Artificial Intelligence Technology Applications and Energy Utilization Efficiency: Empirical Evidence from China

Abstract

Although artificial intelligence (AI) serves as a core driver of the new round of technological transformation, its crucial role in improving energy utilization efficiency has not yet received sufficient attention. This analysis empirically explores how the application of AI technology influences energy utilization efficiency using panel data from Chinese cities over the period from 2008 to 2021. The following are the primary conclusions: (1) AI technology applications are able to enhance energy utilization efficiency, and the outcomes remain valid after extensive reliability tests have been conducted; (2) the investigation of the mechanism demonstrates that AI technology applications can optimize energy utilization efficiency through technological and scale effects; (3) environmental regulation and digital infrastructure serve as positive moderators of the impact of AI technology applications on energy utilization efficiency; and (4) a heterogeneity analysis shows that the positive impact of AI technology applications on energy utilization efficiency is more significant within resource-dependent cities, cities with non-traditional industrial foundations, and those with a strong emphasis on environmental protection. The application of AI technology significantly enhances energy efficiency, which is a finding that remains robust across multiple reliability tests.

1. Introduction

As global energy consumption continues to rise, traditional primary energy sources can no longer meet the increasing energy demands of people, and the use of these sources has further exacerbated environmental and climate problems [1]. Previous studies have demonstrated that enhancing energy efficiency is instrumental in alleviating environmental pollution and climate change [2]. Based on data from the International Energy Agency, a 1% increase in energy utilization efficiency can save billions of dollars in energy expenditure and significantly reduce carbon dioxide emissions. The “World Energy Outlook 2023” report by the International Energy Agency highlights the vulnerability and instability of the international energy framework, emphasizing that controlling the greenhouse effect relies on a transition to clean energy, while the overall demand for renewable energy surpasses that for conventional fossil fuels. Previous studies have indicated that the swift development of China’s economy has come at the cost of environmental degradation [3,4]. In addition to the widespread consumption of fossil fuels, low energy utilization efficiency is a key factor. Therefore, improving energy utilization efficiency is fundamental to promoting sustainable economic growth in China, while also having far-reaching implications for global sustainability.

Marked by the widespread application of advanced digital technologies, including AI, cloud computing, and big data, the Fourth Industrial Revolution is sweeping across the globe in an unprecedented manner. As a primary driver of this revolution, AI technology is leading profound transformations in global science, technology, and industry. The application of AI technology currently shows immense potential and opportunities, utilizing numerous advanced technologies, including blockchain and machine learning, and providing more accessible platforms for resource sharing, knowledge spillover, and efficiency optimization. China has already made some progress in its digital construction and will promote the swift growth of digital cities, fostering the convergence of digitalization and intelligence across various sectors, thereby offering robust support for the nation’s economic transformation. Therefore, China places great importance on the research and market application of AI technology and has made certain advancements in digital construction. In the future, China will further accelerate its digital development and promote the integration of AI with other fields, with the question of how to utilize AI technology to improve energy utilization efficiency being a crucial aspect. What are the effects of AI technology applications on energy utilization efficiency, as well as the core mechanisms and factors that shape this connection? Responses to this question are crucial for efficiently achieving AI-driven high-quality energy development.

Existing studies have examined the connection between these two factors, analyzing it from various viewpoints. First, the influence of AI technology on energy efficiency has been studied. At the micro level, Ngarambe et al. have found that the use of AI technology within the construction field is poised to greatly boost the energy efficiency of structures [5]. He et al. argued that AI development can promote improvements in technological levels by increasing funding for research and talent, thereby effectively enhancing energy efficiency [6]. At a macro level, Sun et al. [7] discovered that AI technology is capable of improving energy utilization efficiency by increasing knowledge and creativity spillover, while Lee et al. [8] suggested that the integration of AI technology with the Internet in the financial sector can indirectly enhance regional energy efficiency levels. Second, the influence of digital technologies and smart systems on energy performance has been explored. On the one hand, from an economic and technological perspective, Borowski found that applying digital technology in industrial production can effectively improve both production and energy efficiency [9], while Lange and Pohl demonstrated that digitalization can enhance energy utilization efficiency to some extent by promoting economic expansion [10]. From the perspective of information symmetry, Chui et al. [11] confirmed that the interconnected energy information and reduced transaction costs provided by intelligence can significantly improve energy efficiency for enterprises. Brynjolfsson and Hitt [12] noted that intelligent management practices provide precise management decisions and reduce labor costs across various processes, resulting in efficient energy utilization for production. Third, the influence of AI technology on energy consumption has been examined. From a negative standpoint, Grant et al. [13] revealed that the application of AI in industrial production can save labor, leading to a boost in energy consumption. Brevini [14] argued that the intelligent transformation of industry could result in high energy consumption and environmental pollution issues for coal-dependent countries. Conversely, from a positive perspective, Forslid et al. [15] showed that the implementation of AI technology in industrial production can yield high productivity and standardized production processes, leading to reduced energy and resource consumption, a finding that is consistent with the findings of Vivanco et al. [16].

Based on the literature reviewed above, it is observed that although an increasing number of researchers have turned their attention to the connection between AI and energy efficiency, there remain two main shortcomings in the analysis of their causal relationship. First, given the deep theoretical connection between the advancement of AI technology and technological innovation, existing studies mostly examine the impact pathway of AI applications on energy efficiency with regard to technological innovation [5,6,7]. This relatively one-dimensional research perspective limits the exploration of the causal mechanisms between the two factors, with insufficient theoretical depth and breadth. Second, in the causal identification regarding AI technology applications and energy efficiency, the current research remains relatively limited in addressing endogeneity. Most scholars primarily use instrumental variable methods combined with two-stage least squares for testing [17,18]. Although this approach alleviates endogeneity bias to some extent, it still faces limitations in verifying the validity of instrumental variables and addressing potential bidirectional causality. Therefore, more diverse and rigorous identification methods need to be introduced in handling endogeneity in future research.

In response, the present study utilizes panel data from Chinese cities covering the period from 2008 to 2021 to measure the level of AI technology applications and energy utilization efficiency indices in these cities. This study investigates the relationship between the two factors, exploring the mechanisms of technological and scale effects, while also introducing environmental regulation and digital infrastructure to analyze their moderating effects.

This research adds to the existing literature on the application of AI technology and energy utilization efficiency in two distinct dimensions. First, building on existing research, it not only provides an in-depth exploration of the technological mechanism but also innovatively incorporates scale effects. It offers a more comprehensive and detailed examination of the mechanisms between AI technology applications and energy utilization efficiency. Second, to address the endogeneity issue in the relationship between AI technology applications and energy utilization efficiency, this study innovatively introduces the “National AI Innovation and Application Pioneer Zone” policy as a policy shock variable for AI technology applications. A difference-in-differences (DID) model is applied to empirical analysis, which further resolves endogeneity concerns and enhances the reliability of the empirical results.

2. Research Hypotheses

Drawing on the research background described above, this study presents five research hypotheses, focusing on the causal relationship between the application of AI technology and energy utilization efficiency, the internal mechanisms of technological and scale effects, and the moderating roles related to environmental regulation and digital infrastructure.

2.1. AI Technology Applications and Energy Utilization Efficiency

The collaborative process of AI and energy utilization efficiency is a process that encompasses both holistic and global dimensions, and its impact on energy can be examined from three key perspectives. First, from the viewpoint of smart manufacturing, the incorporation of AI technology into industrial enterprises primarily revolves around the application of industrial robots. Industrial robots are defined as intelligent machines characterized by automation, repetitiveness, and versatility [19]. These robots can significantly replace human labor in the production process and, due to their inherent precision and timeliness, effectively reduce material wastage and improve the productivity of raw materials [20]. Additionally, industrial robots can observe and assess energy consumption and pollution levels immediately, aiding in timely judgment and decision-making. Second, with regard to production chain optimization, the innovative characteristics of AI can diffusely influence various stages of a business, thereby enhancing the technological innovation of the production chain system and optimizing the allocation and collaborative configuration of resources within the system [21]. Furthermore, at all stages of the production chain—upstream, midstream, and downstream—AI technology enables operations to become more efficient while reducing consumption, thereby contributing to an improvement in energy utilization efficiency. Finally, from the overall perspective of economic operation, under the dual effects of intelligent production and intelligent management, the economy relies on intelligent, rapid, and precise information processing methods, which provide a more convenient means of information exchange for AI technology [22]. As operational carriers become more diverse, resource allocation becomes more rationalized, and related industries become more refined, the convergence of AI technology and energy utilization can become more harmonious and profound. This, in turn, can drive enterprise transformation, promote technological innovation, and ultimately enhance overall energy utilization efficiency. Therefore, this paper posits that AI technology applications are likely to enhance energy utilization efficiency.

Hypothesis 1. 

AI technology applications will improve energy utilization efficiency.

2.2. The Mediating Role of Technological Effects

In terms of the technological effects mechanism that links AI technology applications with energy utilization efficiency, the endogenous growth theory, represented by Romer, was the first to open the “black box” of technological progress and successfully integrate technological innovation into macroeconomic analysis. Existing research mainly divides endogenous growth theory into capital-driven and innovation-driven types. The innovation-driven endogenous growth theory indicates that endogenous technological progress is crucial to economic expansion and long-term progress. As a general-purpose technology that drives current economic and social development, AI inherently involves technological innovation that falls under the category of endogenous technology. The rapid development of AI can effectively stimulate both urban technological innovation and the advancement of AI itself. This endogenous technological driving mechanism of AI can, in turn, enhance both energy utilization efficiency and production efficiency. As Acemoglu and Restrepo have pointed out, AI technology applications and the accelerated development of technological innovation can complement each other, and their deep integration can reduce energy consumption and enhance energy productivity [18]. Therefore, this section focuses on analyzing the impact of technological effects, reflected through technological innovation, and the association between AI technology applications and energy utilization efficiency. First, technological innovation is a crucial driver for enhancing energy utilization efficiency. Song et al. [23] found, through data from several green innovation countries, that technological advancement is vital for improving energy efficiency and reducing consumption during energy production processes. Improvements in technological innovation can attract high-level technical talent and more intelligent enterprises [24], providing more opportunities for technical exchange and progress among urban enterprises, as highlighted by Brynjolfsson and Hitt [12]. High-level technical talent is more likely to seek employment in cities with advanced levels of technological innovation, which can furnish the city’s AI development with more hardware and software, aiding relevant departments in developing more efficient resource allocation methods, making more rational decisions, and boosting energy efficiency. Moreover, technological effects exert considerable spatial externalities on the energy efficiency of adjacent cities [25,26]. Kang et al. and Fernández et al. have analyzed the technological spillover effects at both the urban and enterprise levels within China [27,28], indicating that regional factors contribute to the varying effects of technological innovation on energy efficiency. Enterprises characterized by advanced technological innovation tend to develop first and generate technological spillover effects, thus driving the development of technological levels in surrounding cities and enterprises and collectively enhancing energy efficiency. Furthermore, scholars such as Coe and Helpman have demonstrated that a country can obtain positive technological spillover effects through trade [29], effectively influencing energy productivity. Secondly, AI technology has significant technological spillover effects. As one of the most promising and impactful technologies today [30], its development has profound and substantial implications for technological advancement [31]. Researchers like Tian et al. have confirmed from a micro-level perspective that AI has a considerable and beneficial impact on technological innovation [30]. Luo et al.’s extensive research has been carried out at a macro level, utilizing province-level data collected from China to examine AI technology’s impact on driving technological advancement [32]. Previous research has demonstrated that AI technology, as a rising advanced science and technology, has immense potential to reshape innovation processes and the essence of R&D organizations [33], providing efficient information collection and processing capabilities [34,35] and improving R&D efficiency [36], thereby driving technological development. Additionally, research by Lyytinen et al. has found that improvements in AI technology can potentially positively impact the progression of green technologies [37], while Varian et al. have confirmed that this influence can generate positive technological spillover effects [38], driving the development of green technological innovation levels in surrounding areas. Finally, given the existing literature, it is evident that AI technology can enhance technological innovation levels through various means. Furthermore, enhancing the extent of green technological innovation can substantially boost the efficiency of energy utilization in urban contexts. Therefore, this study posits that there may be a technological effects mechanism influencing the association between AI technology applications and energy utilization efficiency.

Hypothesis 2. 

AI technology applications will improve energy utilization efficiency through the advancement of green technological innovation.

2.3. The Mediating Role of Scale Effects

From the perspective of the scale effects mechanism connecting AI technology applications to energy utilization efficiency, according to the capital-driven endogenous growth theory, the externalities and spillover effects of knowledge capital and human capital provide a solid knowledge foundation for research and production. The synergy between the two facilitates increasing returns to scale, thereby becoming an important driving force for sustained economic growth. Holding other conditions constant, human capital tends to increase with population growth within an economy, which in turn raises the level of knowledge stock and accelerates economic growth, demonstrating a scale effect. AI technology is essentially a knowledge-intensive technology, and its development heavily relies on the support of highly skilled human capital. As AI technology becomes more widely applied, the accompanying accumulation of knowledge and capital will bring significant economic benefits to cities and effectively drive the continuous expansion of economic scale. This expansion provides more funding for cities and enterprises to purchase high-efficiency, low-loss machinery and equipment, thereby enhancing energy utilization efficiency, which reflects the mechanism of the “scale effects.” As Verhoef and Nijkamp found [39], economies of scale can significantly improve urban energy utilization efficiency at certain levels. Therefore, this section aims to explore the mediating role of the scale effect brought about by economic agglomeration. First, economic agglomeration is an important driver of enhanced energy utilization efficiency. On one hand, the increase in economic scale and agglomeration can attract high-level technical talent and more intelligent enterprises to urban areas, providing more opportunities and channels for technical exchange and progress among urban enterprises [40]. Han et al. argued that the development of economic agglomeration can [41], to some extent, promote technological advancement, while Zhao and Lin contended that the formation of economies of scale can lead to improvements in energy efficiency [42], providing a solid foundation for boosting urban energy efficiency levels. On the other hand, Glaeser’s research indicates that cities characterized by greater economic agglomeration tend to have greener transportation and living conditions [43]. Additionally, Glaeser and Kahn’s study shows that increasing urban economies of scale can effectively reduce energy consumption [44]; greener transportation, eco-friendly housing, and decreased energy consumption all contribute to enhanced energy utilization efficiency. Furthermore, Li and Lin found that differences in development stages among countries may lead to differentiated impacts between urban economies of scale and energy efficiency [45], with the effects of economies of scale on energy efficiency being more pronounced in low-income countries. Secondly, advancements in AI technology can greatly expedite the growth of urban economies of scale. Acemoglu and Restrepo suggest that AI technology applications serve as a strong stimulus for economic growth [46], and that economic growth within a limited geographic area can boost economic clustering. Fan and Scott found that advancements in AI technology have a certain attraction for related intelligent enterprises [47], effectively promoting the simultaneous agglomeration of industries and economies within cities, thereby significantly fostering the development of scale economies and scale industries. Additionally, as an emerging information technology, AI can refine the distribution of market factors and the productivity of economic elements, thereby positively impacting the development and growth of scale economies in China [48]. Finally, existing research suggests that the advancement of AI technology can contribute to the expansion of economic agglomeration, and economic agglomeration can enhance urban energy utilization efficiency in various ways. This study posits that there may be a scale effect mechanism influencing the connection involving AI technology applications and the enhancement of energy utilization efficiency.

Hypothesis 3. 

AI technology applications will improve energy utilization efficiency through the advancement of economic agglomeration.

2.4. The Moderating Role of Environmental Regulations

Existing research has shown that the enforcement of environmental regulations not only encourages organizations to pay more attention to pollutant treatment but also prompts them to adjust their usage of AI accordingly. At the same time, Wang et al. identified that the integration of the digital economy with AI technology can effectively enhance energy utilization efficiency [49]. Therefore, this paper examines the moderating effects of environmental regulation and digital infrastructure on the relationship between these two factors.

The over-exploitation of resources and excessive energy consumption have led to rapid industrial development in China, but they have also resulted in severe environmental pollution and climate deterioration. In this context, the government must urgently enforce mandatory environmental regulations to control the management and consumption of energy. Existing research has shown that environmental regulation is vital to both the development of AI and the enhancement of energy utilization efficiency [50]. First, according to the “Porter Hypothesis” proposed by the revised school, the implementation of environmental regulations, while increasing firms’ costs associated with controlling and treating pollutant emissions, also serves to stimulate technological innovation. This innovation, in turn, helps firms improve output levels, thereby offsetting the negative effects of increased costs. As a result, environmental regulations can indirectly encourage firms to reallocate resources toward the exploration and development of advanced AI technologies [51]. Although such regulations increase capital expenditures for enterprises, the labor substitution effect offered by AI helps alleviate financial pressure while enhancing production efficiency and resource allocation capabilities [52]. In summary, moderate environmental regulations can not only incentivize firms to invest in AI research and development but also facilitate advancements in energy efficiency [53]. Second, the concept of natural selection is inherently present in the competitive environment among businesses. Some companies may struggle to adapt to environmental regulations and cease operations, while those that remain often possess advanced equipment and technology, allowing them to conduct production activities and pollution control more efficiently. Furthermore, they can attract more funding to recruit high-tech talent for AI research and effective management, thereby creating a positive cycle that enhances energy utilization efficiency [54]. Consequently, the enforcement of environmental regulation has the potential to stimulate traditional companies to increase the application of AI technology, thereby enhancing their innovation and development capabilities and ultimately improving energy utilization efficiency.

Hypothesis 4. 

Environmental regulation influences the effect of AI technology applications in improving energy utilization efficiency.

2.5. The Moderating Role of Digital Infrastructure

The development of digital infrastructure not only optimizes the flow and allocation of information resources but also drives the expansion of the digital economy and technological capabilities, thereby supporting regional economic development. The advancement of data as a production factor relies on the establishment and enhancement of digital infrastructure. As Nambisan emphasizes, digital infrastructure is fundamental to driving the expansion of the digital economy and digital technologies [55]. On the one hand, Linkov et al. discovered that digital infrastructure can greatly enhance connectivity and intelligence in communication and production [56], with intelligent production becoming a representative technological advancement of the digital age. A robust digital infrastructure, rich datasets, and cutting-edge technological innovations provide essential support for the promotion of AI technology [57]. Furthermore, the advancement of digital infrastructure effectively increases the usability of datasets, which can offer powerful backing for the learning and training of AI [58]. Digital infrastructure plays an important supporting and facilitating role in the development of AI technology. On the other hand, previous research has demonstrated that digital infrastructure serves as a key driver of promoting innovation factors [59], driving economic transformation and upgrades [60], improving enterprise productivity, optimizing resource allocation [61], and enhancing total factor productivity [62]. These factors have become important driving forces for the application of AI technology to improve energy utilization efficiency. The enhancement of digital infrastructure not only accelerates the development of AI technology but also provides sufficient external support conditions for improving energy utilization efficiency. Therefore, this research proposes that the improvement of digital infrastructure effectively promotes the connection between AI technology applications and energy utilization efficiency optimization.

Hypothesis 5. 

Digital infrastructure influences the effect of AI technology applications in improving energy utilization efficiency.

3. Research Methodology

3.1. Benchmark Regression Model

To assess the validity of the previously mentioned theoretical hypotheses, this research develops the baseline regression model outlined below:

$${\mathrm{E}\mathrm{U}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1{\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\alpha }}_2{\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\phi }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(1)

In Model (1), i and t indicate city and time, respectively. EU_it indicates the energy utilization efficiency within city i during time t; AI_it reflects the degree of AI technology applications within city i during time t; and X_it denotes the control variables, which include the degree of financial development (Finance_it), fiscal investment intensity (Govern_it), degree of fiscal decentralization (Finadp_it), industrial structure (Ind_it), market scale (Market_it), and overall upgrading of the industrial structure (Structure_it). ${\mathsf{\alpha }}_0$ is the intercept, ${\mathsf{\alpha }}_1$ and ${\mathsf{\alpha }}_2$ represent the estimated coefficients of the model, ${\mathsf{\mu }}_{\mathrm{i}}$ indicates the geographic fixed effects, ${\mathsf{\phi }}_{\mathrm{t}}$ represents the time fixed effects, and ${\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$ denotes the disturbance term.

3.2. Variable Selection

3.2.1. Explained Variable

This study measures energy utilization efficiency using the total green factor energy efficiency. Prior research has commonly used the DEA model to measure green factor energy efficiency; however, traditional DEA methods have certain limitations in handling undesirable outputs and multi-period data, which could result in distorted measurement outcomes. As demonstrated in the research conducted by Shi et al. [63], this study uses the SBM–Malmquist–Luenberger (SBM-ML) index to evaluate the green total factor energy efficiency for each city. This method integrates both positive and negative outputs within the model. Specifically, labor resources, capital quantity, and energy consumption, along with other relevant factors, are chosen as the input for the analysis. Due to the missing data on energy consumption for various cities, this study follows common practices by choosing liquefied petroleum gas, natural gas, and electricity to estimate energy usage. The release of industrial wastewater, sulfur dioxide emissions, and industrial particulate matter are chosen as undesirable outputs; meanwhile, the regional GDP is designated as the positive output.

3.2.2. Explanatory Variable

Currently, AI technology is primarily applied within the production processes of the manufacturing industry. Thus, this study builds on the approach of Acemoglu and Restrepo by utilizing the density of industrial robot installations in cities as a gauge for AI technology applications [18], specifically the number of industrial robots per 10,000 people. The density of industrial robot installations reflects the level of AI technology applications in that city. Since the International Federation of Robotics (IFR) exclusively releases industrial robot installation data at the country and sectoral levels, this study adopts the common practice from the existing literature and employs the Bartik instrumental variable approach [18]. The urban industrial robot installation density is illustrated below:

$${\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}={\sum }_{\mathrm{j}=1}^{\mathrm{j}}{\displaystyle \frac{{\mathrm{e}\mathrm{m}\mathrm{p}}_{\mathrm{i}\mathrm{j}}^{\mathrm{t}}}{{\mathrm{e}\mathrm{m}\mathrm{p}}_{\mathrm{i}}^{\mathrm{t}}}}·{\displaystyle \frac{{\mathrm{r}\mathrm{o}\mathrm{b}\mathrm{o}\mathrm{t}}_{\mathrm{j}}^{\mathrm{t}}}{{\mathrm{e}\mathrm{m}\mathrm{p}}_{\mathrm{j}}^{\mathrm{t}}}}$$

where ${\mathrm{e}\mathrm{m}\mathrm{p}}_{\mathrm{i}\mathrm{j}}^{\mathrm{t}}$ represents employment in industry j within city i at time t; ${\mathrm{e}\mathrm{m}\mathrm{p}}_{\mathrm{i}}^{\mathrm{t}}$ indicates the total workforce in that industry within city i at time t, the ratio of the two represents the percentage of workers in industry j within city i at time t; ${robot}_j^t$ reflects the total count of industrial robots in industry j in China at time t; ${emp}_j^t$ represents the total number of employees in industry j in China at time t; and the ratio of these two data points represents the urban industrial robot installation density across the country at time t.

3.2.3. Control Variables

To minimize the interference of other external variables, this paper follows the approach of Acemoglu and Restrepo [18] by selecting the following six control variables: financial development level (Finance), fiscal decentralization (Finadp), government intervention (Govern), industrial structure (Ind), advanced industrial structure (Structure), and market size (Market). The details are as follows:

(1) Finance: An advanced state of financial development can effectively drive local financial growth and provide greater financial support to advance AI technologies and related innovations, thereby positively impacting energy utilization efficiency. In this study, Finance is defined as the proportion of the year-end balances of loans and deposits held by financial institutions to the regional GDP.

(2) Finadp: The intensity of fiscal decentralization can influence the behavior of local governments. A high level of fiscal decentralization may improve the efficiency of local resource allocation but may also increase energy consumption due to distorted incentives. Therefore, fiscal decentralization may have contrasting effects on energy efficiency. In this paper, Finadp is quantified by the proportion of government revenue to fiscal expenditure.

(3) Govern: Appropriate government intervention can support technological research and development through policy guidance and by alleviating financial constraints. However, both excessive and insufficient intervention can hinder industrial development and technological innovation. As such, government intervention may result in contrasting effects on energy utilization efficiency. Govern is quantified by the proportion of fixed asset investment to total government fiscal expenditure.

(4) Ind: The industrial structure reflects the demand structure for energy in socio-economic activities. A rational industrial structure can optimize the allocation of production factors, thereby improving energy utilization efficiency. This paper measures Ind by the proportion of the value output of the tertiary sector relative to regional GDP.

(5) Structure: The upgrading of industrial structure is often accompanied by improvements in technological level and production efficiency, indicating a shift toward greener and more modern industrial development. This process contributes to enhanced energy utilization efficiency. In this study, Structure is estimated using the following formula: (added value of the primary industry as a percentage of GDP × 1) + (added value of the secondary industry as a percentage of GDP × 2) + (added value of the tertiary industry as a percentage of GDP × 3). This method more comprehensively and accurately captures the trend of industrial transformation from low- to high-value added activities.

(6) Market: Market size influences product demand and firms’ technological capabilities. It also plays a crucial role in promoting industrial upgrading. The expansion of market size is expected to positively affect energy utilization efficiency. In the present work, the market is assessed through the log of total retail sales of consumer goods.

3.3. Sources of Data and Preliminary Analysis

Due to missing data for some cities, they were not included in this empirical research sample. The original data regarding industrial robots comes from the IFR. Other sources of original data include the China City Statistical Yearbook (2008–2021) and the China Energy Statistical Yearbook (2008–2021), as well as municipal statistical yearbooks and various official reports. Missing values were imputed using linear interpolation.

Table 1. Descriptive statistics.

Variable	Obs	Mean	Std. dev	Min	Max
EU	3374	0.3365	0.1315	0.1026	1.1770
AI	3374	4.4864	4.3937	0.0126	25.9109
Finance	3374	2.4449	1.2390	0.5879	21.3018
Finadp	3374	0.4839	0.2241	0.0544	1.5413
Govern	3374	4.9704	2.1195	0.0526	17.1682
Ind	3374	0.4133	0.1028	0.1180	0.8387
Structure	3374	2.2964	0.1451	1.8312	2.8357
Market	3374	15.5843	1.0621	12.1008	19.0129
Tecn	3374	512.4795	1452.051	0.0000	22275
EA	3374	0.3472	0.8070	0.0013	15.3555
ERI	3374	80.2336	22.4102	0.4900	156.4500
DI	3374	0.1097	0.0997	0.0014	0.8647

displays the summary statistics for the variables.

4. Results

4.1. Benchmark Regression Tests

The regression results, according to the equation in Model (1), are presented in

Table 2. Baseline regression results.

Variable	(1)	(2)	(3)
	EU	EU	EU
AI	0.0032 ***	0.0017 ***	0.0036 ***
	(0.0007)	(0.0005)	(0.0007)
Finance		−0.0257 ***	0.0058 **
		(0.0048)	(0.0026)
Finadp		0.1023 ***	0.0170
		(0.0181)	(0.0216)
Govern		−0.0064 ***	−0.0027 *
		(0.0012)	(0.0015)
Ind		0.5860 ***	−0.4946 ***
		(0.0803)	(0.0823)
Structure		−0.3420 ***	0.4605 ***
		(0.0522)	(0.0865)
Market		0.0460 ***	−0.0264 ***
		(0.0041)	(0.0099)
Constant	0.3223 ***	0.2003 *	−0.1299
	(0.0034)	(0.1164)	(0.2063)
City fixed	Yes	No	Yes
Year fixed	Yes	No	Yes
Observations	3374	3374	3374
R-squared	0.6479	0.2046	0.6525

Note: * p < 0.1, ** p < 0.05, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

. Columns (1) to (3) present the regression estimates of AI technology applications on energy utilization efficiency. Column (1) excludes control variables, whereas Columns (1) and (3) account for both time fixed effects and regional fixed effects, whereas Column (2) does not apply fixed effects for time and region. Column (1) shows that the estimated coefficient for AI appears to be positive, suggesting that in the absence of control variables, AI technology applications are essential for enhancing energy utilization efficiency. After adding control variables, the coefficient for AI remains positive, passing the 1% significance threshold. Based on this finding, Hypothesis 1 is proven.

4.2. Endogeneity Test

4.2.1. Instrumental Variable Method

In order to relieve potential bias in results caused by measurement errors and sample selection bias, this study applies the instrumental variable technique approach through two-stage least squares (IV-2SLS) for empirical analysis. Drawing on the study by Kolko [64], this study uses geographic variables such as terrain ruggedness as an exogenous instrument. Drawing on the approach proposed by Duflo and Pande [65], we construct an interaction factor that involves both terrain ruggedness and time as instrumental variables for AI technology applications. Terrain ruggedness, which is constructed based on the range of elevation, total area, and flat land area within a region, is a fundamental geographic feature of a city and possesses inherent exogeneity. Areas with more complex terrain typically face greater challenges in building information infrastructure and expanding internet access, thereby indirectly hindering the development and application of AI technology. However, as time progresses and technology continues to advance, the constraints imposed by complex terrain are expected to gradually diminish. Therefore, the interaction term between terrain ruggedness and time can effectively capture the evolving trend of AI technology applications levels, indicating a certain degree of correlation between the two. Moreover, terrain ruggedness as a natural geographic variable remains stable over time and is not influenced by endogenous factors such as local policy interventions, industrial structure, or innovation environment—factors that are typically regarded as sources of a model’s error term. As such, terrain ruggedness does not exert a direct impact on energy utilization efficiency. Its interaction with time affects energy utilization efficiency only indirectly through AI technology applications, thus satisfying the exogeneity requirement for a valid instrumental variable.

Table 3. Two-stage regression results using instrumental variables.

Variable	(1)	(2)
	AI	EU
Terrain ruggedness × Year	−0.1315 ***
	(0.0129)
AI		0.0223 ***
		(0.0038)
Constant	524.2929 ***	−0.4564 **
	(49.9756)	(0.1992)
The first-stage F statistic	104.6770 ***
Kleibergen–Paap rk LM		110.5960 ***
Cragg–Donald Wald F		124.6260
Controls	Yes	Yes
City fixed	Yes	Yes
Year fixed	Yes	Yes
R-squared	0.8211	0.6125

Note: ** p < 0.05, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

presents the findings. Using the interaction term between terrain ruggedness and time as the instrumental variable, the first-stage regression results are shown in Column (1). A significant negative relationship is found between the instrumental variable and the application of AI technology, indicating that regions with higher terrain ruggedness tend to impede the adoption of AI, which is consistent with the expectations of this study. Regarding the validity of the instrumental variable, the F statistic of the first-stage regression is 104.677, which is well above the commonly used threshold of 10 for detecting weak instruments, indicating a significant relationship between the instrumental variable and the endogenous regressor. This indicates that the weak instrument problem is not present. The second-stage regression results, as shown in Column (2), reveal that even after accounting for endogeneity, the beneficial impact induced by the AI technology applications in energy utilization efficiency remains statistically significant at the 1% level. Furthermore, the Kleibergen–Paap rk LM statistic is 110.596, which passes the 1% significance level test, confirming that there are no concerns regarding the under-identification of the instrument. The C-D Wald F statistic is 124.626, which far exceeds the Stock–Yogo critical value of 16.38 at the 10% maximal IV size, further confirming the absence of weak instrument problems. These results demonstrate that the interaction term between terrain ruggedness and time is a valid instrumental variable.

To address concerns regarding the exclusion restriction of the instrumental variable, this study combines textual discussion with graphical analysis. On one hand, the existing literature at both the firm and city levels has shown that terrain ruggedness—on which the constructed instrument is based—is determined by natural geographic factors [66,67,68,69], making it an exogenous variable unrelated to the economic system. On the other hand, following the method proposed by Chen et al., a scatter diagram is drawn to examine the association between the instrumental variable and the residuals of the explained variable [70]. Given the large original sample size, this study adopts a random sampling approach, selecting 10% of the observations for visualization. As shown in Figure 1, there is no clear correlation between the two factors, supporting the exogeneity of the instrument.

4.2.2. DID Method

To avoid biases caused by bidirectional causality and omitted variable issues, this study introduces the “National AI Innovation and Application Pioneer Zone” policy as a case study. The DID method is used for additional verification, comparing the differences before and after the experiment for both the experimental and control groups to mitigate potential endogeneity problems. Considering the chronological order of cities entering the pioneer zone for national AI innovation and application, this study constructs the following model for policy evaluation:

$${\mathrm{E}\mathrm{U}}_{\mathrm{i}\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{D}\mathrm{I}\mathrm{D}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_2{\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\phi }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(2)

$${\mathrm{D}\mathrm{I}\mathrm{D}}_{\mathrm{i}\mathrm{t}}={\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}_{\mathrm{i}}\times {\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}}_{\mathrm{t}}$$

(3)

Based on this model, ${\mathrm{D}\mathrm{I}\mathrm{D}}_{\mathrm{i}\mathrm{t}}$ is defined as the primary explanatory variable. The variable ${\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}_{\mathrm{i}}$ is employed to differentiate the intervention group from the comparison group. The variable ${\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}}_{\mathrm{t}}$ represents the policy indicator variable. Thus, ${\mathrm{D}\mathrm{I}\mathrm{D}}_{\mathrm{i}\mathrm{t}}$ is a binary variable that indicates whether a city is a pilot city for the “National AI Innovation and Application Pioneer Zone”; it is set to 1 if city i is designated as such in year t, and it is set to 0 otherwise. The definitions of the remaining variables are the same as those in Model (1).

Table 4. DID baseline regression results.

Variable	(1)	(2)
	EU	EU
DID	0.2724 ***	0.2735 ***
	(0.0454)	(0.0453)
Constant	0.3352 ***	−0.0266
	(0.0013)	(0.2047)
Controls	No	Yes
City fixed	Yes	Yes
Year fixed	Yes	Yes
R-squared	0.6646	0.6687

Note: *** p < 0.01, with the figures in parentheses representing robust standard errors.

shows that the DID coefficients are consistently positive and significant, suggesting that the “National AI Innovation and Application Pioneer Zone” significantly and stably enhances energy utilization efficiency. This outcome effectively demonstrates the reliability and accuracy of the findings in this study.

The parallel trends assumption should be satisfied to ensure the validity of the DID method, which involves testing whether the treatment and control groups maintain consistent trends before the policy is enforced. Most studies utilize event analysis methods for pre-trend tests, using the significance of pre-event coefficients as a reference standard; if most pre-event coefficients are not significant, the finding supports the parallel trends assumption [71]. Given the long period before the founding of the “National AI Innovation and Application Pioneer Zone” in the sample, this research utilizes the methodology of Zhao et al. [72] by choosing the six years leading up to the implementation and the two years following it as the study period. Using the year before the policy as a baseline, the multi-period DID approach is used for the examination of parallel trends, and the result is presented in Figure 2.

The results indicate that before becoming a pilot city for the “National AI Innovation and Application Pioneer Zone,” energy utilization efficiency exhibited a slight upward trend with minor fluctuations, and the relative coefficients for each year were not significant. This suggests that before the establishment of the AI pilot zone, both the treatment and control cities fulfilled the parallel trends assumption.

To mitigate the risk of biases within the regression results due to unobservable factors or the omission of key urban characteristics, this study conducts a placebo test to enhance the reliability of its findings. Specifically, a spatial placebo test is employed. In this test, during the policy implementation year, several cities equal to the treatment group are randomly selected from the full sample and designated as a “pseudo-treatment group.” Regression analysis is then performed based on Equation (2). This process is repeated 1000 times to generate distributions of estimated coefficients and p-values. The results, as shown in Figure 3, indicate that most of the placebo regression coefficients are centered around zero and follow a normal distribution. Furthermore, most of the corresponding p-values surpass 0.1, indicating that the results from random sampling are not statistically significant. Notably, the estimated coefficient is outside the bounds of the placebo estimates and differs substantially from them. These findings suggest that the outcomes reported in this research are unlikely to be due to random chance, and the influence of unobservable variables is effectively ruled out.

Drawing from the abovementioned findings, which reveal that the establishment of the “National AI Innovation and Application Pioneer Zone” significantly improves energy utilization efficiency, this study further investigates whether the policy generates spillover effects on neighboring cities. Following the approach of the existing literature, cities whose administrative boundaries are adjacent to those of the pilot cities are defined as neighboring cities. An explanatory variable, DID_neighbor_it, is constructed for these cities. If city i is a neighboring city in year t, then DID_neighbor_it takes the value of 1; otherwise, the value is 0. Next, the core explanatory variable in Equation (2) is replaced with DID_neighbor_it, and the original pilot cities are excluded from the sample to conduct the statistical analysis.

Table 5. Test of urban spillover effects of the National AI Innovation Pilot Zones.

Variable	(1)	(2)
	EU	EU
DID_neighbor	−0.0032	−0.0047
	(0.0139)	(0.0136)
Constant	0.3305 ***	−0.1019
	(0.0013)	(0.2043)
Controls	No	Yes
City fixed	Yes	Yes
Year fixed	Yes	Yes
R-squared	0.6650	0.6702

Note: *** p < 0.01, with the figures in parentheses representing robust standard errors.

reveals that there is no notable link between neighboring cities and energy utilization efficiency, indicating that the policy has yet to generate notable spillover effects on adjacent areas.

4.3. Robustness Tests

To confirm the robustness of the findings, this study conducts a series of supplementary tests by referring to the existing literature. These include replacing both explanatory and explained variables, applying winsorization to the data, and excluding samples from directly governed municipalities. Specifically, variable substitution helps to mitigate potential specification errors, while winsorization reduces the influence of extreme values. In addition, considering that municipalities typically have stronger economic foundations and more advanced levels of AI development, excluding these samples contributes to avoiding the distortion of regression results resulting from special cases. The findings of the analysis are summarized in

Table 6. Robustness checks.

Variable	Metrics Replacement					Tail Reduction Treatment	Reduce the Sample Size	Replace the Interpolation Method
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	EU	EU	EU	EU	LnEnc	EU	EU	EU
Robotstock	0.0007 ***
	(0.0001)
Company		0.0014 ***
		(0.0002)
Patents1			0.0357 ***
			(0.0034)
Patents2				0.0347 ***
				(0.0037)
AI					−0.0268 ***	0.0040 ***	0.0036 ***	0.0033 ***
					(0.0040)	(0.0008)	(0.0007)	(0.0007)
Constant	−0.1169	0.0227	−0.0821	−0.0607	−0.0121	0.0912	−0.1645	0.0178
	(0.2031)	(0.2034)	(0.2041)	(0.2039)	(0.9380)	(0.2136)	(0.2066)	(0.2027)
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
City fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R-squared	0.6538	0.6769	0.6985	0.6971	0.7338	0.6620	0.6555	0.6799

Note: *** p < 0.01, with the figures in parentheses representing robust standard errors.

. First, the core explanatory variable is substituted: Given that the AI measurement data may contain randomness, industrial robot stock density (Robotstock) was used as a substitute for industrial robot installation density. Column (1) demonstrates that the regression parameter for Robotstock is 0.0007 and is positively significant, suggesting that the findings remain stable during the substitution of the AI measurement. Columns (2), (3), and (4) use the number of local AI companies (Company), the number of AI patent applications (Patents1), and the number of AI patent grants (Patents2), respectively, as substitutes for industrial robot installation density. The results are similar to those in Column (1), further substantiating the robustness of the baseline results. Second, the explained variable is substituted: While energy utilization efficiency is crucial, energy consumption also has a significant impact on determining its overall efficiency. Adopting the methodology outlined by Shi and Li, we use the ratio of energy consumption to actual GDP and take its logarithmic value to serve as the indicator of energy consumption (LnEnc) [63]. The findings presented in Column (5) show that AI technology applications can significantly reduce energy consumption, with the baseline results remaining robust. Third, two-sided winsorization at 1% is applied: Considering that the existence of outliers in the data may lead to bias in empirical results, a two-sided winsorization at 1% was applied to the dataset. The findings in Column (6) verify that the stability of the baseline regression results remains intact. Fourth, outlier samples are excluded: Due to the distinctive economic and geographical advantages of directly administered municipalities, which may skew the overall empirical results, the samples from the four directly administered municipalities are excluded to evaluate the robustness of the empirical findings. The results in Column (7) again affirm the robustness of the previous findings. Fifth, to avoid bias in the research results caused by linear interpolation, this study employs the Random Forest Regression method to interpolate missing data. The regression results are presented in Column (8), where the influence of AI technology applications on energy utilization efficiency remains positively significant, indicating the robustness of the findings.

4.4. Mechanisms

As proposed in the research hypotheses earlier, AI technology applications have the potential to enhance energy utilization efficiency through technological effects and scale effects. Therefore, this section builds on the previous discussion to examine the impact mechanism of AI technology applications on energy utilization efficiency, specifically analyzing how AI affects energy efficiency. Following the research methodology of Baron and Kenny [73], the following model is established for mechanism testing:

$${\mathrm{M}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1{\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\alpha }}_2{\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\phi }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(4)

In the model presented, M refers to the variables that capture Tecn and Aeg, while the interpretations of the remaining variables align with those presented in Model (1).

4.4.1. Technological Effects Mechanism

To measure the technological effects, the level of green technological innovation (Tecn) is selected, specifically following the research methods proposed by Dong and Wang, the quantity of green patents awarded is used as an indicator to evaluate the level of green technological innovation [74]. Column (1) of

Table 7. Mechanistic influence of technological effects and scale effects.

Variable	(1)	(2)
	Tecn	EA
AI	0.0656 ***	0.0379 ***
	(0.0060)	(0.0038)
Constant	−0.7391	−0.7923 *
	(0.9789)	(0.4343)
Controls	Yes	Yes
City fixed	Yes	Yes
Year fixed	Yes	Yes
R-squared	0.6833	0.8441

Note: * p < 0.1, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

shows that the coefficient of AI is significantly positive. This result shows that a 1% rise in AI technology applications corresponds to a 6.56% enhancement in green technological innovation. Therefore, AI technology applications can significantly optimize the technological effects in urban settings. Furthermore, the research hypotheses previously discussed suggest that a rise in green technological innovation enhances energy utilization efficiency, meaning that AI technology applications improve energy utilization efficiency through green technological innovation. Based on these findings, Hypothesis 2 is confirmed.

4.4.2. Scale Effects Mechanism

To measure the scale effects, the level of economy agglomeration (EA) is adopted, specifically drawing on the research methods of Ciccone and Hall, by calculating the ratio of regional GDP to the area of administrative regions to determine the level of economy agglomeration [40]. Table 7 displays the results. Column (2) reports a coefficient of 0.0379 for AI, and it passes the significance test, confirming its statistical relevance and indicating that a 1% advancement in the level of AI technology applications leads to a 3.79% increase in the level of economic agglomeration. Furthermore, the research hypotheses discussed previously indicated that an elevation of economic agglomeration fosters advancements in energy utilization efficiency. Thus, economic efficiency resulting from economic agglomeration is crucial in determining how AI technology applications contribute to improving energy utilization efficiency. Based on these findings, Hypothesis 3 is confirmed.

4.4.3. Analysis of the Mechanistic Effects of Input Variables and Output Variables

As discussed previously, the explained variable in this research is the SBM-ML index. The input variables consist of fixed capital stock (Klstru), energy consumption (EC), and the number of employees (Labor). Additionally, this study distinguishes between desirable and undesirable outputs, with regional industrial wastewater, waste gas, and solid waste considered undesirable outputs, while the GDP is measured as the desirable output. The total emissions of these industrial pollutants are used to assess the level of undesirable output (UO). This section logarithmically transforms and standardizes both input and output factors, further investigating the mechanisms by which each input and output factor operates. It analyzes the channels through which these factors relate AI technology applications to energy utilization efficiency and specifically assesses whether its impact primarily occurs through the input or the output.

Analysis of the Mechanistic Effects of Input Factors

Column (1) of

Table 8. Mechanistic effects of inputs and outputs.

Variable	(1)	(2)	(3)	(4)	(5)
	LnKlstru	LnEc	LnLabor	GDP	UO
AI	−0.014 ***	−0.0282 ***	0.0062 ***	0.0402 ***	−0.0269 ***
	(0.0012)	(0.0041)	(0.0020)	(0.0031)	(0.0055)
Constant	10.9021 ***	10.6623 ***	−0.9288 *	−1.1075 ***	0.1955
	(0.4690)	(1.2186)	(0.4846)	(0.4737)	(0.7791)
Controls	Yes	Yes	Yes	Yes	Yes
City fixed	Yes	Yes	Yes	Yes	Yes
Year fixed	Yes	Yes	Yes	Yes	Yes
R-squared	0.9828	0.8895	0.9449	0.9158	0.7970

Note: * p < 0.1, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

indicates a significantly negative coefficient for AI, showing that the application of AI technology can effectively reduce fixed capital stock. The decrease in fixed capital stock allows enterprises to allocate more available capital for innovation and other capital expenditures. Additionally, a lower level of fixed capital stock can enable industrial enterprises to achieve the same output, thereby improving energy utilization efficiency. Column (2) shows that the coefficient of AI remains significantly negative. This suggests that AI technology applications can significantly reduce energy consumption in production, and the reduction in energy consumption will greatly enhance energy utilization efficiency. Therefore, energy consumption acts as a crucial mechanism in the effect of AI technology applications on energy utilization efficiency. Column (3) shows that the application of AI technology results in a rise in the number of employees; however, there is generally an inverse connection between the employee count and energy utilization efficiency. Thus, the input factor of the number of employees does not exhibit a significant mechanistic effect.

Analysis of the Mechanistic Effects of Output Factors

Column (4) of Table 8 indicates that the regression coefficient for AI is 0.0402, which shows statistical significance at the 10% level, indicating that the application of AI technology can result in growth in the GDP of urban areas. A higher GDP can provide more economic support and conditions for related technological development, thereby contributing to improvements in energy utilization efficiency. In Column (5), the regression coefficient for AI technology applications regarding undesired output is negatively significant, indicating that the application of AI technology can significantly decrease undesirable outputs in production. The reduction of undesirable outputs in the production process is bound to enhance energy utilization efficiency, indicating that undesirable outputs have a certain mechanistic effect.

In summary, from the perspective of production inputs, the application of AI technology significantly improves energy utilization efficiency, primarily by reducing fixed capital stock and energy consumption, while the number of employees does not exhibit a mechanistic effect between these two factors. From the output perspective, the application of AI technology can significantly improve energy utilization efficiency by boosting positive outputs and minimizing negative outputs, with both types of output demonstrating good mechanistic effects.

4.5. Moderating Effects

As previously discussed, the strengthening of environmental regulation (ERI) and the improvement of digital infrastructure (DI) have the potential to strengthen the role of AI technology applications in enhancing energy utilization efficiency. To examine the moderating effects of these two factors, this study references the model constructed by Holland et al. as follows [75]:

$${\mathrm{E}\mathrm{U}}_{\mathrm{i}\mathrm{t}}={\mathsf{\gamma }}_0+{\mathsf{\gamma }}_1{\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}+{{\mathsf{\gamma }}_2{\mathrm{N}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\gamma }}_3{\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}\times {\mathrm{N}}_{\mathrm{i}\mathrm{t}}+\mathsf{\gamma }}_4{\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\phi }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(5)

In this model, ${\mathrm{N}}_{\mathrm{i}\mathrm{t}}$ represents the moderating variables of environmental regulation and digital infrastructure, while ${\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}\times {\mathrm{N}}_{\mathrm{i}\mathrm{t}}$ denotes the explanatory variable of AI technology applications and the interaction term with the moderating variables.

4.5.1. Moderating Effects of Environmental Regulation

This study adopts existing practices by measuring environmental regulation using the overall recycling rate of industrial solid waste. The regression results in

Table 9. Moderating effects of environmental regulations and digital infrastructure.

Variable	(1)	(2)	(3)	(4)
	EU	EU	EU	EU
AI	0.0036 ***	0.0034 ***	0.0035 ***	0.0003
	(0.0007)	(0.0007)	(0.0007)	(0.0012)
LnERI	0.0048	0.0076
	(0.0064)	(0.0063)
LnDI			−0.0126 *	−0.0062
			(0.0069)	(0.0069)
AI		0.0029 ***
		(0.0007)
AI × LnDI				0.0028 ***
				(0.0008)
Constant	−0.0981	−0.0405	−0.1640	−0.2351
	(0.2051)	(0.2077)	(0.2067)	(0.2086)
Controls	Yes	Yes	Yes	Yes
City fixed	Yes	Yes	Yes	Yes
Year fixed	Yes	Yes	Yes	Yes
R-squared	0.6525	0.6536	0.6530	0.6557

Note: * p < 0.1, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

reflect the moderating effects of environmental regulation. Column (1) of Table 9 shows that the coefficient of AI is significantly positive. Column (2) examines the moderating effects of environmental regulation, with the estimated regression coefficient for AI×LnEri being 0.0029 and positively significant. This suggests that environmental regulation can significantly amplify the influence of AI technology applications on energy utilization efficiency. Strengthened environmental regulation can send signals for green development to enterprises, prompting them to focus more on the advancement of AI technology applications and strive to enhance their technological levels to maximize productivity. Thus, this external factor of environmental regulation can foster the advancement of AI technology applications and technological innovation, thereby positively influencing the application of AI technology to enhance energy utilization efficiency through technological effects. Based on these findings, Hypothesis 4 is confirmed.

4.5.2. Moderating Effect of Digital Infrastructure

This study references the measurement techniques outlined by Zhu et al. and identifies five key dimensions for analysis [76]. The study incorporates several indicators to assess the implementation of digital infrastructure, including the internet user population per 100 individuals, the proportion of the workforce in the IT services and software sectors, the individual-based volume of telecommunications services, the mobile phone population per 100 individuals, and the index of digital finance inclusion. To evaluate the degree of digital infrastructure adoption, the entropy method is applied. Table 9 demonstrates the moderating effects of digital infrastructure. Column (3) presents the outcomes that include the moderating variable of digital infrastructure, with the results being consistent with those in Column (1), which were discussed earlier. Column (4) examines the moderating effect of digital infrastructure, revealing an estimated regression coefficient of 0.0028 for AI × LnDI, which passes the significance test at the 10% level. This indicates that the level of digital infrastructure significantly moderates the relationship between AI technology applications and energy utilization efficiency. The level of digital infrastructure can influence urban economic, technological, and AI levels to some extent; therefore, an elevated degree of digital infrastructure can boost the technological and economic agglomeration levels of cities. This, in turn, effectively promotes the application of AI technology to improve energy utilization efficiency through technological and scale effects. The level of a city’s digital infrastructure has a significant moderating effect on the link between AI technology applications and energy utilization efficiency, an outcome that is consistent with the findings of Lee et al. [77]. Drawing from these findings, Hypothesis 5 is confirmed.

4.6. Heterogeneity Analysis

4.6.1. Heterogeneity of Geographical Location

Owing to the disparities in economic advancement stages and the application of AI across regions, this research classifies cities into four distinct regions: eastern, central, western, and northeastern, with the northeastern region including cities from the Liaoning, Heilongjiang, and Jilin provinces. The findings shown in

Table 10. Heterogeneity test results regarding geographical location.

Variable	East	Central	West	Northeast
	EU	EU	EU	EU
AI	0.0038 ***	0.0033 ***	0.0039 ***	0.0019
	(0.0014)	(0.0010)	(0.0012)	(0.0042)
Constant	−0.4844	−0.3207	−0.2934	−2.2362
	(0.5161)	(0.2658)	(0.3160)	(0.5903)
Controls	Yes	Yes	Yes	Yes
City fixed	Yes	Yes	Yes	Yes
Year fixed	Yes	Yes	Yes	Yes
R-squared	0.5776	0.6527	0.7112	0.6165

Note: *** p < 0.01, with the figures in parentheses representing robust standard errors.

reflect the heterogeneity of geographical locations. The findings indicate that the promotion of energy utilization efficiency through the increased application of AI technology is most effective in the western region, and the impact in the eastern region is similarly significant. As it is relatively developed, the central region also shows a significant positive effect, with all three regions’ regression results passing the 1% significance test. However, for the northeastern region, the regression coefficient for AI technology applications associated with energy utilization efficiency is favorable, but it fails to achieve statistical significance in the test. The main reason might be attributed to the northeastern region, which, as a traditional industrial base in China, has a high concentration of conventional industries, particularly heavy and chemical industries, which present considerable challenges for industrial restructuring. In addition, the region’s GDP growth has been slowing in recent years and has consistently lagged behind that of the eastern and central-western regions, resulting in a relatively low level of economic development. This has, to some extent, hindered the promotion of AI technologies and the advancement of the digital economy, thereby weakening the effectiveness of AI technology applications in improving energy utilization efficiency. Consequently, compared to other regions, the northeastern region has significantly lagged in leveraging AI to enhance energy utilization efficiency and has not yet demonstrated a notable positive impact.

4.6.2. Heterogeneity of Urban Types

First, energy utilization efficiency is inseparable from energy consumption, and the resource conditions in different regions can impact their respective baseline regression results. Therefore, this study classifies the 241 cities as either resource-based or non-resource-based based on the “National Sustainable Development Plan for Resource-Based Cities (2013–2020).” Secondly, following the research methods of Shi and Li [63], the selected cities are categorized as old industrial-based cities and non-old industrial-based cities. Finally, policies play a significant role in whether enterprises are willing to use AI to improve energy utilization efficiency. This research divides the sample into key and non-key environmental protection cities based on the “National Environmental Protection 11th Five-Year Plan” published by the State Council in 2007.

Table 11. Heterogeneity test results regarding urban types.

Variable	Resource-Based City	Non-Resource-Based City	Old Industrial Base City	Non-Old Industrial Base Cities	Key Environmental Protection City	Non-Key Environmental Protection City
	EU	EU	EU	EU	EU	EU
AI	0.0022 *	0.0029 ***	−0.0007	0.0053 ***	0.0032 **	0.0001
	(0.0012)	(0.0009)	(0.0008)	(0.0009)	(0.0013)	(0.0009)
Constant	−0.0026	−0.0201	−0.5499	−0.0212	−0.5407	−0.3282
	(0.2827)	(0.2987)	(0.3001)	(0.2709)	(0.5801)	(0.2084)
Controls	Yes	Yes	Yes	Yes	Yes	Yes
City fixed	Yes	Yes	Yes	Yes	Yes	Yes
Year fixed	Yes	Yes	Yes	Yes	Yes	Yes
R-squared	0.6635	0.6346	0.7540	0.6106	0.6860	0.6311

Note: * p < 0.1, ** p < 0.05, and *** p < 0.01, with the figures in parentheses representing robust standard errors.

provides the findings of the heterogeneity test conducted across different urban types. The results confirm that regardless of whether a city is resource-based, energy utilization efficiency is positively influenced; however, the significance level for non-resource-based cities is 1%, whereas for resource-based cities, it is only 10%. The estimated parameter for AI in cities with an old industrial base is not statistically meaningful, while the regression parameter for AI in non-old industrial base cities is 0.0053 and significant at the 1% level. Lastly, in key environmental protection cities, AI technology applications have a positive and statistically significant effect on energy utilization efficiency, whereas the impact in non-key environmental protection cities is minimal and not significant. Several factors contribute to these findings: for resource-based cities, although AI technology applications can improve energy utilization efficiency, their industrial development primarily relies on resource exploitation, which may lead to a stronger positive effect than in other cities. Cities with an established industrial foundation boast a lengthy history of industrial development and a solid industrial foundation, making it more challenging for AI technology applications to influence energy utilization efficiency. In contrast, non-old industrial base cities can better integrate AI technology applications with energy utilization efficiency. In key environmental protection cities, local enterprises are likely to invest more in enhancing energy utilization efficiency and reducing energy consumption, thereby facilitating a better integration of AI technology applications with energy utilization efficiency, while non-key environmental protection cities do not effectively encourage enterprises to focus on energy utilization efficiency.

5. Conclusions and Implications

Based on panel data from Chinese cities from 2008 to 2021, this research systematically investigates the impact of AI technology applications on energy utilization efficiency and its underlying mechanisms. The findings indicate the following conclusions: First, the application of AI technology significantly enhances energy utilization efficiency, and the credibility of these results is validated through endogeneity and robustness tests. Second, the improvement in energy utilization efficiency driven by AI technology applications is primarily realized through technological effects and scale effects. Further analysis reveals that environmental regulations and digital infrastructure beneficially moderate the connection between AI technology applications and energy utilization efficiency. Moreover, a heterogeneity analysis shows that cities with stronger policy support, lower resource dependence, and non-old industrial bases tend to better leverage AI technology applications to boost energy utilization efficiency.

The research findings yield several policy implications: (1) The promotion and application of AI can be accelerated to facilitate high-quality energy development. Specifically, government departments can actively promote the deployment of industrial energy efficiency optimization systems to enable the monitoring and management of energy consumption data; promote the development of predictive maintenance systems to achieve the real-time forecasting of equipment maintenance costs and production efficiency; and improve the construction of intelligent power distribution systems to optimize electricity allocation and supply efficiency. The promotion and application of these AI tools will help achieve energy conservation and efficiency, thereby supporting high-quality energy development. (2) The establishment of digital infrastructure can be accelerated to empower the application of AI technology in improving energy utilization efficiency. Specifically, relevant departments can accelerate the deployment of edge computing nodes to minimize delays and failure risks during system scheduling and improve IoT sensing systems to enable the comprehensive and precise management of energy systems. As the operational foundation for AI technology, the systematic development of digital infrastructure is of great significance for enhancing energy utilization efficiency. (3) The construction of environmental governance systems should be strengthened, and AI-driven green development should be promoted. Government departments can draw on international best practices to improve relevant legal frameworks and policy systems, thereby supporting the high-quality development of AI technology. For example, in May 2023, the U.S. Department of Energy launched the “AI for Science, Energy, and Security” initiative to actively promote the application of AI within the energy field. Seraj et al., using the UK’s residential building energy performance certificate dataset, demonstrated that the utilization of AI models can significantly boost energy efficiency in building processes [78]. Meanwhile, in 2020, Saudi Arabia released its “National Strategy for Data and Artificial Intelligence,” aiming to drive the transition from a petroleum-based economy to an energy powerhouse through the use of AI. These policies provide valuable references for countries promoting AI-powered green development. (4) This research draws on city-level data from China and proposes policy recommendations accordingly, but different developing countries exhibit significant variations in economic development and institutional environments. Therefore, countries should formulate policy measures that are appropriate to their national contexts. For instance, in developing countries with relatively weak digital infrastructure, governments can support enterprise-level AI development by offering public cloud service subsidies, supporting AI pilot demonstration projects, and improving regulatory and policy frameworks. Meanwhile, priority should be given to promoting the deployment of edge computing nodes and portable energy consumption sensors—low-barrier AI tools that can offer robust assistance for the development of AI.

Although this study assesses the implications of AI technology applications in terms of energy utilization efficiency, it continues to exhibit certain constraints that necessitate further enhancement and exploration in subsequent studies. First, considering data availability and representativeness, this study employs industrial robot density as an indicator of the degree of AI technology applications. However, this variable primarily reflects robotics-based AI and does not fully capture other forms of non-robotic AI technologies. Future research could expand the construction of the indicator system to include and examine the influence of non-robotic AI technology applications—such as intelligent recommendation systems, smart scheduling systems, and AI decision engines—within the industrial sector. Second, the analysis in this study is based on city-level data from China, which imposes certain limitations on the scope of the research. Future studies may delve into firm-level or industry-level data to conduct more micro-level investigations. At the same time, incorporating cross-country data could enable comparative analyses, thereby broadening the scope and enhancing the generalizability of the findings. Furthermore, this study relies on data from statistical yearbooks, excluding cities with extensive missing data and applying linear interpolation to fill in partial gaps. Since linear interpolation assumes a time-based linear trend, it may smooth out natural data fluctuations and potentially introduce bias into the estimation results. Future research could address this issue by using more comprehensive datasets and adopting machine learning techniques or time series models to handle missing data, thereby improving the reliability and scientific rigor of the results.