Can Artificial Intelligence Effectively Improve China’s Environmental Quality? A Study Based on the Perspective of Energy Conservation, Carbon Reduction, and Emission Reduction

Abstract

The “technological dividends” brought by AI development provide a new model for the country to achieve green governance, enhance enterprises’ ability to manage pollutant emissions during production and operations, and create a new driving force for improving environmental quality. In this regard, this paper systematically examines the impact of AI on environmental quality in China by employing provincial panel data spanning from 2000 to 2020. Focusing on energy conservation, carbon reduction, and emissions mitigation, the analysis is conducted through the application of a two-way fixed-effects model and mediation effects model to explore both the effects and the mechanisms of AI’s influence on environmental quality. The findings indicate that the development and implementation of AI contribute positively to China’s efforts in energy conservation, carbon reduction, and emissions mitigation, ultimately leading to an enhancement in environmental quality. This conclusion remains valid after multiple robustness checks. Mechanism tests reveal that the optimization of regional energy structures, advancements in green technological innovation, and upgrades in industrial structures serve as crucial pathways through which AI facilitates energy conservation, carbon reduction, and emissions mitigation. Heterogeneity analysis uncovers a notable “path dependence” effect in China’s AI development; regions characterized by higher material capital investment, more advanced technological market development, and greater levels of marketization experience a relatively more pronounced impact of AI on the enhancement of environmental quality. This study offers direct references and practical insights for countries globally to foster AI development, enhance environmental quality, and advance high-quality economic growth amid the ongoing wave of digital and intelligent transformation.

1. Introduction

To bolster comprehensive national strength and expedite modernization, China has long pursued a growth-first development strategy, which has propelled the nation to become the world’s second-largest economy at a remarkable pace. Along with rapid economic development, energy consumption, greenhouse gas emissions, and industrial pollutant emissions have been increasing, posing severe challenges to environmental quality [1]. In this context, the challenge of continuously enhancing ecological environmental quality while promoting a comprehensive green and low-carbon transformation in economic and social development has emerged as a critical issue confronting all sectors of society. In fact, achieving the goals of energy conservation, carbon reduction, and emission reduction is essentially a process of transitioning to technology-intensive industries. As an important hallmark of the new wave of technological advancement, artificial intelligence (AI) has opened up new pathways for green transition development and brought new technological innovations to environmental pollution control [2]. So, can AI effectively achieve energy conservation, carbon reduction, and emissions mitigation, and how can it reasonably improve environmental quality? These are important practical issues worth in-depth analysis and research. Clarifying and answering these questions will not only fundamentally improve environmental quality but also provide significant practical references for China and other countries worldwide in promoting the organic integration of innovation-driven and green development strategies.

AI demonstrates two basic attributes in macroeconomic operations: the attribute of an automation tool and the attribute of a general-purpose technology. As an automation tool, AI can minimize energy and resource consumption in the production process, particularly for repetitive tasks with standardized procedures. This leads to more efficient production and greater control over quality, resulting in cost savings while simultaneously enhancing production efficiency. As a general-purpose technology, AI’s advantages of universality, continuous technological progress, and innovation complementarity stand out [3], giving it the potential to empower and innovate across most production sectors, injecting new momentum into economic growth. Mainstream research on the effects of AI technology predominantly concentrates on its influence on income distribution, labor structure, and employment. However, studies examining the impact of AI technology on environmental quality remain relatively limited [4].

Further review of existing related research reveals that AI technology may be a “double-edged sword” for improving environmental quality. On the one hand, the deployment of AI can enhance output while simultaneously reducing energy consumption [5], advancing green industrial transformation, and decreasing carbon emissions [6]. Furthermore, these technologies introduce innovative paradigms for mitigating urban pollution and steering green, low-carbon transitions [7]. Therefore, many scholars believe that empowering the economic green transition with AI technology can ensure ecological benefits while achieving economic benefits [8,9]. On the other hand, research by some scholars suggests that the widespread application of AI in the industrial field will result in continuous expansion of production scale, which may result in greater energy consumption, increasing the energy burden in economic development [10,11]; it may also lead to increased carbon emissions, causing more severe environmental problems [12,13].

Simultaneously, some studies have also discovered that AI technology’s impact on energy conservation and emission reduction is heavily contingent on whether the regional institutional environment is well-developed, showing significant regional heterogeneity [14]. This implies that the governance of environmental pollution by AI is not merely the result of technological innovation itself but is also constrained by other factors such as regional technological levels, physical capital investment, and marketization levels [15]. Yuan et al. [13] found that in countries or regions where the institutional environment is not yet well-developed, the industrial intelligent transformation brought about by AI technology may not promote carbon reduction and may even have a negative regulatory effect. However, Luo and Feng [7] found that the deployment of AI technologies can effectively reduce regional pollutant emissions. Furthermore, the impact of AI on urban pollutant emissions shows significant variation due to differences in regional endowment factors. Thus, current studies on AI’s environmental impact remain divided, with limited literature assessing the overall influence of AI on energy conservation, carbon reduction, or emissions mitigation. Therefore, it is essential to elucidate and systematically examine AI technology’s environmental effects, along with its regional heterogeneity.

Additionally, the mechanisms and transmission channels through which AI affects environmental quality have also attracted the attention of some scholars. In terms of energy consumption, AI technology can improve energy production techniques, thus lowering energy use in the production processes of enterprises [16]. Regarding carbon emissions, AI technology can effectively decrease carbon dioxide emissions and foster green and sustainable economic growth through mechanisms such as reducing energy consumption [17], improving energy utilization [18], promoting industrial structure upgrading [19], or fostering green technological innovation [20]. Moreover, AI can reduce resource consumption and pollution emissions by restructuring traditional energy consumption patterns [21], enhancing public awareness and participation [22], or optimizing capital productivity [5]. In general, AI technology can mostly achieve the goal of improving environmental quality through technological effects and structural effects.

In brief, previous research mostly studies the influence of AI applications on specific aspects such as carbon emissions or energy consumption, meaning systematic exploration of the environmental effects of AI and its mechanisms at play remains insufficient. Consequently, it is difficult to assess whether and how AI impacts environmental quality. Moreover, the existing literature lacks analysis of the heterogeneous impacts of AI on environmental quality across different regional environments, and the research results are still divided. Therefore, this paper aims to systematically explore whether and how AI can effectively enhance China’s environmental quality from a broader perspective encompassing energy conservation, carbon reduction, and emissions mitigation. This inquiry undoubtedly provides a valuable contribution to the existing body of research.

Therefore, in comparison to existing literature, the central themes and incremental contributions of this study are mainly reflected in the following three aspects. First, by comprehensively considering energy consumption, carbon emissions, and industrial pollutant emissions, this paper systematically examines the underlying logic and effectiveness of AI in enhancing regional environmental quality, providing direct empirical evidence to support society’s green transition. Second, this paper systematically discusses the specific mechanisms through which AI improves environmental quality from multiple aspects, such as regional energy structure, green technological innovation, and industrial structure, providing important references for understanding the environmental effects of AI. Third, this study investigates the heterogeneous impacts of AI on environmental quality across various regional endowment conditions, considering multiple dimensions such as physical capital, technology market development, and the level of marketization. This analysis provides practical leverage for advancing regional AI development and enhancing environmental quality.

The remainder of this paper is structured as follows. Section 2 provides a systematic elaboration of the theoretical framework and research hypotheses. Section 3 details the methodology, encompassing model setup, data description, and data collection. Section 4 presents the empirical findings, including baseline model regression, endogeneity treatment, and robustness checks. Section 5 delves into the mechanisms through which AI contributes to energy conservation, carbon reduction, and emissions mitigation, along with its heterogeneous effects. Section 6 discusses the results, and Section 7 concludes with research findings and policy implications.

2. Theoretical Analysis and Hypothesis

Against the backdrop of rapid digital technology development, AI technology has become an important support for new digital infrastructure and is widely applied in human production and life. AI, relying on its universality and penetration, has deeply integrated with manufacturing, fostering new green economic models and bringing significant opportunities to achieve the “dual carbon” goals. First, the automation control and rapid calculation capabilities of AI technology provide new ways to save energy, reduce energy consumption, and promote green production and consumption [23]. Second, AI can conduct technical detection and tracking, using big data to accurately grasp the production and energy consumption situation of factories, increasing the methods and efficiency of obtaining environmental information, improving enterprises’ resource allocation efficiency, and thus, enhancing operational efficiency to achieve green development [24,25]. Finally, AI technology can predict pollution emissions in industrial production through intelligent simulation and perceptual prediction, enhancing industrial enterprises’ decision-making ability to respond to sudden environmental events and providing a basis for energy conservation, carbon reduction, and emissions reduction [26]. Drawing from the above analysis, Research Hypothesis 1 is proposed.

Hypothesis 1:

The deployment of AI technology facilitates energy conservation, carbon reduction, and emissions reduction, thereby enhancing environmental quality.

Energy structure optimization effect. The adjustment of the energy structure is a crucial factor influencing energy conservation, carbon reduction, and emissions reduction. The rapid integration of AI technology significantly affects the supply–demand dynamics and consumption patterns of energy. The impact of achieving energy conservation, carbon reduction, and emissions mitigation through technological advancement will primarily manifest in the energy consumption structure, particularly through the decreased demand for fossil fuels and the optimization of energy use efficiency [27,28]. From one perspective, the application of AI technology has changed traditional business and consumption models, reducing reliance on heavy industries such as coal, steel, and cement and promoting green economic transformation [29]. Alternatively, the advancement of AI technology can significantly enhance energy efficiency and facilitate the development and utilization of renewable energy, demonstrating substantial potential in reducing carbon emissions [30]. AI has enhanced the efficiency in producing, selling, and utilizing products and services within traditional high-energy-consuming enterprises, thereby altering regional energy consumption structures and enabling technological emission reductions. Although the large-scale application of AI technology may reduce energy prices, leading to increased energy demand and intensified competition between fossil fuels and renewable energy, the overall effect is still positive [31]. Drawing from the above analysis, this section proposes Research Hypothesis 2.

Hypothesis 2:

AI can facilitate energy conservation, carbon reduction, and emissions reduction by advancing the optimization of regional energy structures.

Green technological innovation effect. The deployment of AI technology can stimulate green technological innovation by optimizing the innovation environment, increasing technology spillover, and lowering the costs associated with green R&D innovation. Green technological innovation is a key force in environmental governance [32,33]. Specifically, first, the digital algorithms, intelligent applications, and network communication technologies contained in AI can provide excellent hardware foundations and technical support for enterprises’ green R&D, thereby promoting green technological innovation. Second, AI, through the typical “machine replacement”, not only helps accelerate the realization of economies of scale and expand profit margins but also supports economic entities in investing more personnel and resources in green R&D, thereby promoting green technological progress.

Green technological innovation serves as an intrinsic driver for reducing environmental pollution, with the capacity to alleviate the adverse effects of technological advancements on the ecological environment and to promote the enhancement of environmental quality [34]. On one side, the thorough integration of green technology elements into manufacturing will bring transformative changes to traditional production models, promoting the transition of production processes toward greener and cleaner intelligent development. Alternatively, green technological innovation will also contribute to emission reduction during the end-of-pipe control stage by continually enhancing the pollution control capabilities of stakeholders, optimizing pollution management efficiency, and improving environmental performance [13]. Drawing from the above analysis, Research Hypothesis 3 is proposed.

Hypothesis 3:

AI can facilitate energy conservation, carbon reduction, and emissions reduction by promoting green technological innovation.

Industrial structure upgrading effect. The evolution of industrial structures and the replacement of dominant industries imply the reallocation of resources and are important factors through which AI technology impacts environmental pollution [35,36]. In recent years, high-tech and intelligent service industries powered by AI have emerged as key drivers of high-quality economic development in China. From one perspective, the automation and intelligent tool attributes of AI can optimize the use of fossil energy in traditional industrial sectors, promoting the continuous extension of high-energy-consuming, high-pollution industries and their integration with the tertiary industry. This drives the gradual advancement of traditional industries toward higher levels, resulting in reduced pollutant emissions [37]. Conversely, the general purpose technology attribute of AI drives the formation of new industrial patterns in various industries, promoting the rise of new intelligent industries [38]. As a result, emerging industries with higher productivity and technological levels rapidly develop, guiding the traditional industrial structure dominated by labor and capital-intensive industries to transition toward a cleaner industrial structure dominated by technology and knowledge-intensive industries, thereby reducing energy consumption, carbon emissions, and industrial pollutant emissions. Therefore, this section proposes Research Hypothesis 4.

Hypothesis 4:

AI can facilitate energy conservation, carbon reduction, and emission reduction by promoting industrial structure upgrading.

Drawing from Research Hypotheses 1 to 4, this paper develops a theoretical framework illustrating the mechanism through which AI impacts environmental quality, as depicted in Figure 1.

3. Methods and Data

3.1. Study Area

Considering the availability and consistency of research data, we ultimately selected provincial panel data from 30 provinces, autonomous regions, and municipalities in China for the period 2000 to 2020, excluding Hong Kong, Macao, and Taiwan. Because of significant data deficiencies for certain indicators in Tibet, data from this region were also excluded.

3.2. Model Setting

To verify whether and how AI affects environmental quality, the following baseline model is constructed:

$$E{Q}_{i,t}={\alpha }_0+{\alpha }_{1}A{I}_{i,t}+{\displaystyle \sum {\gamma }_i}{\kappa }_{i,t}+{\sigma }_i+{\tau }_t+{\epsilon }_{it}$$

(1)

Here, the subscripts i and t represent the province and time, respectively; EQ represents regional environmental quality, specifically regional energy consumption, carbon emissions, and industrial pollutant emissions; AI represents the level of artificial intelligence in the region. $\kappa$ represents a series of control variables, including the level of economic development, labor force level, human capital level, infrastructure level, economic openness level, foreign direct investment, and environmental regulation, among others. The same applies hereinafter. ${\sigma }_i$ represents individual fixed effects, ${\tau }_t$ represents time fixed effects, and ${\epsilon }_{it}$ represents random disturbance terms.

To empirically examine the specific mechanisms through which AI contributes to energy conservation, carbon reduction, and emissions reduction across different regions of China, the following mediating effect model will be constructed in this study, based on Hypotheses 2 through 4:

$$IN{V}_{it}={\varphi }_0+{\varphi }_{1}AI{i}_t+{\displaystyle \sum o_i}{\kappa }_{i,t}+{\sigma }_i+{\tau }_t+{\epsilon }_{it}$$

(2)

$$E{Q}_{i,t}={\theta }_0+{\theta }_{1}A{I}_{i,t}+{\theta }_{2}IN{V}_{i,t}+{\displaystyle \sum {\zeta }_i}{\kappa }_{i,t}+{\sigma }_i+{\tau }_t+{\epsilon }_{it}$$

(3)

Among them, Equations (2) and (3) combined with Equation (1) constitute the mediating effect model constructed in this paper. INV denotes the mediating variables, specifically referring to energy structure, green technological innovation, and industry composition. ${\alpha }_1$ represents the overall effect of AI development on regional environmental quality, ${\varphi }_1{\theta }_2$ is the mediating effect transmitted through the mediating variables, and ${\theta }_1$ is the direct effect. If ${\varphi }_1$ and ${\theta }_2$ are both significant, then a mediating effect exists, and Hypotheses 2 to 4 are valid.

3.3. Data Description

3.3.1. Environmental Quality

There are many ways to measure regional environmental quality (EQ). This paper draws on and expands existing typical relevant studies by Du et al. [39], Khan et al. [40], Wen and Liu [41], etc., to comprehensively evaluate regional environmental quality from three aspects: energy consumption, carbon emissions, and industrial pollutant emissions. The specific proxy indicators are as follows.

Energy Consumption (ENC). Using the conversion coefficients of various energy sources to standard coal as provided in the “China Energy Statistical Yearbook,” this study aggregates the energy consumption of nine types of energy, including raw coal, coke, crude oil, gasoline, kerosene, diesel, fuel oil, liquefied petroleum gas, and natural gas, to calculate the total energy consumption for each province in China, and measures it by taking the logarithm.

Carbon Emissions (CO₂). Carbon emissions are calculated using the CO₂ conversion coefficients of standard coal for the nine types of energy mentioned above. The total carbon emissions of each province are obtained by summing the calculated results of the nine energy types and are measured by taking the logarithm.

Industrial Pollutant Emissions (IPE). Drawing on the existing studies by Zhou et al. [42], this study employs the principal component contribution rates of industrial wastewater discharge, industrial SO₂ emissions, and industrial dust emissions to calculate the pollutant emission index for each province.

3.3.2. Artificial Intelligence

For regional artificial intelligence (AI) levels, existing research predominantly assesses AI levels through indicators such as local information dissemination, computing services, and the number of industrial robots installed [36]. However, these indicators capture only one dimension of AI and fail to fully represent the overall development level of regional AI. The number of AI enterprises in a region can, to some extent, reflect the vitality of development in AI research activities, talent aggregation, investment, and application in that region. Therefore, this paper follows the research methodology of Li et al. [25] and others, using the number of AI enterprises per 10,000 residents in each province of China as a proxy indicator for the AI application level in that province. Specifically, the process of collecting the number of AI enterprises in each province involves two key steps. First, the “Tianyancha” platform is utilized to search and filter enterprise data, identifying those whose business scope encompasses AI, robotics, data processing, cloud computing, language recognition, image recognition, and natural language processing as AI enterprises. Enterprises that are active and operational are identified as the statistical objects of AI enterprises. Second, the locations of the enterprises that meet the above search criteria are matched with the provinces, and by aggregating them, the number of AI enterprises in each of the 30 provinces within the sample range of this paper is obtained.

3.3.3. Intermediate Variables

The mediating variables in this paper include regional energy structure (ENS), green technological innovation (GTI), and industrial structure (INU). Following the methodological approach of Ji and Zhang [43], the energy structure is indicated by the proportion of provincial electricity consumption to the national total electricity consumption. Referring to key related studies by Zhang et al. [44] and Zhao et al. [45], the green technological innovation indicator is measured by per capita green patent authorizations across provinces. The higher this value, the higher the level of green technology. The industrial structure is indicated by the proportion of the tertiary industry’s added value relative to the secondary industry. Research by Wang et al. [46] and Zhou et al. [47] demonstrates that the growth of the tertiary industry reflects clean and green production practices. Therefore, this paper underscores, to some extent, that AI facilitates the greening of the industrial structure, thereby promoting sustainable regional environmental progress.

3.3.4. Control Variables

Energy conservation, carbon reduction, and emission reduction in various provinces of China depend on multiple factors. In other words, in addition to the development of AI, other factors also affect energy conservation, carbon reduction, and emission reduction. Based on existing relevant research and the reality of environmental quality development, the following seven control variables are selected: ➀ Economic Development Level (EDL). The EDL is assessed by calculating the natural logarithm of per capita GDP in each province. ➁ Labor Level (LAL). The LAL is assessed by calculating the natural logarithm of the number of employed individuals in each province. ➂ Human Capital Level (HCL). The HCL is indicated by the average number of college students per thousand residents in each province. ➃ Infrastructure Level (INL). The INL is indicated by the natural logarithm of the total freight volume in each province. ➄ Economic Openness Level (EOL). The EOL is assessed by the share of total trade volume relative to GDP in each province. ➅ Foreign Direct Investment (FDI). The level of FDI is indicated by the share of foreign direct investment relative to regional GDP. ➆ Environmental Regulation Level (ERL). The ERL is assessed by the ratio of investment in industrial pollution control to the industrial added value in each province.

The primary data for the aforementioned indicators are obtained from the Wind database and Environmental Statistical Yearbook. The statistical characteristics of the variables are presented in

Table 1. Descriptive statistics of model variables.

Variables	N	Mean	Std.Dev.	Min	Max
ENC	630	9.193	0.835	5.989	11.05
CO2	630	10.09	0.853	6.781	11.93
IPE	630	1.068	0.715	0.0432	3.938
AI	630	0.526	1.007	0.004	8.885
ENS	630	0.033	0.023	0.003	0.110
GTI	630	0.385	0.739	0.001	5.725
INU	630	1.043	0.572	0.494	5.297
EDL	630	−0.026	0.524	−1.324	1.570
LAL	630	7.547	0.803	5.619	8.859
HCL	630	0.159	0.0747	0.0213	0.413
INL	630	11.19	0.912	8.455	12.98
EOL	630	0.295	0.359	0.008	1.721
FDI	630	0.024	0.022	0.001	0.146
ERL	630	0.005	0.004	0.001	0.031

.

4. Empirical Results

4.1. Baseline Model Regression Results

Based on Equation (1), this section systematically investigates whether AI development can effectively achieve energy conservation, carbon reduction, and emissions mitigation. Drawing from the Hausman test results, a two-way fixed effects model is selected. All models account for both time and province fixed effects, with the regression outcomes presented in columns (1) to (3) of

Table 2. Baseline model and its endogeneity treatment results.

Variables	Results of Baseline Model Estimation			Endogeneity Treatment Results
	(1)	(2)	(3)	(4)	(5)	(6)
	ENC	CO2	IPE	ENC	CO2	IPE
AI	−0.032 **	−0.046 ***	−0.093 ***	−0.043 ***	−0.064 ***	−0.127 ***
	(−2.479)	(−3.399)	(−3.967)	(−3.091)	(−4.413)	(−4.772)
EDL	−0.111 *	−0.119 *	−0.370 ***	−0.201 ***	−0.208 ***	−0.433 ***
	(−1.736)	(−1.773)	(−3.207)	(−3.240)	(−3.191)	(−3.614)
LAL	0.379 ***	0.402 ***	0.371 ***	0.412 ***	0.451 ***	0.430 ***
	(4.899)	(4.959)	(2.666)	(5.503)	(5.702)	(2.960)
HCL	0.676 *	0.920 **	0.048	0.861 **	1.047 ***	−0.101
	(1.923)	(2.496)	(0.076)	(2.450)	(2.824)	(−0.153)
INL	0.314 ***	0.323 ***	−0.008	0.278 ***	0.286 ***	−0.025
	(11.734)	(11.527)	(−0.157)	(10.752)	(10.490)	(−0.503)
EOL	0.048	0.074	0.134	0.048	0.080	0.221 **
	(0.829)	(1.221)	(1.280)	(0.834)	(1.323)	(1.993)
FDI	−1.299 **	−1.345 **	2.784 ***	−1.669 ***	−1.801 ***	2.283 **
	(−2.496)	(−2.466)	(2.972)	(−3.263)	(−3.331)	(2.304)
ERL	5.253 **	5.557 **	8.981 **	5.912 **	6.156 **	7.265
	(2.201)	(2.221)	(2.090)	(2.591)	(2.552)	(1.642)
Constant	2.721 ***	3.303 ***	−1.765	2.400 ***	2.907 ***	−2.333 *
	(3.901)	(4.518)	(−1.406)	(3.671)	(4.213)	(−1.842)
Province fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
Kleibergen-Paap rk LM statistic P				0.000	0.000	0.000
Cragg-Donald Wald F statistic				2104.114	2104.114	2104.114
N	630	630	630	630	630	630
Adj.R2	0.968	0.967	0.861	0.934	0.927	0.836

Note: t statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01.

.

The results indicate that, after controlling for relevant influencing factors, AI significantly contributes to energy conservation, carbon reduction, and emission reduction. Specifically, the coefficient of AI’s impact on energy consumption (ENC) is −0.032, which is significant at the 5% level. Additionally, the coefficient of AI’s impact on carbon emissions (CO₂) is −0.046, significant at the 1% level. The coefficient of AI’s impact on industrial pollutant emissions (IPE) is −0.093, which is significant at the 1% level. These findings suggest that the development of AI can effectively lower total energy consumption, carbon emissions, and industrial pollutant emissions, thereby enhancing environmental quality in China. However, the impact of AI on carbon reduction and emissions mitigation is more pronounced. Overall, the baseline model regression results empirically confirm the positive role of AI in energy conservation, carbon reduction, and emissions mitigation, thereby validating Hypothesis 1 of this paper.

Regarding the other control variables, the coefficient of regional economic development level (EDL) is notably negative, suggesting that improvements in regional economic development contribute to energy conservation, carbon reduction, and emissions mitigation within the context of AI. The coefficients for labor level (LAL), human capital level (HCL), and environmental regulation level (ERL) are predominantly significantly positive, suggesting that the current economic development and regulation models still require further adjustment. The matching degree of labor supply, environmental regulation policies, and economic intelligence needs to be improved. The effects of infrastructure level (INL) and foreign direct investment (FDI) on energy conservation, carbon reduction, and emissions mitigation are uncertain, while the impact of economic openness level (EOL) is not significant.

4.2. Endogeneity Treatment

This study employs the instrumental variable method to address endogeneity issues. Given that optical cable, as a fundamental material for AI development, significantly influences AI development levels but has weak correlations with energy conservation, carbon reduction, and other economic variables, we followed the approach of Zhao et al. [48]. Provincial optical cable density, specifically measured by the length of long-distance optical cable per square kilometer in each province, was used as an instrumental variable for AI development levels. After performing weak instrument variable tests and over-identification tests, we conducted empirical analysis using the 2SLS method. The results are presented in columns (4) to (6) of Table 2. The p-value of the K-P LM statistic demonstrates a strong correlation between the instrumental variable and AI application levels, effectively ruling out the issue of weak instruments. Furthermore, the regression coefficients for AI are significantly negative at the 1% level, providing additional support for Research Hypothesis 1.

4.3. Robustness Test

4.3.1. Replacing Explanatory Variables

To enhance the reliability of the AI proxy indicators and the environmental effects observed in this paper, this section further adopts the method used by Liu et al. [36] and others, using the number of industrial robot installations as a proxy for AI development. The baseline model regression is then rerun, with the results presented in columns (1) to (3) of

Table 3. Robustness test.

Variables	Replacing Explanatory Variables			Driscoll–Kraay Standard Errors			Adding Control Variables
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
	ENC	CO2	IPE	ENC	CO2	IPE	ENC	CO2	IPE
AI	−0.092 **	−0.105 **	−0.297 **	−0.032 **	−0.046 **	−0.093 **	−0.023 *	−0.036 ***	−0.075 ***
	(−4.460)	(−4.740)	(−6.332)	(−2.509)	(−2.473)	(−2.480)	(−1.841)	(−2.789)	(−3.340)
EDL	−0.102 **	−0.090 *	−0.041	−0.111	−0.119	−0.370 **	−0.110 *	−0.117 *	−0.306 ***
	(−2.047)	(−1.676)	(−0.359)	(−1.401)	(−1.501)	(−2.302)	(−1.789)	(−1.821)	(−2.743)
LAL	0.520 ***	0.551 ***	0.809 ***	0.379 ***	0.402 ***	0.371 ***	0.384 ***	0.409 ***	0.493 ***
	(6.041)	(5.967)	(4.141)	(4.054)	(3.819)	(2.942)	(5.076)	(5.171)	(3.601)
HCL	1.347 ***	1.735 ***	1.818 **	0.676	0.920 *	0.048	0.326	0.541	−0.700
	(3.567)	(4.282)	(2.122)	(1.432)	(1.743)	(0.075)	(0.961)	(1.526)	(−1.141)
INL	0.174 ***	0.180 ***	−0.005	0.314 ***	0.323 ***	−0.008	0.335 ***	0.346 ***	0.013
	(6.081)	(5.878)	(−0.078)	(4.954)	(4.881)	(−0.214)	(13.058)	(12.894)	(0.272)
EOL	−0.039	−0.052	−0.081	0.048	0.074	0.134	0.101 *	0.131 **	0.171 *
	(−0.790)	(−0.980)	(−0.721)	(0.490)	(0.684)	(0.764)	(1.803)	(2.230)	(1.690)
FDI	−0.964	−0.864	3.867 **	−1.299 ***	−1.345 ***	2.784 **	−1.818 ***	−1.901 ***	2.124 **
	(−1.624)	(−1.356)	(2.872)	(−3.922)	(−3.778)	(2.839)	(−3.637)	(−3.639)	(2.350)
ERL	3.185	3.505	1.667	5.253	5.557	8.981 *	5.000 **	5.308 **	10.476 **
	(1.473)	(1.511)	(0.340)	(1.310)	(1.341)	(1.780)	(2.184)	(2.218)	(2.531)
RDI							−18.658 ***	−20.072 ***	−29.880 ***
							(−7.828)	(−8.056)	(−6.934)
TAX							1.544 **	1.608 **	−2.037
							(2.188)	(2.180)	(−1.597)
Constant	3.929 ***	4.545 ***	−3.119 *	2.186 ***	2.804 ***	−2.232 **	2.632 ***	3.204 ***	−2.231 *
	(5.077)	(5.474)	(−1.777)	(2.858)	(3.317)	(−2.601)	(3.954)	(4.605)	(−1.854)
Provincefixed effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Timefixed effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	450	450	450	630	630	630	630	630	630
Adj.R2	0.981	0.980	0.9110	0.934	0.951	0.926	0.971	0.970	0.873

Note: t statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01. Due to the availability of data on the number of industrial robot installations across China’s provinces, the regression sample data period for the “Replacing Explanatory Variables” section is from 2006 to 2020.

. The results indicate that the regression coefficients of AI on energy consumption (ENC), carbon emissions (CO₂), and industrial pollutant emissions (IPE) are all significantly negative at the 5% level. This consistency demonstrates that, even after substituting the AI proxy indicators, the regression results align with the baseline model, confirming that AI development can effectively enhance environmental quality.

4.3.2. Using More Robust Standard Errors

To mitigate the impact of heteroscedasticity and autocorrelation on statistical inference, this paper adopts the more robust “Driscoll–Kraay standard errors” method for re-estimating the baseline model, in line with Hoechle’s research design [49]. The detailed estimation outcomes are provided in columns (4) to (6) of Table 3. The findings reveal that AI consistently exerts a significant impact on energy conservation, carbon reduction, and emissions mitigation, underscoring the robustness of the conclusions drawn in this paper.

4.3.3. Adding Control Variables

If there is an issue of omitted variables, it would lead to biased and inconsistent estimation results of the test variables. To mitigate the potential influence of omitted variable issues affecting the research outcomes, this paper, drawing on the availability of provincial data in China and pertinent research by Liu et al. [5] and Li et al. [25], further controls for regional institutional factors and other provincial cross-sectional characteristics. Specifically, two additional indicators, tax level (TAX) and R&D intensity (RDI), are incorporated into the control variables. The detailed estimation outcomes are presented in columns (7) to (9) of Table 3. The tax level and R&D intensity are represented by the proportion of tax revenue and internal R&D expenditure to regional GDP, respectively. The results suggest that the regression estimates of AI on energy conservation, carbon reduction, and emissions mitigation remain fundamentally consistent in terms of significance, direction, and magnitude after incorporating the aforementioned control variables, reinforcing the robustness of our findings.

5. Mechanism and Heterogeneity Analysis

5.1. Mechanism Analysis

Following the setup of Equations (2) and (3), this section empirically investigates the specific mechanisms through which AI influences energy conservation, carbon reduction, and emissions mitigation. This analysis is conducted from three perspectives: energy structure, green technological innovation, and industrial structure. The regression outcomes are displayed in

Table 4. Results of the mediation effect test of the energy structure.

Variables	(1)	(2)	(3)	(4)
	ENS	ENC	CO2	IPE
AI	−0.001 ***	−0.010	−0.023 **	−0.072 ***
	(−2.934)	(−0.947)	(−2.072)	(−3.187)
ENS		20.667 ***	21.573 ***	19.963 ***
		(16.990)	(16.882)	(7.817)
Control variables	Yes	Yes	Yes	Yes
Constant	−0.115 ***	4.553 ***	5.275 ***	0.054
	(−6.059)	(8.027)	(8.851)	(0.045)
Province fixed effect	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes
N	630.000	630.000	630.000	630.000
Adj.R2	0.293	0.924	0.915	0.373

Note: t statistics in parentheses; ** p < 0.05, *** p < 0.01.

,

Table 5. Mediation Effect Test Results for Green Technological Innovation.

Variables	(1)	(2)	(3)	(4)
	GTI	ENC	CO2	IPE
AI	0.363 ***	0.027 *	0.026 *	0.008
	(15.421)	(1.802)	(1.679)	(0.287)
GTI		−0.163 ***	−0.199 ***	−0.278 ***
		(−7.344)	(−8.704)	(−6.929)
Control variables	Yes	Yes	Yes	Yes
Constant	1.183	2.379 ***	3.039 ***	−1.903
	(0.973)	(3.686)	(4.569)	(−1.630)
Province fixed effect	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes
N	630	630	630	630
Adj.R2	0.797	0.896	0.887	0.760

Note: t statistics in parentheses; * p < 0.1, *** p < 0.01.

, and

Table 6. Mediation Effect Test Results for Industrial Structure.

Variables	(1)	(2)	(3)	(4)
	INS	ENC	CO2	IPE
AI	0.117 ***	−0.007	−0.019	−0.086 ***
	(8.590)	(−0.519)	(−1.314)	(−3.435)
INS		−0.216 ***	−0.239 ***	−0.063 ***
		(−5.545)	(−5.849)	(−4.880)
Control variables	Yes	Yes	Yes	Yes
Constant	5.777 ***	3.436 ***	4.182 ***	−1.866
	(8.201)	(4.945)	(5.758)	(−1.454)
Provincefixed effect	Yes	Yes	Yes	Yes
Timefixed effect	Yes	Yes	Yes	Yes
N	630	630	630	630
Adj.R2	0.790	0.892	0.879	0.307

Note: t statistics in parentheses; *** p < 0.01.

, respectively.

5.1.1. Energy Structure

Table 4 presents the results of the mechanism test focusing on energy structure. As shown in column (1) of Table 4, the coefficient of AI development on regional energy structure (ENS) is strongly negative with a significance level of 1% (coefficient is −0.001, p < 0.01), suggesting that AI development can promote lower electricity consumption and optimize the energy structure across China’s regions. Furthermore, the ENS is incorporated as a control variable in the baseline regression. The estimation outcomes are displayed in columns (2) to (4). The results show that changes in energy structure (ENS), i.e., increases in regional electricity consumption, have significantly positive regression coefficients on energy consumption (ENC), carbon emissions (CO₂), and industrial pollutant emissions (IPE) at the 1% significance level. Thus, based on Equations (2) and (3), it is evident that the mediating effect of AI on optimizing the energy structure is negative. This indicates that AI can attain regional energy savings, carbon reduction, and emissions mitigation through the optimization of the energy structure, thereby enhancing the environmental quality across China’s regions, and H2 is verified.

5.1.2. Green Technological Innovation

Table 5 presents the results of the mechanism test focusing on green technological innovation. As shown in column (1) of Table 5, the coefficient of AI on green technological innovation (GTI) is strongly positive with a significance level of 1%, suggesting that AI can drive China’s “green intelligent technology revolution” and foster green technological innovation. Additionally, GTI is incorporated as a control variable in the regression. The results, displayed in columns (2) to (4) of Table 5, reveal that the regression coefficients of green technological innovation on energy consumption (ENC), carbon emissions (CO₂), and industrial pollutant emissions (IPE) are all strongly negative. This suggests that advancements in green technological innovation can effectively lead to energy conservation, carbon reduction, and emissions mitigation across China’s regions. Therefore, the mediating effect transmitted through green technological innovation by AI is negative, which means that AI can promote high-quality environmental development in China’s regions through the effect of green technological innovation, and H3 is verified.

5.1.3. Industrial Structure

The mechanism test results based on industrial structure upgrading are presented in Table 6. Column (1) indicates that the impact of AI on industrial structure (INS) shows a robust positive coefficient at the 1% significance level, highlighting the positive role of AI in promoting industrial structure upgrading. Columns (2) to (4) reveal that the coefficients for industrial structure upgrading’s impact on energy consumption (ENC), carbon emissions (CO₂), and industrial pollutant emissions (IPE) are all significantly negative (coefficients are −0.216, −0.239, and −0.063, respectively, p < 0.01), indicating that industrial structure upgrading is conducive to improving environmental quality in China’s regions. In summary, the mediating effect transmitted through industrial structure upgrading by AI is negative, which means that AI can help improve regional environmental quality by guiding industrial structure upgrading, and H4 is verified.

5.2. Heterogeneity Analysis

5.2.1. Heterogeneity of Regional Physical Capital

Regions with higher physical capital investment often have stronger economic development momentum, higher levels of internet technology and AI applications, and better infrastructure development, laying a strong material foundation for the upgrading and widespread application of AI. Thus, in regions with a high level of physical capital development, the impact of AI on enhancing environmental quality will be more pronounced. To assess the impact of regional physical capital development levels, this section uses the aggregate fixed assets of each province as an indicator of physical capital. The 30 provinces are categorized into two groups, “high physical capital” and “low physical capital”, according to the median of the provincial sample throughout the period. The grouped estimation results are presented in

Table 7. Regional heterogeneity of physical capital.

Variables	High Physical Capital			Low Physical Capital
	(1)	(2)	(3)	(4)	(5)	(6)
	ENC	CO2	IPE	ENC	CO2	IPE
AI	−0.026 **	−0.035 ***	−0.063 **	−0.046	−0.051 *	−0.059 *
	(−2.417)	(−3.094)	(−2.056)	(−1.576)	(−1.742)	(−1.937)
Control variables	Yes	Yes	Yes	Yes	Yes	Yes
Constant	8.685 ***	9.603 ***	3.933	1.704 *	2.506 **	0.022
	(10.621)	(11.060)	(1.648)	(1.716)	(2.516)	(0.021)
Province fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
N	315	315	315	315	315	315
Adj.R2	0.982	0.982	0.920	0.968	0.970	0.922

Note: t statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01.

. The estimation results in columns (1) to (3) demonstrate that the effects of AI are more pronounced in regions with high physical capital. This indicates that China needs to continue to address regional economic disparities, accelerate the accumulation of physical capital in less economically developed regions, and better adapt to the intelligent era to further improve environmental quality.

5.2.2. Regional Technology Market Heterogeneity

The advancement of regional technology markets not only indicates the maturity of the innovation environment and technology transfer system but also signifies the potential for future economic growth, which is likely to have a significant impact on AI development. The proportion of technology market transactions to GDP across each province is used as a proxy for the maturity of the technology market. The 30 provinces are categorized into “high” and “low” technology market groups based on the median.

Table 8. Regional heterogeneity of the technology market.

Variables	High Technology Market			Low Technology Market
	(1)	(2)	(3)	(4)	(5)	(6)
	ENC	CO2	IPE	ENC	CO2	IPE
AI	−0.014 **	−0.029 ***	−0.022 **	−0.020	−0.022	−0.075 *
	(−2.036)	(−7.415)	(−2.173)	(−1.199)	(−1.305)	(−1.790)
Control variables	Yes	Yes	Yes	Yes	Yes	Yes
Constant	6.820 ***	7.634 ***	3.139 *	0.344	0.523	−3.124
	(7.549)	(8.105)	(1.806)	(0.345)	(0.510)	(−1.616)
Province fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
N	315	315	315	315	315	315
Adj.R2	0.979	0.978	0.884	0.979	0.980	0.899

Note: t statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01.

displays the grouped estimation results.

The findings in columns (1) to (3) suggest that in provinces where the technology market is more advanced, the regression coefficients for AI are significantly negative at the 5% level or higher. By examining columns (4) to (6), it is evident that the regression coefficients for the group with a lower level of technology market development are mostly insignificant or exhibit lower levels of significance. The above research findings suggest that the effects of AI are more pronounced in provinces where the technology market is more advanced. This means that the improvement in the development level of technology markets helps enterprises use modern technologies such as AI and big data to improve outdated production methods, thereby improving environmental quality.

5.2.3. Heterogeneity of Regional Marketization

The 30 provinces in the sample are categorized into two groups, “high marketization” and “low marketization”, according to the median value of the China Marketization Index, and regression is performed. The specific relevant results are detailed in

Table 9. Regional heterogeneity of marketization.

Variables	High Marketization			Low Marketization
	(1)	(2)	(3)	(4)	(5)	(6)
	ENC	CO2	IPE	ENC	CO2	IPE
AI	−0.001 **	−0.009 ***	−0.070 ***	0.266 ***	0.281 ***	−0.238 ***
	(−2.351)	(−6.446)	(−2.998)	(5.071)	(5.268)	(−2.726)
Control variables	Yes	Yes	Yes	Yes	Yes	Yes
Constant	10.923 ***	11.525 ***	6.587 ***	−0.659	−0.174	0.064
	(12.957)	(13.076)	(3.799)	(−0.731)	(−0.189)	(0.043)
Province fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
Time fixed effect	Yes	Yes	Yes	Yes	Yes	Yes
N	315	315	315	315	315	315
Adj.R2	0.981	0.980	0.934	0.977	0.978	0.894

Note: t statistics in parentheses; ** p < 0.05, *** p < 0.01.

.

The findings, which are presented in columns (1) to (3) of Table 9, indicate that in provinces with higher levels of marketization, the regression coefficients for AI are also strongly negative, with a significance level of 5% or greater. By examining columns (4) to (6), it becomes evident that when the marketization level of a province is low, the energy-saving, carbon-reducing, and emission-reducing effects of AI are uncertain. The likely reason is that the level of marketization indicates the extent of development in regional product and factor markets. Regions with relatively higher marketization levels often have more high-tech and knowledge-intensive industries, making it easier for AI applications and development to form scale effects, thereby significantly improving productivity quality. In contrast, provinces with relatively lower marketization levels rely more on traditional industries and labor-intensive jobs, and the demand for AI applications may be relatively small; hence, the impact of AI on energy conservation, carbon reduction, and emissions mitigation remains unclear.

6. Discussion

The connection between economic growth and green development has long been central to scholarly debates on economic models. Achieving harmony between socioeconomic systems and the ecological environment is essential for sustainable global economic development. As a significant indicator of the latest wave of technological advancement, AI technology not only offers innovative frameworks for nations to achieve green governance but also strengthens the ability of enterprises to manage pollutant emissions in their production and operational processes. Consequently, AI is emerging as a new catalyst for enhancing environmental quality. In this context, the question of whether and how AI can facilitate improvements in environmental quality has become a critical and urgent issue.

In light of this, this paper delves into the effects, underlying mechanisms, and heterogeneous characteristics of AI on environmental quality within the framework of China’s energy conservation and carbon reduction initiatives. Examining these aspects aids in furthering the digital and intelligent evolution of China’s economy and supports the advancement of high-quality environmental transformation. The findings indicate that the development and deployment of AI can effectively contribute to the enhancement of environmental quality in China by optimizing the energy structure, fostering green technological innovation, and upgrading industrial structures. Furthermore, the positive impact of AI on environmental quality demonstrates “path dependence” and significant variability across regions, influenced by factors such as physical capital, technological market development, and levels of marketization. These empirical results are comprehensively analyzed in the “Theoretical Analysis and Hypothesis” section of this paper, with particular attention to both direct effects and the three intermediary mechanisms: energy structure optimization, green technology innovation, and industrial structure upgrading.

Clearly, the research perspective, findings, and potential policy implications of this paper differ from existing literature, such as Zhao et al. [50], Liu et al. [5], Brevini [11], and Hao and Wu [16]. Specifically, this paper first integrates the perspectives of energy conservation, carbon reduction, and emissions reduction to comprehensively investigate the extent to which AI can effectively enhance environmental quality, thereby extending and enriching the existing body of related research. Second, it develops a theoretical framework to analyze the relationship between AI development and environmental quality and empirically examines the mechanisms through which AI influences environmental quality via three key channels: energy structure optimization, green technology innovation, and industrial structure transformation. The findings align closely with China’s ongoing economic transformation and industrial upgrading, providing valuable insights into related issues in existing research. Third, to derive more targeted policy implications for promoting AI and enhancing environmental quality, this paper thoroughly investigates the heterogeneity of regional endowment factors, including physical capital, technological markets, and levels of marketization. This analysis undeniably contributes a significant supplement to the current body of research.

Naturally, this study has certain limitations, notably the absence of a corresponding mathematical model to underpin the construction of the empirical model, which remains an area for future exploration. Nonetheless, the theoretical framework presented in this paper, especially the examination of how AI can enhance environmental quality through the three channels of energy structure, optimization, green technology innovation, and industrial structure upgrading, provides direct approaches and valuable insights for the development and refinement of theoretical research on high-quality environmental progress.

7. Conclusions and Policy Recommendations

7.1. Conclusions

From the perspectives of energy conservation, carbon reduction, and emission reduction, this paper systematically examines a range of practical questions concerning the extent to which AI enhances environmental quality in China. Utilizing panel data from 30 provinces, municipalities, and autonomous regions in China over the period from 2000 to 2020, the study employs two-way fixed effects models and mediation effect models to conduct the analysis.

The findings demonstrate that the development and implementation of AI can substantially lower China’s energy consumption, carbon emissions, and industrial pollutant discharge, thereby enhancing environmental quality. This conclusion holds true even after conducting multiple robustness tests.

Mechanism tests reveal that the adjustment of energy structure, the innovation in green technology, and the upgrading of the industrial structure are the specific pathways through which AI exerts its influence on energy conservation, carbon reduction, and emissions reduction, thereby contributing to the effective improvement of China’s environmental quality. Heterogeneity analysis reveals a significant “path dependence” effect in the development of AI in China. In regions with higher physical capital investment, better-developed technology markets, and higher levels of marketization, the effects of AI on energy conservation, carbon reduction, and emission reduction are more pronounced, indicating a more significant improvement in environmental quality due to AI.

7.2. Policy Recommendations

Based on the above research findings, the following policy recommendations can be derived.

Firstly, all regions should elevate the development and implementation of AI to maximize its potential for emissions reduction within the framework of an increasingly intelligent economy. Local governments should take advantage of the opportunities generated by the development of economic intelligence. On the one hand, they can increase financial support for cutting-edge AI research, build talent training systems based on the demand for intelligent skills, and improve the construction of intelligent energy internet infrastructure. On the other hand, they can develop policy frameworks that encourage the widespread integration of intelligent technologies, foster their application in achieving industry-wide technological advancements, and offer technical support for energy conservation, carbon reduction, and emissions reduction across all regions.

Secondly, all regions should expedite the development of intelligent green technologies by enterprises and systematically advance the transformation and enhancement of intelligent industrial frameworks. On the one hand, local governments can offer policy support for green-focused technological innovation by encouraging intelligent enterprises to engage in green technology research through tax incentives and R&D subsidies, thereby facilitating the rapid implementation of green technological advancements. On the other hand, the government can steer industrial upgrades with an emphasis on energy conservation, carbon reduction, and emissions reduction by promoting the synergy and advancement of intelligent technologies within traditional industries. This approach would facilitate the widespread adoption of intelligent technologies across sectors, boost the share of clean industries within the industrial framework, and gradually establish a new industrial paradigm focused on energy conservation, carbon reduction, and emissions reduction.

Thirdly, all regions should work to minimize regional capital disparities while advancing the growth of technology markets and elevating the degree of marketization. The heterogeneity analysis indicates that physical capital, the growth of technology markets, and the degree of marketization are essential foundations for AI to enhance environmental quality in Chinese regions. To fully and thoroughly advance high-quality environmental development across China, it is essential to expedite the equalization of physical capital investment, technology market growth, and marketization levels across various regions. For example, actively improving digital infrastructure construction, fully leveraging the diffusion and spillover effects of physical capital and technology, promoting the formation of a comprehensive AI application support system, and reducing the risks of digital and intelligent transformation for enterprises.