inflection point of the Environmental Kuznets Curve, with carbon emissions yet to enter a declining phase. This necessitates further structural upgrading to promote a green transition in the development model [2]. As a representative new mode of production technology, AI can facilitate the synergistic allocation of various factors during the upgrading and transformation of traditional elements, thereby promoting a green transition in the development model. This positions AI as playing a crucial role in achieving the goals of carbon peak and carbon neutrality. Nevertheless, the current development of AI in China still faces certain limitations, including constrained implementation of innovative achievements, insufficient levels of some critical technologies, and increased energy consumption due to computing power demands [3]. Moreover, a balanced assessment of AI’s environmental impact requires acknowledging its potential negative consequences. First, the training and operation of large AI models consume substantial electricity, and if powered by fossil fuels, this could offset or even exceed the emission reductions achieved through AI applications. Second, AI-driven automation may accelerate industrial production and consumption, potentially inducing rebound effects where efficiency gains lead to increased overall energy use (Jevons paradox). Third, the digital divide across regions means that less developed areas may lack the absorptive capacity to benefit from AI innovation, exacerbating regional inequalities in carbon reduction capabilities. Fourth, the rapid obsolescence of AI hardware generates electronic waste and associated carbon footprints. These countervailing effects suggest that the net carbon reduction impact of AI innovation is an empirical question rather than a foregone conclusion.

These challenges introduce potential obstacles in the process of leveraging AI for carbon emission reduction. Given the inexorable trend of AI innovation and development, a thorough investigation into the impact of AI innovation on regional carbon intensity and its underlying mechanisms—specifically, whether and how AI innovation levels affect regional carbon intensity—is of significant importance for formulating scientific environmental policies and sustainable development strategies. This study selects provinces in Mainland China as the research object based on four explicit and scientifically defensible criteria: (i) data completeness and quality—China maintains consistent provincial-level reporting (30 provinces, 2010–2023) on carbon emissions, AI patents, and industrial structure, enabling balanced panel spatial analysis; (ii) global policy relevance—as the world’s largest carbon emitter accounting for approximately 30% of global CO₂ emissions (EDGAR database), China’s carbon reduction process directly influences the achievement of the Paris Agreement goals; (iii) AI technological leadership—China ranks first globally in AI patent filings since 2019 (China National Intellectual Property Administration), providing substantial variation in the explanatory variable; and (iv) regional heterogeneity—China exhibits large disparities in economic development, industrial structure, and innovation capacity across its eastern, central, and western regions, making it an ideal setting to test spatial heterogeneity hypotheses. In addition, China’s dual-carbon targets (carbon peak by 2030 and carbon neutrality by 2060) create a strong policy context for studying AI’s role in carbon reduction. First, Mainland China possesses complete, continuous, and publicly available provincial-level panel data on carbon emissions, AI patent counts, and industrial structure, which effectively supports spatial econometric analysis. From an academic perspective, this study aims to explore the universal patterns between AI and green development. The research findings will help enhance cross-strait academic exchanges in the fields of environmental and techno-economic studies.

To recap, the specific objectives of this paper are: (1) to quantify the direct impact of AI innovation on carbon intensity using provincial panel data from China (2010–2023); (2) to decompose the spatial effects (direct, indirect, and total) via the Spatial Durbin Model; (3) to test whether industrial structure upgrading serves as a mediating mechanism; and (4) to conduct heterogeneity analyses across economic belts and R&D intensity levels. By answering these research questions, this study contributes to the literature on AI and environmental sustainability by shifting the focus from industrial robot applications to broader AI innovation capacity and by uncovering the spatial paradox of local carbon reduction versus neighboring carbon increase.

2. Literature Review

The literature on AI and sustainability has evolved along two complementary streams. The first stream, rooted in program evaluation and policy assessment, emphasizes the need for rigorous frameworks to evaluate AI-enhanced systems. Pudney et al. [4] proposed a conceptual framework for evaluating AI-enhanced critical infrastructure systems, highlighting that AI’s environmental impacts often manifest indirectly through system optimization and behavioral changes rather than direct emissions. Similarly, Potluka et al. [5] examined how digitalization transforms evaluation practices, arguing that the spatial and temporal diffusion of digital technologies creates new challenges for assessing their net environmental effects. These evaluation perspectives suggest that AI innovation may generate spatially heterogeneous outcomes—a theoretical insight that has rarely been empirically tested in the context of carbon emissions. The second stream, which forms the core of this study, focuses on empirical estimations of AI’s environmental impact. Current research on the environmental impact of AI has primarily concentrated on the industrial sector, specifically examining the influence of industrial intelligence on the environment. The proxy variables for AI used in these studies are predominantly related to industrial robots, such as installation density, quantity deployed, and exposure rates. Most studies have suggested that the application of AI in industry exerts a positive environmental impact, contributing to carbon emission reduction and improvements in green total factor productivity [6]. For instance, using China’s provincial panel data from 2006 to 2019 and measuring AI development levels by industrial robot installation density, it was found that AI significantly reduced carbon emissions in China’s eastern region, although its carbon reduction effect was not significant in the central and western regions [7]. However, this study—like most others—relies on industrial robot data, which neglects AI innovation in non-industrial sectors and ignores spatial spillovers. Our review suggests that no prior research has simultaneously accounted for AI patents as a broad innovation measure, spatial dependence, and mediating mechanisms. Therefore, the present study advances beyond existing literature by integrating these three dimensions. Employing a similar methodology with provincial panel data, another study reached comparable conclusions, also asserting that AI significantly reduces carbon emissions, primarily through improving energy structure and fostering technological innovation [8]. Some investigations have adopted a more granular approach, utilizing city-level panel data. It was demonstrated that industrial robots effectively reduced urban industrial carbon emissions [9]. Other research has adopted an industry-level perspective, examining the impact of industrial AI on sectoral carbon emissions. Evidence indicated that industrial robots significantly reduce industry-level carbon emissions through pathways such as improving energy efficiency, optimizing energy structure, and altering factor input ratios [10].

The existing literature reveals that research on AI’s environmental impact has largely been confined to the industrial domain. However, AI applications have long expanded into numerous other fields, particularly the tertiary sector. Data related to industrial robots, being exclusively focused on industry, struggles to comprehensively reflect a region’s overall AI development level and, more importantly, fails to capture AI’s inherent characteristics of continuous innovation and iteration. Consequently, a more representative measure of AI innovation is required. In studies examining the environmental impact of technological innovation, R&D expenditure [11,12] and patent counts [13,14,15] are frequently employed as indicators of technological innovation levels. Nevertheless, no specific statistics are currently available for regional AI R&D investment. AI patents, in contrast, span multiple sectors including agriculture, industry, and services, and manifest in various forms such as software and hardware. This allows them to offer a more integrated and comprehensive reflection of a region’s AI innovation capacity and development level. Moreover, patents capture the frontier of technological invention rather than mere adoption, which is critical for studying innovation-driven effects on carbon intensity. While industrial robot density reflects the deployment of a specific automation technology, AI patents encompass algorithms, software, hardware, and cross-sectoral applications, making them a superior proxy for a region’s capacity to generate new AI solutions that can be applied to energy efficiency, industrial optimization, and green transition. Admittedly, patents do not measure actual usage or commercialisation; we acknowledge this limitation in Section 6.4. Nevertheless, for the research question of how AI innovation (as opposed to adoption) spatially affects carbon intensity, patents are the most appropriate available measure given the absence of regional AI R&D expenditure data. Therefore, this study focuses on AI innovation as the central research variable, utilizing the number of AI patents as its measurement indicator. The objective is to investigate the spatial effects and impact pathways of AI innovation on regional carbon intensity.

3. Theoretical Analysis and Research Hypotheses

The Environmental Kuznets Curve (EKC) provides a macro-level backdrop: the “technique effect” suggests that technological progress reduces emissions per output. However, standard EKC frameworks treat technology as an aggregate factor. This study focuses specifically on AI innovation as a distinct form of general-purpose technology with unique characteristics (scalability, rapid iteration, cross-sectoral applicability), thereby unpacking the black box of technology in environmental economics. Rather than testing the EKC hypothesis itself, we use its technique effect as a theoretical foundation for hypothesizing that AI innovation reduces carbon intensity. According to the EKC, when economic development reaches a certain stage, technological effects can mitigate the impact of production on the natural environment and reduce pollution emissions per unit of output. Research across various fields has demonstrated the positive role of AI technology in optimizing environmental indicators, particularly energy consumption and carbon emissions. Rather than simply enumerating sectoral applications, the carbon reduction potential of AI innovation can be theoretically categorized into three mechanisms: (i) the efficiency effect—AI optimizes resource allocation and energy use within existing production structures; (ii) the substitution effect—AI enables the replacement of high-carbon inputs with low-carbon alternatives; and (iii) the structural effect—AI drives the transformation of the industrial mix toward less carbon-intensive sectors. These mechanisms manifest across multiple sectors. In the construction sector, AI can retrofit existing buildings [13], optimize building operations management [16], and enhance energy efficiency. It can also contribute to reducing energy consumption and carbon emissions in the construction industry through the simulation and optimized design of new energy-efficient buildings [17]. In the transportation sector, AI can improve public transport system scheduling [18], optimize vehicle routes [19], and enhance the performance of autonomous vehicles [20], thereby reducing carbon emissions from transportation. In the manufacturing sector, AI can optimize production processes from a holistic perspective, improve operational efficiency [21], and reduce carbon emissions generated during industrial manufacturing. In the energy sector, AI can increase solar energy capture rates [22], reduce operational costs of wind power [23], and promote the adoption of clean energy, consequently lowering the proportion of fossil fuels and reducing carbon emissions. Thus, AI technology not only facilitates the upgrading and transformation of traditional factors of production but also functions as a new type of factor, playing a unique role across various fields and industries. It enhances operational efficiency and promotes the synergistic allocation of resources, fundamentally driving a green transition in the development model and reducing carbon emissions. Empirical evidence from recent studies supports the magnitude of these mechanisms. For example, Liu et al. [6] estimated that a 1% increase in industrial robot density reduces industrial carbon intensity by approximately 0.2% in China. In the energy sector, AI-based optimization has been shown to reduce building energy consumption by 10–15% [13,14] and improve solar power capture by 5–8% [20]. At the provincial level, the average AI patent count in our sample ranges from 0 to over 30,000, with a standard deviation of 1.896 in log terms (

Table 1. Variable definitions, data sources, and expected signs.

Variable	Definition	Source	Expected Sign
lnCI (Carbon intensity)	ln(CO2 emissions/GDP)	EDGAR + China Statistical Yearbook	Dependent variable
lnAII (AI innovation)	ln(Number of AI patents + 1)	China National Intellectual Property Administration	Negative (direct effect)
lnER (Environmental regulation)	ln(Investment in pollution treatment /industrial value added)	China Environmental Statistical Yearbook	Ambiguous
lnES (Energy structure)	ln(Electricity consumption/total national consumption)	China Energy Statistical Yearbook	Positive
lnIND (Industrialization level)	ln(Industrial value added/GDP)	China Statistical Yearbook	Positive
lnTB (Tax burden)	ln(Tax revenue/GDP)	China Statistical Yearbook	Positive
lnAIS (Industrial structure upgrading)	Σ(industry share × m) for m = 1, 2, 3	China Statistical Yearbook	Mediator (positive on CI reduction)

). This substantial variation allows us to estimate the effect size of AI innovation on carbon intensity. Drawing on these quantitative benchmarks, we propose:

Hypothesis 1.

The level of AI innovation exerts a reducing effect on regional carbon intensity.

Technological innovation can influence regional social structures, economic development, and environmental conditions. As the scale of technological innovation expands, this influence extends beyond the local region to surrounding areas, generating spatial spillover effects. In developed countries and regions, technological innovation tends to have stronger environmental attributes and more frequent factor mobility, which can reduce carbon emissions in neighboring areas. However, in developing countries and regions, the environmental friendliness of technological innovation is relatively lower, potentially leading to increased carbon emissions in surrounding areas [24]. AI technological innovation represents a significant source of contemporary technological advancement. Spatial correlations exist in AI innovation levels across different regions, generating spatial spillover effects that influence carbon intensity. With the rapid development of AI technology, its knowledge and technological achievements diffuse across regions, triggering activities such as industrial transfer, capital flow, and population migration, thereby affecting carbon intensity in surrounding areas. Specifically, spatial spillovers of AI innovation on carbon intensity can occur through three distinct channels. First, the technology diffusion channel: advanced AI technologies developed in high-innovation regions may be adopted by neighboring regions, potentially reducing their carbon intensity if the technologies are green, or increasing it if they enable more production without corresponding emission controls. Second, the industrial relocation channel: as AI innovation raises productivity and environmental standards in the home region, high-carbon, low-technology industries may relocate to neighboring regions with looser regulations, increasing carbon intensity there (the “pollution haven” effect). Third, the factor mobility channel: capital and skilled labor tend to flow toward regions with higher AI innovation, leaving neighboring regions with older, dirtier capital stocks and less skilled labor, thereby raising their carbon intensity. These channels are not mutually exclusive and may operate simultaneously. In the Chinese context, given regional competition and uneven development, the net effect of these spillovers is hypothesized to be positive (i.e., increasing carbon intensity in neighboring regions). China exhibits imbalanced and insufficient development across different regions, accompanied by competitive dynamics. The maturity and commercialization level of AI technology require further improvement. Therefore, it is posited that the carbon reduction effect of China’s current AI innovation level is confined locally, while exhibiting pollution transfer effects on surrounding areas. On one hand, an increased level of AI innovation can promote the development of high-end industries locally [25], leading to the agglomeration of capital and labor in these sectors. Traditional high-carbon industries face greater competitive pressure and consequently transfer to surrounding areas with lower AI innovation levels, increasing carbon intensity in those regions. On the other hand, significant disparities exist across regions in AI technology application capacity, infrastructure development, and industrial structure. Some areas, due to limited technological absorption capacity, struggle to effectively receive and transform AI innovation achievements spilling over from other regions to exert carbon reduction effects. Conversely, inappropriate technology application may even lead to increased carbon intensity [3]. Therefore, Hypothesis 2 is proposed.

Hypothesis 2.

The level of AI innovation exerts a spatial spillover effect on carbon intensity in surrounding regions, manifested as an increase in carbon intensity in those areas.

Spatial heterogeneity refers to the differences in certain geographical characteristics or phenomena across various regions, manifesting in multiple dimensions. Regarding economic development levels, China’s eastern region, with its relatively advanced economy, features an industrial structure dominated by high value-added manufacturing and services, alongside higher innovation capacity and technological levels. The western region exhibits relatively lower economic development levels compared to the eastern region, with an industrial structure characterized by resource-intensive and labor-intensive industries, where economic growth relies more heavily on resource exploitation and traditional industries. These disparities in economic development levels lead to divergent regional performances in technology application, industrial upgrading, and environmental protection. In terms of policy and institutional environments, the eastern region demonstrates stronger policy implementation, stricter environmental regulations, and higher corporate innovation motivation. The western region lags in policy support and institutional innovation, with weaker enforcement of environmental regulations, leading to greater challenges in technology diffusion and green development. Heterogeneity among economic belts results in differential effects of technological innovation on carbon emissions across regions [26]. Concurrently, R&D intensity varies across regions. In areas with high R&D intensity, AI technological innovation can be more effectively translated into practical outcomes, promoting a green transition in the development model and thus more fully realizing carbon reduction potential. AI represents the most cutting-edge technological innovation, and its penetration, impact intensity, and influence pathways differ across China’s regions, thereby exerting varying effects on carbon emissions. Therefore, Hypothesis 3 is proposed.

Hypothesis 3.

The impact of AI innovation level on carbon intensity exhibits heterogeneity across different economic belts and regions with varying R&D intensity.

Industrial structure refers to the systematic configuration of the relative proportions and interrelationships among various industrial sectors within an economy, including the proportional relationships between primary sectors such as agriculture, industry, and services, as well as the distribution among sub-sectors within each industry. The rationality and optimization degree of industrial structure constitute key indicators for measuring economic development quality and sustainability. Due to significantly different carbon emission characteristics across industries, industrial structure directly determines the carbon intensity within an economic system. Heavy industry and resource-intensive sectors are typically associated with high energy consumption and carbon emissions; the higher their proportion, the greater the carbon intensity. Conversely, the development of service industries, high-tech sectors, and information industries entails lower energy consumption and carbon emissions; their higher proportion can effectively reduce carbon intensity. Therefore, optimizing industrial structure, particularly by increasing the share of low-carbon and high value-added industries, represents an important pathway for achieving carbon emission reduction [27,28,29]. AI technology can spawn emerging industries and business models while also driving the intelligent transformation of traditional industries. With the widespread application of AI in a region, the industrial structure gradually transitions toward high value-added and knowledge-intensive directions. Traditional labor-intensive and resource-intensive industries are progressively replaced by technology-driven sectors, steering the economic model toward more efficient, greener, and sustainable development. Simultaneously, industrial transfer between regions may trigger pollution transfer. As AI innovation levels increase in a particular region, high-carbon industries often face higher production costs and market competitiveness pressures, thus tending to seek more efficient production methods to maintain competitiveness. They may choose to transfer to surrounding regions with lower AI innovation levels, thereby increasing carbon intensity in those areas [30]. Furthermore, the transfer of high-carbon industries may bring increased employment opportunities and economic activities, potentially raising carbon emissions in surrounding regions. To establish a convincing causal chain, we argue that AI innovation precedes industrial structure upgrading for three reasons. First, AI is a general-purpose technology that creates new industries (e.g., AI software, autonomous systems) and transforms existing ones through automation and data-driven optimisation, directly altering the composition of output across sectors. Second, AI innovation reduces information and coordination costs, enabling resources to shift from low-productivity, high-carbon industries to high-productivity, low-carbon services and high-tech manufacturing. Third, empirical studies have documented that technological innovation leads to structural change with a time lag, mitigating reverse causality concerns. Therefore, Hypothesis 4 is proposed.

Hypothesis 4.

The level of AI innovation influences regional carbon intensity through the adjustment of industrial structure.

4. Research Design

4.1. Model Construction

4.1.1. Baseline Model

To investigate the impact of AI innovation level on carbon intensity, a panel regression model was constructed for empirical analysis. The structure of this model is presented in Equation (1):

$$\ln {\mathit{CI}}_{\mathrm{it}} = \alpha + \beta {\ln \mathit{AII}}_{\mathit{it}} + \sum \gamma X_{\mathit{it}} + { \lambda }_t + {\mu }_i +{ ϵ}_{\mathit{it}}$$

(1)

where ln denotes the natural logarithm of the variable; i represents the region; t represents the year; ${\mathit{CI}}_{\mathrm{it}}$ denotes carbon intensity; ${\mathit{AII}}_{\mathit{it}}$ denotes the level of AI innovation; $X_{\mathit{it}}$ represents control variables, including environmental regulation, energy structure, industrialization level, and tax burden; ${\lambda }_t$ represents time fixed effects; ${\mu }_i$ represents regional fixed effects; and ${ϵ}_{\mathit{it}}$ represents the random error term.

4.1.2. Spatial Econometric Model

Traditional econometric models neglect the spatial correlations among variables. However, previous studies have indicated that both environmental and technological indicators exhibit spatial dependence; therefore, conclusions derived from traditional models may be subject to limitations and biases. To comprehensively analyze and demonstrate the spatial characteristics of carbon intensity and AI innovation level, the Spatial Durbin Model (SDM) was employed to investigate the spatial effects of AI innovation on carbon intensity. The structure of this model is presented in Equation (2):

$$\ln {\mathit{CI}}_{\mathrm{it}} = \alpha + \rho \mathrm{Wln}{\mathit{CI}}_{\mathrm{it}} + {\beta }_1{\ln \mathit{AII}}_{\mathit{it}} +{ \beta }_2{\mathrm{W}\ln \mathit{AII}}_{\mathit{it}} + \sum {\mathsf{\gamma }}_1{X}_{\mathit{it}} + \sum {\mathsf{\gamma }}_2{\mathrm{W}X}_{\mathit{it}} +{ \lambda }_t + {\mu }_i + {ϵ}_{\mathit{it}}$$

(2)

where $\rho$ represents the spatial autoregressive coefficient; $\mathrm{W}$ represents the known spatial weight matrix; and the other coefficients carry the same meanings as in Equation (1). In the Spatial Durbin Model, the effects of explanatory variables on the dependent variable can be decomposed into direct effects, indirect effects, and total effects. Intuitively, the SDM allows us to answer two related questions: (1) Does AI innovation in a given province reduce that province’s own carbon intensity? (direct effect). (2) Does AI innovation in a given province also affect carbon intensity in neighbouring provinces? (indirect or spillover effect). The latter captures phenomena such as technology diffusion (which could lower neighbours’ emissions) or industrial relocation (which could raise neighbours’ emissions). The spatial weight matrix W defines which provinces are considered “neighbours”—in our main specification, based on geographical distance. The spatial autoregressive coefficient ρ indicates whether carbon intensity itself clusters spatially, i.e., whether a province with high carbon intensity tends to be surrounded by other high-intensity provinces. This intuitive decomposition is the key advantage of SDM over traditional regression models.

4.1.3. Mechanism Model

Following the two-step approach for mediating effects suggested by Jiang [28], the impact of the mechanism variable on the dependent variable should be direct and evident, substantiated by extensive literature. After regressing the dependent variable on the explanatory variable, the mechanism variable was regressed on the explanatory variable to investigate whether a mediating effect exists among the explanatory variable, mechanism variable, and dependent variable. The structure of the spatial mediating effect mechanism model is presented in Equation (3):

$$\mathrm{l}\mathrm{n}AI{\mathrm{S}}_{\mathrm{i}\mathrm{t}}=\alpha +\rho \mathrm{W}\mathrm{l}\mathrm{n}{CI}_{\mathrm{it}}+{\beta }_1{\ln \mathit{AII}}_{\mathit{it}}+{\beta }_2{\mathrm{W}\ln \mathit{AII}}_{\mathit{it}}+\sum {\mathsf{\gamma }}_1{X}_{\mathit{it}}+\sum {\mathsf{\gamma }}_2{\mathrm{W}X}_{\mathit{it}}+{\lambda }_t+{\mu }_i+{ϵ}_{\mathit{it}}$$

(3)

4.2. Variable Selection

4.2.1. Dependent Variable

The dependent variable is carbon intensity (CI), expressed as the ratio of carbon emissions to regional gross domestic product (GDP), reflecting the carbon emissions per unit of economic output. Carbon emissions data were obtained from the Emissions Database for Global Atmospheric Research (EDGAR), and GDP data were sourced from the China Statistical Yearbook.

4.2.2. Explanatory Variable

The core explanatory variable is the level of AI innovation (AII), represented by the number of AI patents. Data on the number of AI patents were retrieved from the China National Intellectual Property Administration, based on AI patent codes identified in the Reference Relationship Table between Strategic Emerging Industry Classification and International Patent Classification (2021).

4.2.3. Control Variables

Control variables, widely substantiated in existing research as influencing carbon emissions, were employed to account for the effects of environmental governance level, energy consumption, and industrialization level on carbon intensity. These include: Environmental Regulation (ER), expressed as the ratio of completed investment in industrial pollution treatment to industrial added value. Energy Structure (ES), expressed as the ratio of electricity consumption to total national electricity consumption. Industrialization Level (IND), expressed as the ratio of industrial added value to regional GDP. Tax Burden (TB), expressed as the ratio of tax revenue to regional GDP. Data for control variables were sourced from the China Statistical Yearbook, provincial statistical yearbooks, the China Energy Statistical Yearbook, and the China Environmental Statistical Yearbook.

4.2.4. Mechanism Variable

A mechanism variable was employed to investigate the pathways through which AI innovation reduces carbon intensity. The mechanism variable is industrial structure upgrading (AIS). To measure the degree of industrial structure transformation and upgrading, following the approach of Yuan and Zhu [31], an industrial structure hierarchy coefficient was adopted to quantify industrial structure upgrading. This coefficient describes the evolutionary process of the three industries at a quantitative level through relative changes in their proportional shares. This indicator reflects the proportional changes in China’s three industries, from the initial dominance of the primary industry to the subsequent gradual ascendancy of the secondary and tertiary industries, embodying the quantitative connotation of industrial structure advancement. It is expressed as the weighted sum of the proportions of the three industries in regional GDP, calculated according to Equation (4):

$$\mathit{AI}{\mathrm{S}}_{\mathrm{it}}={\displaystyle \sum _{m = 1}^3}{y}_{\mathit{imt}} \times m, m = 1,2,3$$

(4)

where $y_{\mathit{imt}}$ represents the proportion of the m-th industry in region i at year t relative to regional GDP. It should be noted that industrial structure upgrading represents one possible mediating channel, rather than the only pathway through which AI innovation affects carbon intensity. Other mechanisms, such as energy efficiency improvements and technological spillovers, may also play important roles. All variables, their definitions, data sources, and expected signs are summarized in Table 1 below.

5. Results and Discussion

5.1. Descriptive Statistics

To ensure data availability and integrity, a balanced panel dataset of 30 Chinese provinces (autonomous regions and municipalities) from 2010 to 2023 was employed in this study. To mitigate estimation bias arising from heteroscedasticity and multicollinearity, natural logarithms were applied to all variables. Descriptive statistics for all variables in the model are presented in

Table 2. Descriptive statistics of variables.

Variable	Observations	Mean	Standard Deviation	Minimum	Maximum
lnCI	420	7.278	0.680	5.224	8.826
lnAII	420	6.288	1.896	0.000	10.309
lnER	420	−6.215	0.971	−9.697	−3.474
lnES	420	−3.621	0.673	−5.576	−2.336
lnIND	420	−1.169	0.301	−2.295	−0.587
lnTB	420	−2.589	0.316	−3.342	−1.611
lnAIS	420	0.873	0.050	0.757	1.046

.

5.2. Baseline Model Results Analysis

The baseline regression results of AI innovation on carbon intensity are presented in

Table 3. Baseline regression results.

Variable	(1) Pooled OLS	(2) Region-FE	(3) Two-Way FE
lnAII	−0.251 *** (0.012)	−0.175 *** (0.016)	−0.067 ** (0.028)
lnER	0.182 *** (0.021)	0.033 *** (0.009)	0.022 *** (0.008)
lnES	0.210 *** (0.048)	0.288 (0.211)	0.177 (0.189)
lnIND	0.212 *** (0.056)	0.198 *** (0.051)	0.170 * (0.091)
lnTB	−0.363 *** (0.073)	0.208 *** (0.056)	0.168 * (0.090)
Constant	10.054 *** (0.246)	10.399 *** (0.684)	9.392 *** (0.889)
Observations	420	420	420
R2	0.752	0.877	0.922
Number of regions	30	30	30
Region fixed effects	No	Yes	Yes
Year fixed effects	No	No	Yes

Note: ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively. Standard errors are in parentheses. This applies to all subsequent tables.

. The pooled OLS regression results indicated that the impact of AI innovation on carbon intensity was significantly negative. In the panel regressions, after incorporating region-fixed effects and both time and region-fixed effects, the coefficients for AI innovation remained significantly negative at the 1% and 5% levels, respectively. This suggests that an increase in the level of AI innovation can reduce regional carbon intensity.

To interpret the magnitude: a 1% increase in AI patent count is associated with a 0.067% reduction in carbon intensity (Column 3, two-way FE). This effect is economically meaningful given that the average annual reduction in carbon intensity across Chinese provinces over the study period was approximately 3.2%. Regarding the concern that richer regions may have both more AI and lower carbon intensity, we note that the two-way fixed-effects model controls for time-invariant regional characteristics (including geography and long-term development levels) and common time shocks, thus isolating the within-region variation. The negative coefficient of lnAII therefore reflects the effect of changes in AI innovation within a province over time, not cross-sectional correlation.

The control variables exhibit some instability across specifications, which is expected given their different roles. For instance, Environmental Regulation (lnER) shows a positive coefficient in all models, suggesting that current pollution treatment investments may not yet effectively reduce carbon intensity—possibly because they target end-of-pipe pollutants rather than CO₂. Energy Structure (lnES) loses significance after including fixed effects, indicating that its variation is largely cross-sectional. The high R² (0.922) in the two-way FE model is not unusual for panel data with region and year fixed effects, as these dummies alone often explain most of the variance in carbon emissions (e.g., regional dummies capture climate, industrial base, and energy mix). The descriptive nature of this section is intentional, as the causal interpretation and spatial dynamics are reserved for the spatial econometric results that follow.

5.3. Spatial Econometric Model Results Analysis

5.3.1. Spatial Correlation Analysis

Before conducting spatial econometric analysis, it was necessary to test the variables for spatial correlation. The Moran’s I index is commonly employed to examine spatial correlation. Based on a geographical distance weight matrix, the Moran’s I indices for provincial carbon intensity and AI innovation level in China from 2010 to 2023 were calculated, and the results are presented in

Table 4. Moran’s I indices for provincial carbon intensity and AI innovation level in China, 2010–2023.

Year	lnCI	lnAII
2010	0.2594 ***	0.2706 ***
2011	0.2667 ***	0.2228 ***
2012	0.2673 ***	0.2179 ***
2013	0.2626 ***	0.2379 ***
2014	0.2531 ***	0.2679 ***
2015	0.2665 ***	0.3000 ***
2016	0.2584 ***	0.2865 ***
2017	0.2657 ***	0.3021 ***
2018	0.2588 ***	0.2953 ***
2019	0.2535 ***	0.2760 ***
2020	0.2521 ***	0.2929 ***
2021	0.2521 ***	0.2835 ***
2022	0.2463 ***	0.2825 ***
2023	0.2511 ***	0.3016 ***

Note: *** denotes significance at the 1% level.

. During the study period, the Moran’s I indices for carbon intensity and AI innovation level were significantly positive, all passing the 1% significance test. This indicates that the variables exhibit significant positive spatial autocorrelation, suggesting similarity in carbon intensity and AI innovation level between geographically adjacent regions.

5.3.2. Spatial Econometric Model Specification Tests

Prior to conducting regression using spatial econometric methods, LM tests were performed to diagnose the presence of spatial effects in the variables. Subsequently, LR tests and Wald tests were conducted to confirm the appropriate model specification. The test results presented in

Table 5. Spatial econometric model specification test results.

Test	Statistic	Test	Statistic
LM-error	20.71 ***	LM-lag	8.574 ***
Robust LM-error	18.957 ***	Robust LM-lag	6.821 ***
LR test (SDM vs. SAR)	77.13 ***	Wald test (SDM vs. SAR)	86.08 ***
LR test (SDM vs. SEM)	76.04 ***	Wald test (SDM vs. SEM)	71.01 ***

Note: *** denotes significance at the 1% level.

indicate that the Spatial Durbin Model should be selected for analyzing the impact of AI innovation on carbon intensity.

5.3.3. Spatial Durbin Model Regression Results

The Spatial Durbin Model with two-way fixed effects was employed for regression, and the results are presented in

Table 6. Spatial Durbin Model regression results.

Variable	(1) Main Effect	(2) Spatial Spillover Effect	(3) Spatial Autoregressive Coefficient
lnAII	−0.037 *** (0.008)	0.060 *** (0.015)
lnER	0.025 *** (0.007)	−0.003 (0.014)
lnES	0.260 *** (0.047)	−0.771 *** (0.127)
lnIND	0.100 *** (0.031)	0.238 *** (0.076)
lnTB	0.090 *** (0.026)	0.046 (0.060)
rho			0.277 *** (0.047)
Observations	420	420	420
R2	0.752	0.877	0.922
Number of regions	30	30	30
Region fixed effects	No	Yes	Yes
Year fixed effects	No	No	Yes

Note: Standard errors are in parentheses. *** denote significance at the 1% levels.

.

The estimated coefficient for the main effect of AI innovation level was −0.037, which was significantly negative, consistent with the estimation results from the two-way fixed-effects panel model. This finding reaffirms the carbon reduction effect of AI innovation. Concurrently, the spatial spillover effect coefficient for AI innovation level was 0.060, significantly positive at the 1% level. This indicates that the AI innovation level in surrounding regions influences carbon intensity in a given region, confirming the existence of spatial spillover effects. The spatial autoregressive coefficient was significantly positive, suggesting that, due to spatial spillover effects, carbon intensity in one region is positively influenced by carbon intensity in neighboring regions. The spatial effects of AI innovation on carbon intensity can be decomposed into direct effects, indirect effects, and total effects. The decomposition results are presented in

Table 7. Decomposition of spatial effects.

Variable	(1) Direct Effect	(2) Indirect Effect	(3) Total Effect
lnAII	−0.034 *** (0.008)	0.069 ** (0.020)	0.035 (0.022)
lnER	0.025 *** (0.007)	0.005 (0.019)	0.030 (0.021)
lnES	0.226 *** (0.046)	−0.911 *** (0.154)	−0.685 *** (0.154)
lnIND	0.113 *** (0.032)	0.348 *** (0.102)	0.462 *** (0.115)
lnTB	0.094 *** (0.027)	0.092 (0.076)	0.186 * (0.087)
Number of Observations	420	420	420
Number of Regions	30	30	30
Region Fixed Effects	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes

Note: Standard errors are in parentheses. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

.

The direct effect of AI innovation level was significantly negative, indicating that an increase in AI innovation level in a region significantly reduces carbon intensity in that region. The indirect effect of AI innovation level was significantly positive, suggesting that an increase in AI innovation level in one region increases carbon intensity in surrounding regions. While the increase in AI innovation level is beneficial for reducing carbon intensity through its direct effect, its positive indirect effect may adversely impact carbon intensity. This finding implies that the current carbon reduction effect of AI innovation still possesses certain limitations, being insufficiently profound and not yet universal across all relevant domains. Therefore, when formulating carbon emission reduction policies, it is essential to consider not only the direct effect of AI innovation on carbon intensity but also the potential indirect effects it may trigger, particularly negative aspects such as pollution transfer.

5.4. Heterogeneity Analysis

Considering China’s geographical and economic characteristics, the full sample was divided into the western region and non-western regions (with non-western regions including the eastern, central, and northeastern regions) for heterogeneity analysis. The grouped regression results are presented in

Table 8. Heterogeneity analysis results.

Variable	(1)	(2)	(1)	(2)
	Western Region	Non-Western Regions	Higher R&D Intensity Regions	Lower R&D Intensity Regions
Main Effect	0.006 (0.018)	−0.069 *** (0.009)	−0.103 *** (0.016)	−0.039 * (0.020)
Spatial Spillover Effect	0.023 (0.029)	0.013 (0.016)	−0.118 *** (0.028)	0.052 *** (0.019)
Direct Effect	0.010 (0.019)	−0.069 *** (0.009)	−0.094 *** (0.016)	−0.034 * (0.020)
Indirect Effect	0.041 (0.039)	−0.005 (0.019)	−0.069 * (0.037)	0.055 * (0.024)
Control Variables	Yes	Yes	Yes	Yes
Region Fixed Effects	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes
Number of Observations	154	266	210	210
Number of Regions	11	19	15	15

Note: Standard errors are in parentheses. ***, * denote significance at the 1%, and 10% levels, respectively.

.

The heterogeneity analysis results indicated that the effect of AI innovation level on carbon intensity differed across economic belts. The main effect coefficient for AI innovation was significantly negative in non-western regions, whereas the coefficient for the western region was close to zero and not statistically significant. Regional economic development levels may influence the application of AI innovation achievements and their carbon reduction effects. Non-western regions possess more developed economic foundations and higher levels of AI innovation, enabling AI innovation achievements to be more widely applied across various fields, thereby reducing carbon intensity. In contrast, economic development in the western region is relatively lagging, and the application of AI innovation achievements may be comparatively limited, resulting in insignificant carbon reduction effects. Differences in industrial structure and energy structure between regions may also affect the effectiveness of AI innovation on carbon intensity. Non-western regions exhibit more diversified and service-oriented industrial structures, making carbon emission reduction relatively easier to achieve. The western region, however, relies more heavily on energy-intensive industries, where the application of AI innovation achievements yields relatively smaller carbon reduction effects for these sectors.

The spatial spillover effect coefficients for AI innovation were positive in both the western and non-western regions but did not pass significance tests. This finding suggests that the current spatial diffusion characteristics of AI innovation may depend more on heterogeneous linkages between the western and non-western regions, rather than on balance diffusion within each respective region. In the full sample regression results, the significant spatial spillover effect may have arisen from the transmission of AI innovation’s impact on carbon emissions from non-western regions to the western region. While enhancing their AI innovation levels, non-western regions may transfer high-carbon industries to the western region. Grouped regressions shield this cross-regional diffusion channel, making it difficult to capture significant spatial spillover effects.

Research and development investment facilitates technological innovation and improves green economic benefits [32]. The research and development of AI innovation achievements require R&D investment as support. Across different regions, disparities in R&D intensity may affect the quality and application level of AI innovation achievements, thereby leading to variations in the carbon reduction effects of AI innovation. Following the approach of Zhao and Qian [33], the ratio of each region’s internal R&D expenditure to GDP was employed to measure R&D intensity, and the full sample was divided into higher R&D intensity regions and lower R&D intensity regions for heterogeneity analysis. The specific implementation method involved calculating the mean R&D intensity for each region from 2010 to 2023. The median value of mean R&D intensity was 0.015. The 15 provinces with mean R&D intensity above the median were classified as higher R&D intensity regions, while the 15 provinces with mean R&D intensity below the median were classified as lower R&D intensity regions. The grouped regression results are presented in Table 8.

In higher R&D intensity regions, the direct and indirect effect coefficients for AI innovation were significantly negative, with the spatial spillover effect reducing carbon emissions in surrounding areas. This finding suggests that among higher R&D intensity, the possibility of carbon emission transfer triggered by increased AI innovation levels is relatively low, exhibiting a greener and more stable development trend. In lower R&D intensity regions, however, the direct effect of AI innovation was negative while the indirect effect was positive, with the effect coefficients consistent with the full sample regression results. This indicates that in higher R&D intensity regions, AI innovation significantly reduces carbon intensity, while technology diffusion effects simultaneously reduce carbon intensity in surrounding areas. Research and development investment can promote AI technology research and development, accelerate breakthroughs and applications of AI technology, thereby improving the quality of AI innovation achievements and enhancing their application effectiveness in areas such as energy conservation, emission reduction, and intelligent control, thus generating carbon reduction effects. Concurrently, R&D investment not only drives AI technology innovation but also contributes to forming a more complete and active innovation ecosystem. Collaboration and exchange among research institutions, higher education institutions, and technology enterprises supported by R&D investment facilitate technology transfer and application, enabling AI innovation achievements to penetrate more widely across various industries and fields, further enhancing their carbon reduction effects.

5.5. Mechanism Test

Both the two-way fixed-effects panel model and the Spatial Durbin Model demonstrated that AI innovation level can reduce carbon intensity. However, to better guide practical activities, the pathways of emission reduction warrant further investigation. Industrial structure has been widely substantiated as an important factor influencing carbon intensity, with industrial structure optimization capable of reducing carbon intensity. Therefore, a mechanism test was conducted to examine the effect of AI innovation level on industrial structure.

Table 9. Mechanism test results.

Variable	(1) lnCI	(2) lnAIS
Main Effect	−0.037 *** (0.008)	0.003 ** (0.001)
Spatial Spillover Effect	0.060 *** (0.015)	−0.006 * (0.003)
Direct Effect	−0.034 *** (0.008)	0.004 ** (0.001)
Indirect Effect	0.069 ** (0.020)	−0.005 * (0.003)
Control Variables	Yes	Yes
Region Fixed Effects	Yes	Yes
Year Fixed Effects	Yes	Yes
Number of Observations	420	420
Number of Regions	30	30

Note: Standard errors are in parentheses. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

reports the mechanism test results.

From the direct effect coefficients, it can be observed that the promoting effect of AI innovation level on local industrial structure upgrading was significant. This demonstrates that AI innovation can achieve carbon emission reduction by promoting industrial structure upgrading. While our two-step approach cannot fully rule out reverse causality (e.g., industrial upgrading might also attract AI innovation), the temporal structure of our panel (AI patents measured before the annual outcome) and the inclusion of year and region fixed effects reduce the likelihood of simultaneity bias. Future research using instrumental variables for AI innovation would further strengthen causal identification. Concurrently, the indirect effect coefficient of AI innovation on industrial structure upgrading was significantly negative, indicating that an increase in AI innovation level in one region hinders industrial structure upgrading transformation in surrounding regions. This phenomenon may be attributed to the development of AI technology causing the transfer of primary and secondary industries from the focal region to surrounding areas. Governments may choose to prioritize supporting technological innovation and industrial upgrading in core regions, resulting in surrounding areas being disadvantaged in terms of policy support and resource allocation. Through adjusting industrial structure, AI innovation level can indirectly influence carbon intensity. Notably, the direct and indirect effects operate in opposite directions, a phenomenon that warrants attention in regional integrated development.

5.6. Robustness Tests

To verify the robustness of the Spatial Durbin Model estimation results, robustness tests were conducted by excluding municipality samples, altering the time window, and replacing the spatial weight matrix. The obtained results are presented in

Table 10. Robustness test results.

Variable	(1) Excluding Municipalities	(2) Altered Time Window	(3) Replaced Spatial Weight Matrix
Main Effect	−0.029 ** (0.012)	−0.030 ** (0.012)	−0.055 ** (0.024)
Spatial Spillover Effect	0.056 *** (0.018)	0.071 *** (0.019)	0.080 *** (0.025)
Direct Effect	−0.025 * (0.013)	−0.026 ** (0.012)	−0.054 *** (0.017)
Indirect Effect	0.068 * (0.025)	0.086 ** (0.028)	0.081 *** (0.027)
Control Variables	Yes	Yes	Yes
Region Fixed Effects	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes
Number of Observations	364	330	420
Number of Regions	26	30	30

Note: Standard errors are in parentheses. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

.

From the results presented in Column (1) of Table 10, after excluding municipality samples, the coefficient directions for all effects remained consistent with the full sample regression results and were all significant. This indicates that the spatial effects of AI innovation on carbon emissions are robust. From the results presented in Column (2) of Table 10, it can be seen that after altering the time window (excluding the COVID-19 pandemic impact period from 2020 to 2022), the coefficient directions for all effects remained consistent with the full sample regression results and were all significant. This suggests that the spatial effects of AI innovation on carbon emissions were not affected by the exogenous shock of the COVID-19 pandemic, demonstrating robustness. From the results presented in Column (3) of Table 10, it can be seen that after replacing the spatial weight matrix, the coefficient directions for all effects remained consistent with the full sample regression results and were all significant. This indicates that the spatial effects of AI innovation on carbon emissions do not depend on a specific spatial weight matrix, confirming robustness.

5.7. Discussion

The empirical results of this study yield several findings that merit further discussion in light of existing literature and theoretical expectations.

The negative direct effect of AI innovation on local carbon intensity (direct effect = −0.034, p < 0.01) aligns with the growing body of evidence that AI applications can enhance energy efficiency and facilitate the green transition [6,7,8,24]. However, our effect size is smaller than estimates based solely on industrial robot density [7], suggesting that a broader measure of AI innovation (including patents across all sectors) captures a more diverse set of applications, some of which may have weaker carbon reduction impacts. This finding extends the literature by showing that the carbon reduction benefit of AI is not limited to manufacturing but is detectable at the provincial level when using a comprehensive innovation measure. Similar evidence from cross-country studies [22] and sectoral analyses [10] supports the general conclusion that technological innovation reduces carbon intensity, though the magnitude varies by context [4,5]. Accordingly, Hypothesis 1 is supported.

Turning to spatial dynamics, the positive indirect effect (0.069, p < 0.05) provides novel evidence of a spatial “leakage” effect: AI innovation in one region increases carbon intensity in neighboring regions. This finding contributes to the environmental economics literature by demonstrating that technological innovation, while beneficial locally, can exacerbate interregional environmental inequality through industrial relocation and factor mobility [24,26]. Our results echo the pollution haven hypothesis [25] but with an important twist—the polluting industries are not fleeing environmental regulation per se, but rather the productivity-driven agglomeration forces induced by AI innovation. This interpretation is consistent with studies on industrial intelligence and structural change [23], which show that AI-driven productivity gains concentrate economic activities in technologically advanced regions, pushing traditional industries outward. Thus, Hypothesis 2 is supported.

The heterogeneity results reveal that the carbon reduction effect of AI is conditional on regional absorptive capacity. In higher R&D intensity regions, the indirect effect becomes significantly negative (−0.069, p < 0.1), meaning that AI innovation in these regions actually reduces carbon intensity in surrounding areas. This suggests that when regions have strong innovation ecosystems, the technology diffusion channel dominates over the industrial relocation channel. This finding supports the concept of “green spillovers” from technologically advanced regions [22,26] and offers a policy lever: investing in R&D can transform AI’s spatial spillovers from positive (harmful) to negative (beneficial). Similar threshold effects have been documented in the context of green credit and firm innovation [30], as well as digital technology and green performance [31], reinforcing the notion that absorptive capacity is a critical moderator. Hypothesis 3 (heterogeneity across economic belts and R&D intensity) is partially supported: the effect is significant in non-western regions but insignificant in the western region; and the direction of spillovers flips from positive to negative under high R&D intensity.

The mechanism test confirms industrial structure upgrading as a mediator. The positive direct effect of AI on industrial structure upgrading (0.004, p < 0.05) and the negative indirect effect (−0.005, p < 0.1) indicate that the same AI innovation that upgrades local industry mix may hinder such upgrading in neighbors. This paradox—AI helping one region’s green transition while delaying another’s—has not been previously documented and calls for coordinated regional policies. The mediating role of industrial structure is consistent with the broader literature on structural change and carbon emissions [27,28,29], which demonstrates that shifting the industrial mix toward services and high-tech sectors reduces carbon intensity. Our study extends this literature by showing that AI innovation is a driver of such structural change, and that this driver operates unevenly across space. Therefore, Hypothesis 4 is supported.

In summary, this study makes three distinctive contributions to the literature. It shifts the focus from industrial robots to AI patents, capturing economy-wide innovation capacity. It uncovers the spatial paradox of AI innovation: local carbon reduction versus neighboring carbon increase. It identifies R&D intensity as a critical moderator that can reverse the sign of spatial spillovers. These findings suggest that the net global effect of AI on carbon emissions depends crucially on the spatial distribution of innovation capacity and the presence of technology diffusion mechanisms, an insight that aligns with recent evaluation frameworks emphasizing the contextual nature of AI’s sustainability impacts.

6. Conclusions and Policy Recommendations

6.1. Research Conclusions

Based on provincial panel data from China covering the period from 2010 to 2023, this study employed the Spatial Durbin Model to investigate the spatial effects of AI innovation level on regional carbon intensity. The conclusions drawn are as follows: (1) An increase in the level of AI innovation in a region significantly reduces local carbon intensity. (2) An increase in the level of AI innovation in a region may increase carbon intensity in surrounding areas. (3) AI innovation exhibits a significant carbon reduction effect in non-western regions, whereas its impact on carbon intensity in the western region is not statistically significant. (4) In regions with higher research and development (R&D) intensity, an increase in AI innovation level reduces carbon intensity in both local and surrounding areas. (5) AI innovation can indirectly influence carbon intensity through the adjustment of industrial structure, indicating that industrial structure mediates the relationship between AI innovation level and carbon intensity.

6.2. Theoretical Contributions

This study advances theoretical understanding in three ways. First, it extends the Environmental Kuznets Curve framework by unpacking the “technique effect” into AI-specific mechanisms (efficiency, substitution, structural effects). Rather than treating technology as an aggregate factor, we show that the type of innovation matters for environmental outcomes. Second, we contribute to spatial spillover theory by identifying three distinct channels (technology diffusion, industrial relocation, factor mobility) through which AI innovation affects neighboring regions’ carbon intensity, and by demonstrating that the net spillover sign can be positive or negative depending on regional absorptive capacity (R&D intensity). Third, we propose and empirically test a mediation model linking AI innovation, industrial structure upgrading, and carbon intensity, thereby clarifying the mechanism behind AI’s environmental impact. These theoretical refinements offer a more nuanced understanding of technology–environment relationships in the context of rapid technological change.

6.3. Policy Implications

Building on the theoretical contributions above, several policy implications emerge. First, to maximize the local carbon reduction effect of AI, governments should invest in AI R&D and facilitate technology diffusion within regions. Second, to mitigate negative spatial spillovers, cross-regional coordination is essential: high-innovation regions should support neighboring regions in adopting green AI technologies, rather than simply transferring polluting industries. Third, differentiated strategies are needed: non-western regions (eastern, central, northeastern) should leverage AI for industrial upgrading, while western regions may require technology transfer and capacity building before AI can yield significant carbon benefits. Fourth, given the mediating role of industrial structure, policies that promote AI-driven industrial upgrading (e.g., subsidies for green AI applications, carbon pricing) can amplify the carbon reduction effect. Fifth, the finding that higher R&D intensity reverses the sign of spatial spillovers suggests that investments in regional innovation systems are not only good for local growth but also generate positive environmental externalities for neighbors. Given the correlational nature of our analysis, these policy recommendations should be interpreted as conditional on the empirical patterns observed; future research with stronger causal identification is needed to confirm the direction of effects. In particular, the recommendation for cross-regional coordination directly addresses the identified risk of spatial emission spillovers: high-innovation regions should actively support neighboring regions’ green transitions to avoid simply displacing carbon emissions.

6.4. Limitations and Future Research

Several limitations of this study should be acknowledged to guide future research. First, regarding data, AI patents measure the quantity of innovation but not its quality, economic value, or actual deployment. A province with many patents may not necessarily have widespread AI adoption in carbon-relevant sectors. Future research could complement patent data with survey-based adoption metrics or AI investment data. Second, the provincial level of analysis masks within-province heterogeneity, particularly between urban and rural areas or between developed and lagging cities within the same province. City-level analysis, when data become available, would provide finer-grained insights. Third, concerning methodology, while the Spatial Durbin Model accounts for spatial dependence, it does not fully address potential endogeneity or reverse causality. For instance, regions with ambitious carbon reduction policies might simultaneously invest more in AI innovation. Our fixed-effects specification controls for time-invariant confounding but not for time-varying omitted variables. Future studies could employ instrumental variables (e.g., historical telephone density or university AI programmes) or quasi-natural experiments (e.g., AI policy shocks) to strengthen causal identification. Fourth, the energy consumption of AI computing itself—including the carbon footprint of training large models—is not accounted for due to data constraints. As AI models grow, this limitation becomes increasingly relevant. Fifth, the study focuses on China; generalising the findings to other developing or developed countries requires caution given China’s unique institutional and technological context. Despite these limitations, the robustness checks (alternative spatial weight matrices, exclusion of municipalities, altered time windows) increase confidence in the main findings.