Research on Whether Artificial Intelligence Affects Industrial Carbon Emission Intensity Based on the Perspective of Industrial Structure and Government Intervention

Abstract

Artificial intelligence serves as the fundamental catalyst for a new wave of technological innovation and industrial transformation. It holds vital importance in reaching carbon reduction targets and the objectives of “carbon peak and neutrality”. This factor contributes significantly to the reduction in carbon emissions in the industrial domain. This article utilizes panel data from 30 provinces in China, covering the years 2013 to 2021, to develop an evaluation framework for assessing the progress of artificial intelligence development. Through the use of double fixed-effect models, mediation effect models, and threshold effect models, the empirical analysis examines the industrial carbon reduction effects of artificial intelligence and its operating mechanisms. Research indicates that the advancement of AI can significantly reduce carbon emission intensity within the industrial sector. This conclusion remains valid following comprehensive robustness tests. Furthermore, there exists temporal and regional variability in AI’s impact on industrial carbon reduction, particularly more pronounced after 2016 and in central and western regions. AI influences carbon emission reduction in China’s industrial sector through the advancement and optimization of industrial structures. Here, the increase in senior-level operations acts as a partial masking effect, while optimization serves as a partial mediator. The relationship between AI and industrial carbon emission intensity is non-linear, being influenced by the threshold of government intervention; minimal intervention weakens AI’s effect on carbon intensity reduction. These findings enhance our understanding of the factors influencing industrial carbon emissions and contribute to AI-related research. They also lay a solid empirical groundwork for promoting carbon emission reduction in the industrial domain via AI. Additionally, the results offer valuable insights for formulating policies aimed at the green transformation of industry.

1. Introduction

In the context of global warming, the increasing emission of global greenhouse gases (GHGs) represented by carbon dioxide (CO₂) has led to the melting of glaciers, sea level rise, frequent droughts and floods, presenting a serious risk to human existence and advancement [1]. China first suggested a “dual carbon” target of “carbon peaking” by 2030 and “carbon neutrality” by 2060 at the 73rd session of the UN General Assembly [2]. Industry is the pillar of China’s economic development, but its “high-cost, high-consumption, high-pollution” development mode causes serious ecological and environmental problems, and its carbon emissions consist of up to 70% of China’s total carbon emissions, becoming the largest carbon emission source [3]. In the realm of fostering low-carbon development, advocating for the reduction in carbon emissions within the industrial sector emerges as a pivotal undertaking crucial to combatting the challenges posed by global climate change.

Artificial intelligence, viewed as the central impetus behind the latest technological revolution and industrial evolution, possesses the ability to actively oversee the real-time emission of pollutants, and optimize it to help enterprises complete the intelligent transformation and enhance the entire process from R&D to marketing; thus, the purpose of energy saving and emission reduction is achieved, which is regarded as a new way to achieve accelerated carbon emission reduction in the industrial sector [4]. However, most of the studies on the subject use developed economies as the object of research for empirical summaries, and there is a lack of empirical evidence from China. However, it is difficult to provide an inspiring Chinese program for the development of AI and green low-carbon transition in developing countries. In addition, it has been found that with the popularity of machine learning systems, AI systems may consume too much electricity and produce large amounts of carbon emissions [5]. So, is it possible to reduce carbon emissions through AI? What are the mechanisms and logic? How does the impact of AI on carbon emissions differ across heterogeneous factors? Studying the diverse factors that influence the effects of AI on carbon emissions is crucial in the current context of rapid digital technology advancement and the economic shift towards low-carbon practices. This analysis holds both theoretical importance and practical relevance for the Chinese industry in achieving carbon reduction targets.

The novelties of this work in comparison to previous research are as follows: Firstly, from the research perspective, it is the first time that AI is linked with carbon emissions in the trial field, which provides a new idea for subsequent research. Secondly, it establishes a system of AI evaluation indexes, which is innovative in measuring the level of the development of AI. Thirdly, theoretically, it analyzes the three levels of the direct effect, the indirect effect, and the threshold effect of AI on carbon emissions in the industrial sector in multiple dimensions, which lays the foundation for an in-depth understanding of the impact of AI on carbon emissions in the industrial sector. Fourthly, on the basis of empirical analysis, the role of industrial structure advanced, industrial structure rationalization and government intervention in the development of AI on industrial carbon emissions is explored. Finally, this paper analyzes the heterogeneity in time and region to provide insights for the formulation of ‘dual-carbon’ policies in different regions.

The remainder of this article is structured as follows: Section 2 reviews the relevant literature on AI and industrial carbon emissions; Section 3 provides a theoretical framework incorporating four hypotheses as well as the relevant theoretical analyses; Section 4 carries out the research design, including the construction of the econometric model and the selection of the variables and the data descriptions; Section 5 presents the empirical results of the relevant models designed in the previous section as well as the relevant analyses; and Section 6 presents the conclusions and recommendations.

2. Literature Review

Artificial intelligence was first proposed by Marvin Minsky, the ‘father of artificial intelligence’, in 1956, and scholars have put forward different understandings and definition standards for artificial intelligence according to different research fields and directions, but the core concept is basically the same. Artificial intelligence (AI) is an emerging interdisciplinary field that revolves around the development of algorithms and computational models to enable machines to perform tasks that typically require human intelligence. This innovative branch of technology aims to create systems capable of learning, reasoning, and problem-solving, with the ultimate goal of mimicking human cognitive functions. At present, there is no agreement in the academic community on the measurement standard of AI and how to quantitatively analyze the concept of AI, and the main measurement methods include the single-indicator method and the multi-indicator construction method. The single-indicator method mainly consists of three kinds of methods: The first is to use the “value added of information transmission, computer services and software industry” to represent; this method is built on the basis of the positive causal relationship between capital investment and the level of development, and the investment in artificial intelligence to represent the scale of the development of artificial intelligence [6]. The second is to use the method of figurative description; robots are the product of artificial intelligence, and their number reflects the level of development of artificial intelligence to a certain extent; so, the logarithm of robot density is used to measure the level of artificial intelligence [7]. The third is to use the number of patents on artificial intelligence as a metric [8], to take the technology of artificial intelligence as a comprehensive technological change in economic and social fields, and to explore its impact on the entire economic paradigm. The multi-indicator construction method involves the use of multiple indicators to evaluate, and there are fewer applications in constructing an evaluation system for AI development. Sun Zao and Hou Yulin [9] measured industrial intelligence from the three aspects of infrastructure, production and application, and competitiveness and benefits; Lv Rongjie and Hao Lixiao measured China’s inter-provincial AI development index from the four dimensions of institutional environment, infrastructure, technological innovation, and production and application [10].

With the rapid advancement of artificial intelligence, many scholars have analyzed the impact it has had. At first, the impacts on the economic and social aspects were analyzed. Artificial intelligence has emerged as a significant emblem of the contemporary scientific and technological revolution within the era of extensive data, exerting a profound influence on the economic advancement of nations globally [11], and it is considered an important way to promote economic growth [12]. Some scholars have proposed that the replacement effect of robots decreases the demand for labor [13]. However, some scholars have contended that advancements in robotics not only create a significant number of new employment opportunities but also lead to the replacement of existing jobs [14]. Nevertheless, some scholars have contended that the impact of robotics’ applications on employment leans more towards promotion rather than substitution. They argue that the use of robots has substantially raised the rate of labor employment within industrial enterprises [15]. More recently, with the escalating severity of global climate change and environmental issues, there has been a growing emphasis on the utilization of artificial intelligence in environmental monitoring and governance [16]. Artificial intelligence technology has the capability to enhance the operational conditions of businesses and facilitate the advancement of wastewater treatment facilities, and effectively improve the pollutant control efficiency [17], which is of great significance for reducing environmental pollution [18].

China has consistently prioritized the matter of climate change, demonstrating a strong commitment to addressing this global challenge. The reduction in carbon emissions has emerged as a central area of scholarly inquiry, drawing significant academic interest and scrutiny. Existing studies have mainly analyzed carbon-emission-influencing factors from three aspects. One area of study pertains to the economic factor, which primarily investigates the correlation between economic growth and carbon emissions [19]. Another key focus is on demographic factors, where researchers predominantly examine the impacts of population aging [20], population agglomeration [21] and population mobility [22] on carbon emissions. Thirdly, we start from the institutional aspect, focusing on the carbon emission reduction effects of the carbon-trading system [23], carbon tax [24], and environmental regulatory policies [25]. Against the background of generalized technological development, a few scholars began to pay attention to the impact of ICT on carbon emissions and energy consumption [26], and believed that there is industrial and national heterogeneity in its reduction in carbon emissions [27]. With the continuous development of AI and its application in various industries, its impact on industry carbon emissions has also attracted more and more attention from academics, but no consensus has been formed. Some scholars point out that the use of AI can effectively reduce carbon emissions, but its role is phased and there is industrial heterogeneity [28], while some scholars argue that there is an inverse U-shaped relationship between AI and total carbon emissions, positing that the emission reduction effect of AI will become increasingly pronounced with its advancement [29]. However, there exists a noticeable deficiency in research concerning the impact of artificial intelligence on particular sectors, including industrial carbon emissions, along with related studies on its operational mechanisms, which form the central theme of this paper.

3. Theoretical Analysis and Research Hypotheses

3.1. Artificial Intelligence’s Direct Effects on Industrial Releases of Carbon

Artificial intelligence can reduce industrial carbon emissions in three ways: effectively allocating resources, improving energy use efficiency, and fueling technological innovation.

First, AI can enhance the effective distribution of energy resources and regulate carbon emissions by minimizing the usage and inefficiencies of fossil fuels while encouraging the amalgamation and enhancement of renewable energy sources. On the one hand, artificial intelligence technology is introduced into energy consumption management, and an energy consumption control platform based on digital infrastructure is constructed to achieve full-cycle and full-process spatial and temporal monitoring of energy consumption, discover energy consumption and inefficiency, and carry out scientific prediction [30], decision-making [31] and regulation of energy consumption and load peaks through digital technology so as to achieve more reasonable energy consumption scheduling and supply decisions. Through intelligent power system scheduling, accurate regulation of new energy consumption is achieved, simultaneously reducing the reliance on high-carbon energy sources [32].

Secondly, AI has the potential to enhance the effectiveness of energy consumption [33], such as by assisting in decision-making, optimizing production processes, etc., thereby directly reducing carbon emissions. Artificial intelligence applies advanced technologies such as smart grid, intelligent control, energy saving, new materials, Internet of Things, etc., to reduce energy consumption in the production and transmission process and improve the efficiency of energy use. At the same time, AI has the characteristics of self-learning, self-adaptation, and of being self-driven, which can effectively solve complex and complicated production processes, thus realizing the reconfiguration of services and production processes. By intelligently monitoring the entire production process, pollution can be controlled at its source. Real-time monitoring of high-pollution, high-energy-consumption production processes allows for the accurate identification of these areas, enabling suggestions for improving production techniques and processes, thus forming a moderate pollution control system. This fosters the enhancement in green production efficiency in enterprises.

Finally, the development of artificial intelligence technology will promote technological innovation in related fields, which has very significant advantages in reducing emissions. On the one hand, artificial intelligence is a mechanical device with its own technological progress effect, which broadens the boundaries of production possibilities in the manufacturing industry, guides enterprises to carry out green technological innovation, produces more environmentally friendly products, improves the productivity of factors of production, and increases green output [34]. At the same time, through technological innovation, the development of intelligent and energy-saving integrated technologies can be accelerated and energy consumption can be reduced [7]. As a result, the diffusive innovation characteristics of AI promote the construction of the technological innovation system of the entire industrial chain, and its internal innovation elements promote the development of green technological innovation through agglomeration and reorganization and synergistic configuration.

Therefore, the following hypothesis is proposed:

Hypothesis H1: 

The development of artificial intelligence affects the intensity of industrial carbon emissions.

3.2. Artificial Intelligence’s Indirect Effects on Industrial Releases of Carbon

According to Graetz and Michaels, intelligence and automation enhance industrial production efficiency through disruptive and innovative effects, respectively, while the advanced and rationalized industrial structure corresponds to the vertical enhancement and horizontal transfer of production efficiency within and between industries [7]. Therefore, from the standpoint of industrial structure, this article posits that artificial intelligence has the potential to enhance industrial carbon emission reduction through industrial structure advancement and industrial structure rationalization.

The utilization of artificial intelligence (AI) to advance the industrial structure is primarily demonstrated by its capability to significantly elevate the scientific and technological standards of the entire industry, thereby enhancing the efficiency of utilizing production resources [35]. First, the application of AI promotes the optimization of capital structure, and ‘squeezes out’ the tangible, human and institutional capital from high-energy-consumption and high-pollution industries to flow to knowledge- and technology-intensive clean industries [12], thus realizing the elimination of outdated production capacity in depth and breadth. Changes in capital structure will weaken the squeezing effect on innovation resources, prompting potential innovators and R&D funds to flow from inefficient industries to efficient industries, accelerating technological innovation and improving innovation performance, thus laying the foundation for improving the green efficiency of the industry. Secondly, AI promotes industrial upgrading, which helps break through the low-end lock-in of the manufacturing industry and promotes the development of smart manufacturing, especially intelligent production services [36], and especially the development of smart productive services through the role of effective demand and new jobs [37]. This new type of industry can not only achieve higher energy saving and environmental benefits, but also produce products with ‘green’ characteristics, which provides a strong support for the development of China’s green economy. Finally, artificial intelligence represents a form of disruptive technological advancement that has the ability to consistently alter the current technical production methods. This process facilitates the ongoing enhancement in industrial standards and the establishment of a novel economic framework. A key characteristic of artificial intelligence lies in its capacity to significantly decrease energy consumption within low-end manufacturing sectors by enhancing energy efficiency through autonomous adjustments and deep learning. This development holds substantial importance for the evolution of China’s sustainable economy.

Artificial intelligence can enhance the quality of inter-industry aggregation and promote the rationalization of industrial structure by changing the factor input structure and coordinating the inter-industry combination relationship [38]. On the one hand, due to the extensive penetration characteristics of artificial intelligence, it can accelerate the circulation of production factors and shorten the technological divide between industries. Its empowerment of traditional industries can enhance the original advantageous technology and production mode, improve the added value of traditional industries, and significantly improve the productivity and resource allocation efficiency of enterprises [39]. On the other hand, the innovative characteristics of AI itself also make the innovative elements able to aggregate and integrate, and realize the collaboration with other factors of production so as to accelerate the green technological progress of the industry chain itself [38]. This enhancement, while promoting the application of green technology, will also create a spillover effect, prompting the entire industrial chain to be greener and more sustainable [40], thus building a green technology production system and laying the foundation for the win–win development of economic growth, resource conservation and environmental protection [41].

Based on the above analysis, the following assumptions are made:

Hypothesis H2: 

Artificial intelligence empowers industrial carbon emission reduction through the advanced industrial structure.

Hypothesis H3: 

Artificial intelligence empowers industrial carbon emission reduction through the rationalization of the industrial structure.

3.3. Artificial Intelligence Has a Threshold Effect on Industrial Carbon Emissions

In the case of AI, unregulated and unrestrained development may have negative impacts such as destructive competition and failure to improve artificial productivity. Therefore, when examining the correlation between artificial intelligence (AI) and industrial carbon emissions, it is vital to integrate the presence and role of government intervention into the analytical structure.

The government can directly intervene in the development of AI through industrial policy guidance. On the one hand, AI cannot be separated from the government’s “tangible hand” to carry out top-level design and policy layout. Especially in the context of China’s lack of innovation in basic theories and underlying algorithms, and the inadequacy of the industrial ecosystem and industrial chain, the government should actively participate in guiding and supporting the development of artificial intelligence. On the other hand, the significant “polarization” adjustment of AI on the labor structure [9] may lead to further imbalance in the demand for labor and labor share in the society, which will aggravate social inequality, and it is not realistic to rely solely on technological progress to solve the negative problems of economic growth and population caused by industrial intelligence. It is not realistic to rely solely on technological progress to solve the negative problems of economic growth and population generated by industrial intelligence. If the government does not control the development of AI, capital will flow from the traditional material sector to the intelligent sector, and the distribution of benefits will be emphasized over the other.

Government intervention provides public management and institutional safeguards for carbon governance. The realization of green governance and economic green transformation puts higher demands on the national governance system and the government’s ability to govern, not only requiring incentives from market forces, but also relying on the government to ensure the smooth progress of the process through effective regulation and mechanism design. At the same time, it is necessary to intervene in the exit to compensate for the limitations of the market’s “spontaneous order”. For example, the implementation of a functional carbon emission trading scheme [42] and the implementation of environmental regulations [43] can influence carbon emissions. Therefore, considering the analyses presented from the viewpoints of ‘industrial policy’ and ‘green governance’, the following hypothesis is posited.

Hypothesis H4: 

The suppression effect of AI on industrial carbon emission intensity is non-linear due to government intervention.

3.4. Theoretical Framework

Figure 1 illustrates the theoretical framework of this paper, presenting the relationships between the components discussed above. The framework integrates the four main hypotheses presented in the previous sections, emphasizing the direct impact of AI development on carbon emission intensity in the industrial sector (Hypothesis H1), the indirect impact through the advanced industrial structure and the rationalization of the industrial structure (Hypothesis H2 and Hypothesis H3), and the threshold effect of government intervention (Hypothesis H4). It visually summarizes the interactions and dependencies within the framework, helping to understand the complex dynamics at play, and provides a cohesive structure for analyzing how AI can drive carbon reduction in industry.

The theoretical framework serves as a foundational guide for the empirical analysis in the following sections. By systematically outlining the relationships between AI and carbon emissions, as well as the critical role of government intervention, this framework sets the stage for a comprehensive investigation into the mechanisms that underpin the effectiveness of AI in achieving carbon reduction targets in the industrial sector. In the next section, we will use this framework to empirically demonstrate the relationship between AI and carbon emission.

4. Analytical Empiricism

4.1. Model Setting

4.1.1. Baseline Regression Model

In order to verify whether hypothesis H1 is valid, this paper intends to establish a static panel regression model to analyze the impact of AI on carbon emission intensity in China’s industrial sector, and construct an econometric model as shown in Equation (1):

$${lnCI}_{it}={\alpha }_0+{\alpha }_1{AI}_{it}+{\alpha }_2{Control}_{it}+{\epsilon }_{it}$$

(1)

where ${CI}_{it}$ is the industrial carbon emission intensity of region i in period t, ${AI}_{it}$ is the artificial intelligence technology, ${Control}_{it}$ is a series of control variables, ${\alpha }_0$ is a constant term, ${\alpha }_1$ and ${\alpha }_2$ denote the regression coefficients of the core explanatory variables and the control variables, respectively, and ${\epsilon }_{it}$ is a random error term. Further, if the coefficient ${\alpha }_1$ is significantly positive, it indicates that the AI technology improves carbon emission intensity; if it is significantly negative, the opposite is true.

4.1.2. Mediating Effect Model

Through the analytical framework in Section 3, we propose Hypothesis H2 and Hypothesis H3; that is, the development of AI can affect the carbon emission intensity of China’s industrial sector through the advanced industrial structure and the rationalization of industrial structure. In order to verify hypotheses H2 and H3, this study adopts the methodology proposed by Zhao [44] and constructs the following mediating effect model:

$$ln{M}_{it}={\beta }_0+{\beta }_1{AI}_{it}+{\beta }_2{Control}_{it}+{\epsilon }_{it}$$

(2)

$${lnCI}_{it}={\rho }_0+{\rho }_1{AI}_{it}+{\rho }_2{ln{M}_{it}+\rho }_3{Control}_{it}+{\epsilon }_{it}$$

(3)

Among them, $M_{it}$ is the mediating variable, which is the industrial advanced structure and rationalization, representing the transmission pathway of AI on industrial carbon emission intensity. ${\beta }_0$ and ${\rho }_0$ denote constant terms, ${\beta }_1$, ${\beta }_2$, ${\rho }_1$, ${\rho }_2$ and ${\rho }_3$ denote the parameters to be estimated, and the other variables are the same as in Equation (1).

4.1.3. Threshold Effect Model

Considering that the level of government intervention may constrain the inhibitory effect of AI on the intensity of industrial carbon emissions so that the effect of AI development on industrial carbon emissions is presented as nonlinear (hypotheses H4), a panel threshold effect model is further set up on the basis of model (1) with reference to Hansen [45]:

$$ln{CI}_{it}=a_0+a_1{AI}_{it}·I\left({GI}_{it}\le {\gamma }_1\right)+a_2{AI}_{it}·I\left({\gamma }_1<{GI}_{it}\le {\gamma }_2\right)+\cdots +a_n{AI}_{it}·I\left({GI}_{it}>{\gamma }_n\right)+\theta {Control}_{it}+{\epsilon }_{it}$$

(4)

Model (4) is a multiple-threshold-effect model about AI on industrial carbon emission intensity, ${GI}_{it}$ is a threshold variable, I(·) functions as a schematic indicator, returning 1 when the condition is met and 0 when it is not, ${\gamma }_1$, ${\gamma }_2$, …, ${\gamma }_n$ denote n threshold variables, and the meanings of other variables are the same as those in Model (1). $a_1$, $a_2$, …, $a_n$ are the correlation coefficients of AI on industrial carbon emission intensity under different threshold intervals, and θ is the correlation coefficient of control variables.

4.2. Variable Selection and Data Description

4.2.1. Explained Variable

The explanatory variable in this paper is industrial carbon emission intensity, which is measured using the ratio of total industrial carbon emissions to GDP. Regarding the measurement of total carbon emissions, according to the common practice in academia [46], the formula in the IPCC National Greenhouse Gas Inventory 2006 is used for the measurement.

$$C=\sum _{i=1}^8{C}_i=\sum _{i=1}^8{M}_i\times N_i\times {EC}_i\times {\displaystyle \frac{44}{12}}$$

(5)

The formula includes the following terms: $M_i$ is the energy consumption (unit: tonne), $C_i$ is the energy carbon emission, i is the type of energy, ${EC}_i$ is the energy carbon emission coefficient, and $N_i$ is the energy conversion coefficient of standard coal. Carbon emissions were measured from eight types of energy sources, such as raw coal, coke, diesel and natural gas, which are mainly involved in the industrial production process, and the specific types of energy sources, as well as their carbon emission coefficients and standard coal conversion coefficients, are shown in

Table 1. Carbon emission factor and energy standard coal conversion factor.

Type of Energy	Standard Coal Conversion Factor	Carbon Emission Factor
raw coal	0.7143	0.7559
coking coal	0.9714	0.8550
crude oil	1.4286	0.5857
petrol	1.4714	0.5538
gasoline	1.4714	0.5714
diesel	1.4571	0.5912
fuel oil	1.4286	0.6185
petroleum	1.3300	0.4483

Note: The conversion factor for standard coal from natural gas is in kg of standard coal per cubic meter. The unit of the carbon emission factor is tons of carbon per ton of standard coal, and the conversion factor for standard coal from other energy sources is kg of standard coal per kilogram.

.

4.2.2. Main Explanatory Factor

The main explanatory variable of this paper is the level of AI development; in view of the limitations of the single-indicator method, this paper draws on the theory of technological innovation evaluation and the research of Lv Rongjie [10], and selects the four dimensions of institutional environment, infrastructure, technological innovation and production and application to construct the indicator system of AI development level as shown in

Table 2. Indicator system for evaluating the level of development of artificial intelligence.

Primary Indicator	Secondary Indicator	Description
Institutional environment	Strength of financial support	Strength of government attention
	Government expenditure on science and technology	Number of mentions of intelligence, robotics, and “Internet+” in government work reports
Infrastructure	Software base	Software product revenue
	Hardware base	Fiber optic cable line length/provincial area
	Information technology service base	Information technology service revenue
	Data supply base	Number of mobile Internet users
Technological Innovation	Talent Scale	R&D personnel in high-tech industries above scale
	Material conditions for innovation	Internal expenditure on R&D funding for high-tech industries
	Innovation efficiency	Technology market turnover
	Innovation results	Number of artificial intelligence patents
Production Applications	Production in intelligent manufacturing enterprises	Gross industrial output value of high-tech industries
	Artificial intelligence enterprises	Number of artificial intelligence enterprises
	Artificial intelligence applications	Robot installation density

, and uses entropy power method to measure the level of AI development of China’s provinces and regions.

4.2.3. Mediating Variables

The industrial structure reflects the interconnection and relative importance of various industries within an economy, and is an important indicator of the degree of regional development. Artificial intelligence technology achieves industrial carbon emission reduction by adjusting the industrial structure and rationally scheduling resources. This paper evaluates the industrial framework in terms of two aspects of the industrial structure—advanced industrial structure and rationalization.

Advanced industrial structure, also known as industrial structure upgrading, refers to the process of transformation of the industrial structure system from the low-level form to the high-level form, which generally follows the law of industrial structure evolution from the low-level to the high-level evolution. The advanced industrial structure is represented by the industrial structure hierarchy coefficient. It illustrates the evolution of the three primary industries quantitatively through changes in their share ratios, calculated using a specific formula.

$$TS={\sum }_{m=1}^3{y}_{imt}\times m,m=\mathrm{1,2},3$$

(6)

In Equation (6), the variable $y_{imt}$ denotes the share of the mth industry within the $i$ region relative to the GDP during period $t$. This metric illustrates the transition among China’s three primary sectors, shifting from the prevalence of primary industry to the leading roles of secondary and tertiary industries. It serves as an indicator of a sophisticated industrial framework.

Rationalization of the industrial structure refers to an industrial structure that is coordinated among industries, has a strong ability to transform the industrial structure and good adaptability, can adapt to changes in market demand and brings the best benefits, which is specifically manifested in the process of the quantitative and proportional relationship, economic and technological linkages and interactions among industries tending to be coordinated and balanced. In order to measure the degree of rationalization of industrial structure, we adopted the Tyrell index, which is calculated as follows:

$$TL=\sum _{i=1}^n\left(Y_i/Y\right)ln\left[\left(Y_i/L_i\right)/\left(Y/L\right)\right]$$

(7)

In the formula, $Y$ and $L$ indicate the outcome value and the quantity of employed people, respectively. Where $n$ is the quantity of different industries and $i$ is the industry. When $TL$ is 0, it means that the economy of the region has reached equilibrium and its industrial organization makes sense; when $TL$ is not 0, it means that the region’s industrial structure is out of equilibrium and its industrial structure is in an unreasonable state.

4.2.4. Threshold Variable

Government fiscal spending serves as a critical tool for national macroeconomic regulation and oversight, significantly influencing China’s economic progress. Thus, the fiscal expenditure-to-regional GDP ratio gauges the extent of government intervention. The government employs the “visible hand” to offset the “invisible hand”, aiming to enhance resource allocation efficiency and minimize carbon emission intensity.

4.2.5. Control Variables

In order to minimize the impact on the estimation results due to the omission of important variables, this paper sets environmental regulation, degree of openness to the outside world, level of foreign investment, and urbanization rate as control variables on the basis of the established literature.

Environmental regulation (ER): The proportion of investment in industrial pollution control relative to the industry’s added value serves as a metric for environmental regulation. To adopt the new green development paradigm and tackle the challenge of harmonious coexistence between humans and nature, the level of environmental regulation in China has intensified. Enterprises are incentivized to adopt low-carbon clean technologies, minimize resource waste, and enhance economic and social returns through technological innovation and product enhancement. However, this can also lead to higher product costs and increased carbon emissions.

International trade (IT): The total value of imports and exports as a percentage of GDP indicates the degree of openness to global markets. The volume of goods traded can significantly influence the overall amount and intensity of regional carbon emissions. While it may raise carbon emission levels, as China could import high-pollution industries, it may also facilitate the transfer of production processes and technology dissemination, potentially lowering carbon emissions regionally.

Foreign direct investment (FDI): Evaluated by the ratio of actual FDI utilization to regional GDP, the connection between FDI and carbon emissions remains contentious. Some researchers argue that FDI causes “pollution haven” effects and “carbon leakage”. Conversely, others suggest that FDI often brings advanced technologies and management expertise that could support the international transfer of green technologies and the adoption of clean, low-carbon solutions in China, thereby aiding in carbon emissions reduction.

Urbanization rate (UR): Measured by the share of the urban population to the total population, urbanization presents dual impacts. As urbanization progresses, populations concentrate in cities, leading to increased energy consumption and emissions due to heightened urban infrastructure and development. Yet, urbanization might also contribute to carbon emission reduction by enhancing energy efficiency and diversifying energy sources.

4.2.6. Data Sources and Description

This study aims to analyze 30 provinces and cities in mainland China from 2013 to 2021, based on the availability and validity of data (Tibet, due to incomplete information, is therefore not included in the statistics). Among them, the data for measuring AI technology come from Patenthub’s World Patent Database, and the data related to carbon emissions in the industrial sector come from China Statistical Yearbook, China Energy Statistical Yearbook, China Industrial Statistical Yearbook and the EPS database. In order to unify the scale, the data of some indicators are processed by the logarithmic method. The descriptive statistics of each variable are shown in

Table 3. Descriptive statistics for variables.

Type	Variable	Variable Name	Sample Size	Mean Value	Standard Deviation	Maximum Value	Minimum Value
Explained Variable	CI	Industrial Carbon Intensity	270	0.5060	0.4057	1.9972	0.0034
Core Explanatory Variable	AI	Artificial Intelligence Development	270	0.0905	0.1132	0.7477	0.0056
Mediating variables	TS	Advanced Industrial Structure	270	2.4119	0.1182	2.8343	2.1323
	TL	Rationalization of Industrial Structure	270	0.1402	0.0848	0.4036	0.0082
Threshold variable	GI	Government Intervention	270	0.0034	0.0037	0.0310	0.0001
Control Variables	ER	Environmental Regulation	270	0.0178	0.0140	0.0796	0.0001
	FDI	Foreign Investment Level	270	0.2529	0.2643	1.3418	0.0076
	IT	International Trade	270	0.6074	0.1163	0.9377	0.3789
	UR	Urbanization Rate	270	0.2522	0.1023	0.6430	0.1066

.

5. Empirical Results and Interpretations

5.1. Evaluation of the Fixed-Effect Baseline Regression’s Results

The fixed-effect model is suitable for solving the problem of temporal and individual correlation in the data, and effectively solves the endogeneity problem due to the omission of variables, etc. Through the study of the change rule of China’s industrial carbon emission intensity, the model further explores whether the development of artificial intelligence can effectively reduce the intensity of carbon emission in China’s industrial field. This study uses Stata 17.0 to conduct an empirical analysis of model (1). The regression findings presented in

Table 4. The baseline regression’s outcomes.

	(1)	(2)	(3)	(4)	(5)
	lnCI	lnCI	lnCI	lnCI	lnCI
AI	−1.548 ***	−1.291 ***	−0.701 ***	−0.726 ***	−0.643 **
	(0.228)	(0.229)	(0.253)	(0.253)	(0.252)
ER		−16.742 ***	−17.159 ***	−17.134 ***	−16.440 ***
		(4.126)	(3.951)	(3.938)	(3.902)
IT			0.861 ***	0.847 ***	0.739 ***
			(0.184)	(0.183)	(0.186)
FDI				−2.016	−3.321 **
				(1.254)	(1.344)
UR					1.578 **
					(0.628)
Constant	−0.609 ***	−0.533 ***	−0.817 ***	−0.767 ***	−1.591 ***
	(0.029)	(0.034)	(0.069)	(0.075)	(0.336)
Fixed provinces	yes	yes	yes	yes	yes
Fixed time	yes	yes	yes	yes	yes
N	270	270	270	270	270
$R^2$	0.293	0.190	0.050	0.032	0.095

Note: **, *** denote significance levels of 5% and 1%, respectively, with t-values in parentheses.

reveal a significant negative coefficient for AI at the 1% level; for every 1 unit increase in the level of AI development, the carbon emissions generated by an increase of 1 unit of GDP decrease by 0.643 units; according to Equation (5) the total carbon emissions are associated with energy consumption due to the different standard coal conversion factors and carbon emission factors of different energy sources. Thus, we assume that the reduction in each type of energy consumption is uniform, according to the average emission factor calculations, which lead to us to conclude that when the carbon emissions decreased by 1%, the consumption of energy decreased by 0.56%. This suggests that advancements in AI development optimize resource allocation and energy efficiency across regions, leading to reduced energy consumption and lower carbon emission intensity. This result initially supports hypothesis H1. Models (2)–(5) maintain fixed years and regions while incrementally adding control variables. The regression results remain highly significant. The control variable coefficients indicate that environmental regulation positively influences economic and social benefits, contributing significantly to sustainable development while reversing the role of carbon emission intensity. The coefficient for openness to trade is notably positive. This may stem from increased goods import and export activities, which raise total carbon emissions in China. Additionally, highly polluting industries have shifted from developed nations to China, designating it as a pollution haven. The urbanization rate, reflecting social development, promotes population mobility and integration. However, it simultaneously escalates energy demands, resulting in higher carbon dioxide emissions. Foreign investment levels facilitate the adoption of energy-efficient and emission-reducing technologies, thereby suppressing industrial carbon emission intensity. The findings confirm the validity of hypothesis H1, underscoring the robustness of the article’s conclusions.

5.2. Robustness Check

Although the dual fixed-effect model is used earlier, it is suitable for addressing issues of time and individual correlation in the data, effectively resolving endogeneity problems caused by omitted variables. However, in order to verify the validity of the benchmark regression results, this paper conducts a series of robustness tests, such as re-selecting samples, replacing control variables and replacing explanatory variables.

5.2.1. Re-Selection of Samples

Taking into account that the four municipalities of Beijing, Tianjin, Shanghai, and Chongqing are directly under the control of the central government, these have more obvious resource advantages and are in the leading position in the development of AI, which may have an impact on the research conclusions. Therefore, we excluded the data of the above four municipalities and re-regressed the remaining 234 samples, and

Table 5. Robustness test results.

	(1)	(2)	(3)
	lnCI	lnCI	lnCI
AI	−0.500 **	−0.810 ***	−0.125 ***
	(0.240)	(0.249)	(0.000)
ER	−18.251 ***	—	−16.422 ***
	(3.483)	—	(3.788)
ER’	—	1.503 ***	—
	—	(0.425)	—
IT	−0.037	0.558 ***	0.576 ***
	(0.259)	(0.193)	(0.193)
FDI	−2.012	−4.054 ***	−3.432 ***
	(1.503)	(1.367)	(1.324)
UR	−0.731	2.528 ***	1.484 **
	(0.867)	(0.669)	(0.619)
Constant	0.039	−2.484 ***	−1.508 ***
	(0.440)	(0.395)	(0.331)
Fixed provinces	yes	yes	yes
Fixed time	yes	yes	yes
N	234	270	270
$R^2$	0.137	0.018	0.059

Note: **, *** denote significance levels of 5% and 1%, respectively, with t-values in parentheses.

‘s Column (1) displays the outcomes. The findings show that AI technology lowers industry-wide carbon emissions, with the coefficient of AI being significantly negative at the 5% level. This is consistent with the findings of the baseline regression and implies that the research conclusions are strong.

5.2.2. Replacement of Control Variables

The regression is re-expanded for the new variables by measuring environmental regulation in the control variables by government fiscal expenditure, and the results are shown in Column (2) of Table 5. It can be found that the coefficient of AI is significantly negative, suggesting that AI technology can reduce carbon intensity, which is consistent with the results of the baseline regression, and again suggests that the findings are robust.

5.2.3. Replacement of Explanatory Variables

The AI patent data serve as an explanatory variable for conducting the robustness test. These data were obtained from the State Intellectual Property Office, based on the AI patent classification number found in the ‘Strategic Emerging Industries Classification and International Patent Classification Reference Relationship Table (2021)’. Column (3) of Table 5 indicates that the AI coefficient remains negative at the 1% significance level. This suggests that advancements in AI development continue to contribute to reductions in carbon emission intensity, thereby affirming the robustness of the earlier empirical findings.

5.3. Heterogeneity Analysis

5.3.1. Analysis of Temporal Heterogeneity

AI’s inclusion in the ‘13th Five-Year Plan’ in 2016 marked a significant development. Since then, various governmental bodies, including the state, the Development and Reform Commission, the Ministry of Industry and Information Technology, and the Ministry of Science and Technology, have introduced several plans and initiatives for advancing artificial intelligence. To ensure the rigor of this article, the year 2016 is taken as the time point, and the time is divided into two parts, 2013–2015 and 2016–2021, which, in turn, explores the temporal heterogeneity.

Table 6. Heterogeneity test results.

	(1)	(2)	(3)	(4)
	lnCI	lnCI	lnCI	lnCI
AI	−0.041	−1.022 ***	−0.542	−1.484 **
	(1.417)	(0.263)	(0.389)	(0.694)
Constant	−0.905	−0.910 **	−3.311	0.605
	(0.810)	(0.401)	(0.576)	(0.631)
Controls	yes	yes	yes	yes
Fixed provinces	yes	yes	yes	yes
Fixed time	yes	yes	yes	yes
N	90	180	99	171
$R^2$	0.568	0.231	0.112	0.069

Note: **, *** denote significance levels of 5% and 1%, respectively, with t-values in parentheses.

, Column (1), displays the regression outcomes for 2013–2015, revealing insignificant coefficients. Conversely, Column (2) presents the regression results for 2016–2021, showing significantly negative coefficients. This indicates that the influence of AI technology on carbon emission intensity post-2016 substantially diverges from that of 2013–2015, displaying a more inhibitory effect on carbon emission intensity. To examine potential reasons, first, national policy incentives have driven significant advancements in science and technology since 2016. This progress enhances energy efficiency and lowers energy consumption per production unit, leading to reduced carbon emissions. Second, AI catalyzes industry structural changes, facilitating an economic shift away from traditional high-carbon industries towards low-carbon or carbon-neutral alternatives. Lastly, AI streamlines production processes, improving resource utilization and minimizing energy consumption.

5.3.2. Analysis of Regional Heterogeneity

China is a large nation with significant disparities in industrial development across its regions. Variations also exist in the economic environment, energy resources, and technological capabilities. These differences may lead to regional variability in how AI development affects industrial carbon emission intensity. This study categorizes regions into eastern and central/western for econometric analysis, with the results summarized in Table 6. The coefficient measuring AI’s impact on carbon emission intensity in the eastern region (Column 3) failed to pass the significance tests. In contrast, carbon emission intensity in central and western regions (Column 4) is more influenced by artificial intelligence technology; for every unit of artificial intelligence technology, an increase of 1 unit of GDP produces a reduction of 1.484 units of carbon emissions. The reason for this may be that the eastern part of the country, as a region of rapid economic development, has obvious advantages in terms of economic level, talent aggregation and technological research and development and investment, etc., and the AI infrastructure, energy structure and carbon emission management have reached a relatively mature stage; so, the role of AI in carbon emission reduction may face the problem of diminishing marginal returns. The central and western regions are in the process of economic development and industrialization, and carbon emission efficiency and management are relatively under-developed; so, AI may directly improve energy efficiency, optimize resource utilization, and reduce carbon emissions in this context. In addition, the industrialization process in the Midwest relies on coal and other high-carbon emission energy sources, and in the AI infrastructure, such as smart grid, smart transportation, etc., which is not perfect, and AI has great potential in these aspects, so the effect of AI on carbon reduction in the Midwest is more significant.

5.4. Mediating Effect Tests

The findings from the benchmark regression illustrate that the advancement of AI can substantially diminish carbon emission intensity. To further investigate the precise mechanism through which AI influences industrial carbon emission intensity, testing hypotheses H2 and H3, this paper incorporates two mediating variables, industrial structure advanced and industrial structure rationalization, into the benchmark regression of model (1) for analysis. As observed in Columns (1) and (2) of

Table 7. Intermediation effect regression results.

	(1)	(2)	(3)	(4)
	ln(TS)	lnCI	ln(TL)	lnCI
AI	0.032 **	−2.349 ***	−1.265 ***	−2.163 ***
	(0.013)	(0.427)	(0.315)	(0.448)
ln(TS)		−9.072 ***
		(1.341)
ln(TL)				0.420 ***
				(0.079)
constant	0.794 ***	6.088 ***	−0.375	−0.396
	(0.017)	(1.013)	(0.420)	(0.263)
Controls	yes	yes	yes	yes
	yes	yes	yes	yes
Fixed time	yes	yes	yes	yes
Sobel test value	−0.371 ***		−0.557 ***
(p-value)	(0.000)		(0.000)
N	270	270	270	270
$R^2$	0.684	0.671	0.646	0.651

Note: **, *** denote significance levels of 5% and 1%, respectively, with t-values in parentheses.

, when the advancement of the industrial structure serves as the mediating variable, the coefficient for AI in Column (1) is significantly positive at the 5% level. In Column (2), both AI and the advancement of the industrial structure exhibit negative coefficients at the 1% significance level. This indicates that AI mitigates the intensity of industrial carbon emissions by encouraging the advancement of industrial structure. It suggests that the advancement of the industrial structure may obscure this relationship, which does not entirely align with Hypothesis 2, and could imply either a suppressive effect on industrial carbon emissions or an inflated role of AI. This might occur because the process of advancing the industrial structure also produces significant carbon emissions. In Columns (3) and (4) of Table 7, when the rationalization of the industrial structure is examined as the mediating variable, the coefficient for AI in Column (3) shows a significant negative correlation at the 1% level, indicating a notable mediating effect between AI and the rationalization of industrial structure. The positive correlation between the rationalization of industrial structure and industrial carbon emission intensity in Column (4) arises from the application of the Terrell index. When the Terrell index equals 0, the industrial structure is rational; otherwise, it is deemed irrational. This highlights that AI facilitates the reduction in carbon emission intensity by aiding in the rationalization of industrial structure, i.e., confirming the validity of hypothesis H3.

5.5. Threshold Effect Tests

To determine whether government intervention will cause a non-linear reduction in industrial carbon intensity through AI and to test hypothesis H4, we utilized the software Stata 17 to take 300 random samples using the bootstrap method. The empirical findings are as follows.

From

Table 8. Threshold estimation results.

Threshold Type	Threshold Value	F-Value	p-Value	Threshold Value
				10%	5%	1%
Single threshold	0.168	78.44	0.000	20.814	23.753	32.664
Double threshold	0.240	16.27	0.130	18.022	22.260	33.452

, we observe that the F-value for the single threshold is significant at the 1% level, allowing us to reject the null hypothesis. Additionally, the F-statistic for the double threshold exceeded the 10% threshold in the double-threshold test, suggesting that the double-threshold hypothesis does not hold. Consequently, we opted for the single-threshold regression model to evaluate the non-linear impact of AI on industrial carbon emission intensity.

The results in

Table 9. Panel threshold regression results.

Variable	Coefficient	Confidence Interval
	CI
AI (GI < threshold)	−1.677 ***	[−2.144, −1.209]
	(0.237)
AI (GI > threshold)	−3.998 ***	[−4.649, −3.347]
	(0.330)
threshold value	0.168	[0.166, 0.173]

Note: *** denote significance levels of 1%, respectively, with t-values in parentheses.

show that the inhibitory effect of AI on industrial carbon emission intensity is constrained by the threshold variable of government intervention, and the effect of government intervention on industrial carbon emission intensity is significantly negative when the government intervention fails to pass the threshold value of 0.168, and the low level of government intervention, due to the weaker support for innovation or the relaxation of ecological and environmental policy regulations, makes the effect of AI on the reduction in carbon emission intensity weaker. The effect of AI on the reduction in industrial carbon intensity increases significantly after the level of government intervention crosses the threshold. This empirical result verifies that Hypothesis H4 is valid.

6. Conclusions

Artificial intelligence (AI) serves as the primary catalyst for the latest wave of scientific and technological advancements, along with industrial transformations. It plays a pivotal role in achieving carbon emission reductions and reaching the goals of ‘peak carbon’ and carbon neutrality. This article utilizes provincial panel data from 2013 to 2021 in mainland China to construct an evaluation index system for AI development. It employs a double fixed-effect model, a mediated-effect model, and a threshold effect model to conduct theoretical analyses and empirical research on how AI influences industrial carbon emissions and the underlying mechanisms.

The findings of this study reveal several key points: First, the baseline regression analysis indicates that AI significantly impacts industrial carbon emissions. This result suggests that advancements in AI contribute to enhanced carbon emission reduction within the industrial sector. These conclusions remain valid after robust testing, including the substitution of explanatory variables, sample re-selection, and variations in control variables. Second, research into the mechanisms through which AI operates reveals its crucial role in rationalizing industrial structure, thus facilitating reductions in industrial carbon emissions. Although an advanced industrial structure does contribute to AI’s role in emission reductions, it also exhibits a masking effect, which diverges from the original hypothesis. Third, the influence of AI on the intensity of industrial carbon emissions is found to be non-linear, contingent upon the level of government intervention; lower levels of intervention weaken AI’s effectiveness in reducing carbon emission intensity. Fourth, regarding temporal heterogeneity, AI’s impact on carbon reduction becomes more pronounced post 2016. Lastly, the western and central regions experience a greater impact from human intelligence technology compared to the eastern regions in terms of carbon emission intensity.

On this basis, the following countermeasures are proposed:

First, we should increase research and development on artificial intelligence and expand application scenarios. We should make more efforts to research AI, especially to make breakthroughs in ‘neck-breaking’ technologies. The deep integration of AI technology and manufacturing technology can transform traditional crude industries, eliminate backward production capacity, and realize the ‘green’ development of industries. At the same time, we should speed up the promotion of cleaner production processes in industry, and carry out the ‘marginal production’ type of technological transformation in order to achieve emission reduction at the source and improve the tail gas treatment capacity of enterprises.

Secondly, we must fully leverage the role of science and technology as a dual constraint on economic growth. Scientific and technological advancements can enhance the carbon emission efficiency of industries. They can also offer sustainable pathways and mechanisms for the green development of artificial intelligence. China must urgently reinforce the connection and collaboration between industrial chain development and AI technological advancements. This effort will foster the synergistic development of industry and intelligent ecosystems, establishing a robust industrial foundation for achieving low-carbon emission reductions.

Thirdly, we need to enhance policy support. Building on the study of temporal heterogeneity, we propose policy recommendations that aim to accelerate China’s economic progress. To harness the benefits of AI, various regions in China should capitalize on their unique industrial policy advantages. Policies must guide and direct the development of AI effectively. It is crucial to emphasize AI’s role in optimizing industrial structures to bolster industry standards and strengthen the position of the industrial chain. This focus should promote the upgrading of industrial structures and prioritize the development of strategic emerging industries, ultimately cultivating China’s competitive edge in science and technology.

Fourth, it is essential to implement differentiated policies tailored to local circumstances. The varying levels of development, industrial structures, and energy consumption across regions lead to different impacts of AI technology on carbon emission reductions. Therefore, formulating targeted policies based on each region’s comparative advantages is necessary. It is vital to define the applicable realms of artificial intelligence scientifically. In the central region, heavily influenced by the ‘carbon lock-in’ effect, it is important to avoid ‘catch-up’ strategies and expedite the transformation of energy consumption patterns. In the eastern region, recognized as the ‘leader’, there is an urgent need to leverage emerging technologies like artificial intelligence to achieve low-energy consumption. The eastern region must actively pursue the transformation and upgrading towards low-carbonization, digitization, and intelligence to realize a green transformation of industries. In this framework, there is a pressing need to utilize artificial intelligence to facilitate the greening and the development of intelligence of traditional sectors, and improve the green performance evaluation systems for intelligent manufacturing enterprises in line with major strategies like ‘East Counts, West Counts’.

While this study has made significant progress in examining the impact of AI technology on industrial carbon emissions in China, it is not without its limitations. Firstly, due to the geographic, economic and cultural differences between Tibet and the rest of China, obtaining accurate data remains challenging, and the lack of data from Tibet may have some impact on the applicability of this study’s findings, especially when as relates to indicators with distinct regional characteristics. Second, the focus of this study is limited to the role of AI technology in influencing carbon emissions in the industrial sector in domestic provinces and cities, limiting the external validity and generalizability of this study’s conclusions to other countries and other industries and subsectors under the industrial sector. Once again, the combined effect between AI and control variables will be refined with moderating-effect interaction terms in future studies due to space limitations. Finally, in addition to the government, other institutions such as universities, trade associations, and other organizations play an important role in the development of AI in terms of setting standards and providing guidance, which we did not study due to space constraints.

In future research, we will verify the generalizability of the findings after obtaining Tibetan data. In addition, we will study the effect of AI under industrial segments in other industries and at the microlevel, and expand our focus to the global level to analyze the effect of AI on carbon emission reduction in various countries, which will provide richer theoretical support for alleviating the global warming crisis. Furthermore, we will focus on other emerging technologies. Finally, we will look at the synergy between other emerging technologies and AI technology to find the best path for synergistic emission reduction, and discuss the influence of other organizations on the development of AI.