Artificial Intelligence and Carbon Emissions: Mediating Role of Energy Efficiency, Factor Market Allocation and Industrial Structure

Abstract

Artificial intelligence (AI) plays an important role in promoting energy transformation and achieving global green and low-carbon goals. Based on the panel data of 285 prefecture-level cities in China from 2011 to 2022, this paper empirically examines the impact of AI on carbon emission (CE) and its internal mechanism. It is found that the impact of AI on CE in general shows an “inverted U-shaped” relationship, which is first promoted and then suppressed, and this result still holds after a series of robustness tests. The mechanism test shows that AI affects CE in three main ways: improving energy efficiency, optimizing factor market allocation, and industrial structure. The heterogeneity results show that the “inverted U-shape” relationship of AI on CE is significant in resource cities insignificant in non-resource cities, significant in low-carbon pilot cities, and insignificant in non-low-carbon pilot cities, significant in areas with a high level of industrialization, and insignificant in areas with a low level of industrialization. This study provides valuable insights for the application of AI and the formulation of energy conservation and emission reduction policies.

1. Introduction

Relying on comparative advantages such as the resource dividend and resource skew, China has swiftly integrated itself into the global value chain system, and the process of urbanization and industrialization has been accelerating [1]. However, due to the past strategy of prioritizing the advancement of heavy industry and relatively lax environmental regulations, China’s energy consumption and CE have been steadily increasing, driving the ecological environment toward a critical threshold. Given the escalating challenges of energy and environmental constraints, China has advocated accelerating the shift to a green development model and steadily advancing carbon peaking and carbon neutrality goals. Therefore, shifting away from the high-speed, resource-intensive growth model and exploring a more inclusive development approach that achieves a balance between environmental conservation and high-quality economic development is a strategic imperative to support China’s sustainable development in the new era.

Simultaneously, AI is a pivotal technology driving a new wave of technological and industrial revolutions. AI plays a crucial role in fostering new industries, upgrading traditional ones, optimizing industrial structures, and driving productivity growth, and can provide new chances for controlling CO₂ emissions. However, one of the major challenges that AI must face in reducing CO₂ emissions is the rebound effect, also referred to as “Jefon’s paradox” [2]. On the one hand, AI, as a general-purpose technology, can be applied in the fields of healthcare, transportation, energy, communication, and electricity, which can help improve the effectiveness of work in these fields and reduce CE; on the other hand, the development, training, and application of AI itself requires a large amount of digital infrastructure and consumes a large amount of energy. For example, the carbon footprint of training a single large language model is predicted to be up to 300,000 kg of carbon dioxide equivalent [3]. Therefore, whether AI can be an effective path for China to achieve the “dual-carbon” goal needs to be further verified.

The unique contribution of this paper is as follows: First, it reveals the intrinsic mechanism of AI affecting CE through improving energy efficiency and optimizing factor market allocation and industrial structure. Second, based on the panel data of 285 prefectural-level cities in China from 2011–2022, it empirically examines and finds that there exists an “inverted U-shaped” relationship between AI and CE. Third, it explores the heterogeneity of AI affecting CE from the angles of resource-based cities and low-carbon pilot cities. And the results of the study show that the “inverted U-shaped” impact of AI on CE is significant in resource-based cities, not significant in non-resource-based cities, significant in low-carbon pilot cities, and not significant in non-low-carbon pilot cities.

2. Literature Review

Studies on the influence factors of CE intensity primarily focus on energy consumption structure technological advancement and industrial structure. First of all, regarding the impact of energy efficiency, established studies generally agree that energy efficiency and CE intensity show a positive correlation. Energy consumption CE is the main source of CE, and the enhancement of energy efficiency has a significant inhibitory effect on CE [4]. Zhu and Han [5] employed the DEA method to evaluate energy efficiency and discovered that the enhancement of energy efficiency is a key factor leading to the reduction in CE intensity, and the research of Fisher [6] likewise confirms this point. Therefore, some regions have taken the improvement of energy efficiency as a key strategy for CE reduction [7]. Second, regarding the influence of technological progress, Waris et al. [8] analyzed the evolving relationship between innovation in renewable energy patents and CO₂ emissions in ASEAN countries; the findings reveal that the increase in renewable energy-related technology patents significantly leads to increased emissions. However, some scholars argue that technological advancement carries a “double-edged effect”, which may increase CO₂ emissions, and the reduction effect may be halved or rebound directly. Because technological innovation improves energy efficiency to a certain extent, there will be greater energy demand in the market, and if the increase is greater than the part of energy cuts and savings, it may ultimately cause an increase in overall consumption [9]. Again, regarding the impact of industrial structure, Wang et al. [10] found that industrial structure is the influence factor that leads to the change of CE by comparing the opposite trend of CE in China and India. China’s industrial structure is gradually transforming into the service industry and high-end manufacturing industry, so the CE is slowly decreasing, while India’s economy is still dominated by high-energy-consuming heavy industry, and the CE is continuing to increase.

As the smart economy era begins, the green development effect of AI has attracted considerable interest in the academic community, prompting a large number of studies on its impact on CE, the study of AI on CE, mainly from the perspective of direct and indirect impact. Regarding the direct impact effect, Elhenawy et al. [11] found that the application of AI technology can enhance the productivity and impact of the CO₂ capture process. Altintas et al. [12] found that AI can assist in the realization of carbon capture through the application of pre-combustion carbon capture. Liu et al. [13] incorporated population, affluence, and technology into the STIRPAT model and discovered that AI has a reducing effect on carbon intensity, and this effect has obvious stage characteristics, while Wang et al. [14] used panel data from 67 countries and regions to show that AI significantly reduces carbon emissions, and this effect increases with the increase in trade openness. In addition, thanks to the advancement of AI sensors and data processing technology, its influencing pathway on CE is more diversified. So, there are also studies indirectly proving the influence of AI on the CE intensity of enterprises. As for the impact of the components of AI on CE, scholars mainly focus on the specific applications of AI such as industrial robots, industrial intelligence, and intelligent manufacturing, and study AI’s impact on CE intensity. Li et al. [15] discovered that the use of industrial robots can improve productivity and optimize the factor structure, and the study by Wang et al. [16] shows that the advancement of industrial intelligence can significantly lower the carbon intensity of both local and neighboring regions, while the study by Tang et al. [17] finds that smart manufacturing is essential in promoting CE reduction, and that government intervention enhances this effect.

However, the advancement of AI technology can indeed drive changes in carbon emission levels, but its energy-intensive nature means that the application and development of AI may bring additional energy needs [18]. Especially when training deep learning models and running large-scale data centers, the computing power of AI consumes a lot of power resources [19]. These power needs, if they rely primarily on traditional fossil fuels, can lead to increased carbon emissions. However, in the long term, AI also has the potential to indirectly reduce carbon emissions by improving energy efficiency, optimizing production processes, and promoting innovation in green technologies. Thus, the relationship between AI and carbon emissions has a duality.

3. Mechanistic Analysis of the Impact of Artificial Intelligence on Carbon Emissions

3.1. Artificial Intelligence Lowers Carbon Emissions by Enhancing Energy Efficiency

AI improves energy efficiency through automated production, waste reduction, and green technology innovation, thereby reducing CE. Firstly, AI has the effect of technological advancement and upstream and downstream correlation [20], which can improve energy efficiency via automated production. The application of automation technology can finely regulate the temperature, humidity, speed, and other parameters in the production process, so that each link can achieve optimal efficiency with the lowest energy consumption and avoid excessive energy consumption, which effectively reduces reliance on fossil fuels and stimulates the extensive adoption of clean, renewable energy, reducing corporate CE. Second, AI enhances energy efficiency by reducing waste. AI, combined with the Internet of Things and sensors, can monitor energy consumption and resource use during the production process in real time, precisely analyze energy efficiency and identify energy waste links [21], and based on data analysis, provide optimization suggestions to enhance the production process and help enterprises achieve the goal of low-carbon development. Finally, AI improves energy efficiency and reduces CE through innovation in green technology. Enterprises can integrate AI technology into their operations to boost green technology innovation, and enterprises apply AI to upgrade energy-intensive equipment to make it more high-quality, high-efficiency, low-consumption, safe, and reliable, to achieve effects of energy saving and emission reduction, enhance energy utilization efficiency, and thus realize energy saving and emission reduction.

3.2. Artificial Intelligence Reduces Carbon Emissions by Optimizing Factor Allocation

AI reduces CE by optimizing labor and capital allocation. On the one hand, AI replaces part of the human labor in high-energy-consuming industries with high CE, reduces the reliance of economic development on high-carbon industries, and fundamentally optimizes the effectiveness of resource allocation. With the extensive use of AI automation technology, the production process of high-energy-consuming industries tends to be intelligent and precise, and this technological substitution not only enhances the production efficiency, but also helps to weaken the traditional industry’s dependence on the carbon-emission-intensive labor mode [22], which in turn reduces CE. On the other hand, because AI itself has low-pollution and knowledge-intensive attributes, its widespread application can fuel the expansion of the green economy and knowledge-driven industries, guide the transfer of capital to the green economy and knowledge-driven industries, improve the capital allocation efficiency in the factor market [23], and make the capital flow faster and more accurately to green technology innovation and low-carbon industries. In addition, financial technology driven by AI improves the transparency and liquidity of the capital market, enhances the accuracy of risk assessment, makes capital flows more aligned with the path of green development, shortens the return cycle of investment in renewable energy technologies, carbon capture technologies, and smart grids, and thus indirectly reduces the CE intensity of economic activities.

3.3. AI Reduces Carbon Emissions by Optimising Industrial Structure

AI optimizes the industrial framework and reduces CE by boosting industrial upgrading, playing the pulling role of consumption, and achieving coordinated regional development. First of all, AI technology, represented by big data, cloud computing, and the Internet of Things, can change the traditional high-pollution, high-emission, and rough production mode, build a low-risk, low-input whole-process composite innovation model [24], form a personalized, green enterprise production chain, and empower enterprises, especially those in high-energy-consuming industries, to achieve industrial transformation and upgrading, to reduce CE. Secondly, with the enhancement of residents’ living standards and environmental awareness, resource- and environment-friendly products have gradually become the mainstream of the consumer market. The use of AI can meet the residents’ demand for new products and new forms of business, and these demands will also produce a forcing effect, gradually eliminating high-pollution, high-emission rough production enterprises and optimizing the industrial structure in a benign interaction. Finally, AI also helps promote the green and coordinated growth of the regional economy. Drawing on big data analysis and regional economic models, AI can accurately identify the industrial characteristics and carbon emission characteristics of different regions, providing strong support for local governments to formulate industrial structure adjustment and green development policies [25], thus promoting CE reduction.

In addition, the impact of AI on CE has stage differences and may have a nonlinear relationship. In the early stages of AI development, due to the need for a large amount of digital infrastructure, there is a high consumption of energy, such as electricity, and at this time, the large-scale application of AI will offset the CE reduction effect, or even exacerbate CE. However, as the development of AI technology moves into a mature stage, its application in various fields is becoming more and more widespread, especially in improving energy efficiency, optimizing resource allocation and industrial structure, etc., the positive effects of which are gradually emerging, and to a certain extent, reducing CE. Therefore, the impact of AI on CE is not a simple linear relationship but may exhibit complex nonlinear characteristics.

Based on this, this paper introduces the following hypotheses:

H1: 

AI has an “inverted U” relationship with CE.

H2: 

There are three ways AI promotes carbon reduction: by improving energy efficiency, optimizing resource allocation, and optimizing industrial structure.

4. Empirical Analysis of the Impact of Artificial Intelligence on Carbon Emissions

4.1. Modeling

STIRPAT is an expandable stochastic environmental impact assessment model based on the IPAT model, which evaluates the relationship between the three independent variables, namely population, economy, and technology, and the dependent variable. The STIRPAT model has the characteristics of good applicability, strong interpretation, and wide application in the study of the factors affecting CEs. Therefore, this paper is grounded in the STIRPAT model and draws from the work of Shahbaz et al. [26], focusing on the effect of AI on CE. The comprehensive development level of AI is selected as the core explanatory variable, and a collection of control variables is added; the improved econometric model is shown as follows:

$$ln{C}_{it}={\alpha }_0+{\alpha }_{1}ln{AI}_{it\_dis}+{\alpha }_{2}ln{AI}_{it\_dis}^2+X_{it}^{\prime }\gamma +{\mu }_{it}$$

(1)

Here, $C_{it}$ is the dependent variable, which represents the total CE of different regions in different years; ${AI}_{it\_dis}$ is the primary explanatory variable, which signifies the overall development index of AI in different regions in different years. To determine whether there is a nonlinear relationship between the explanatory variable and the core explanatory variable, the square of the comprehensive development index of AI is introduced here; $X_{it}^{\prime }$ represents the control variable.

4.2. Description of Variables and Data Sources

4.2.1. Explanatory Variables: Prefecture-Level Cities’ ${\mathrm{C}\mathrm{O}}_2$ Emissions $C_{it}$

As a key indicator, urban carbon dioxide emissions can directly reflect the actual situation of energy consumption and greenhouse gas emissions in the process of urbanization. Therefore, we use the CO₂ emissions of prefecture-level cities to measure the explained variables in this paper.

4.2.2. Core Explanatory Variable: AI Composite Development Index ${AI}_{it\_dis}$

The measurement of the AI level generally starts from two perspectives: the industry-level AI level mostly adopts a single indicator such as penetration rate and the number of industrial robots, and the regional-level AI level is mostly measured by the AI comprehensive development index. Scholars both domestically and internationally examine the development of AI in depth from different dimensions, but the measurement dimensions have not yet formed a unified cognition. This paper draws from the work of Gu and Ma [27] and the China Academy of Information and Communication Research (CAICR) on the comprehensive development index of AI to evaluate the overall regional development index of AI from the three dimensions of the AI development environment, AI technological innovation, and the growth of the AI industry. Firstly, the advancement of regional AI cannot be separated from the support of the development environment, human, material, and financial resources, and other factors of production are essential to the development environment of the AI industry; investment in R&D funds, the extent of infrastructure, practitioners, and the support of regional policies will significantly affect the level of regional AI [28]. Secondly, as AI represents skill-biased technological progress [29], the technological innovation gap is the core element for evaluating the extent of AI development, which cannot be separated from the input of AI patents, papers, enterprise R&D, and other elements. Finally, the level of AI development should be realistically characterized through industrial development, and the business income in the field of AI and the degree of penetration of AI technology are all important embodiments of the development of the regional AI industry. The specific indicators selected in this paper are shown in

Table 1. Indicator system for comprehensive development index of AI.

Primary Indicators	Secondary Indicators	Specific Indicators
AI Development Environment	R&D expenditure	Financial science and technology expenditures
	Infrastructure	Internet broadband subscribers per 100 population
	Staffing inputs	Employees in the information transmission, computer services, and the software industry
	Policy support	Frequency of intelligent keywords used in government work report
AI Technology Innovation	Enterprise R&D	Enterprise R&D investment
	Number of patents	Number of software copyright registrations
	Enterprise size	Number of AI companies
AI Industry Development	Business income	Total business per capita in software and information technology services
	Degree of penetration	AI book value/total number of employees

.

Drawing on the research method of Ma [30] and other researchers, this paper utilizes the entropy weight method to measure the comprehensive AI development index, which is a kind of objective assignment method, using the idea of entropy value to determine the significance of each index, and judging the influence of the indices based on the extent of their dispersion, and the greater the indices’ discrete degree, the greater the influence of the indices in the evaluation. The entropy weight method can avoid the defects of the principal component analysis method to a certain extent so that the discrepancy between the data is objectively reflected through the weights. The specific calculation formula is outlined as follows:

The evaluation object contains m samples, each sample possesses indicators, and $X_{ij}$ denotes the value of the jth indicator of the ith sample. Firstly, the indicators are standardized, and the relative values of the indicators are calculated to make them comparable:

$$X_{ij}^{\prime }={\displaystyle \frac{X_{ij}-\mathrm{m}\mathrm{i}\mathrm{n}\{X_{1j},X_{2j}\dots X_{mj}\}}{\mathrm{m}\mathrm{a}\mathrm{x}\{X_{1j},X_{2j}\dots X_{mj}\}-\mathrm{m}\mathrm{i}\mathrm{n}\{X_{1j},X_{2j}\dots X_{mj}\}}}$$

(2)

Afterwards, the weight assigned to the ith sample value under the jth indicator for that indicator is calculated:

$$P_{ij}{\displaystyle \frac{X_{ij}^{\prime }}{\sum _{i=1}^m{X}_{ij}^{\prime }}},j=\mathrm{1,2}\dots n$$

(3)

The entropy value of the jth indicator is computed:

$$e_j=-k{\sum }_{i=1}^m{P}_{ij}ln{P}_{ij},j=\mathrm{1,2}\dots n;k=1/ln\left(m\right)$$

(4)

The redundancy of information entropy is computed:

$$d_j=1-e_j,j=\mathrm{1,2}\dots n$$

(5)

The weights of the indicators have been determined:

$$w_j={\displaystyle \frac{d_j}{{\sum }_{j=1}^n{d}_j}}$$

(6)

$${AI}_{i\_dis}={\sum }_{j=1}^n{w}_j{X}_{ij}$$

(7)

$${AI}_{i\_pro}={\sum }_{i=1}^n{AI}_{i\_dis}$$

(8)

Finally, the AI comprehensive development index of each prefecture-level city ${AI}_{i\_dis}$ is calculated, and its values are accumulated to obtain the AI comprehensive development index of each provincial-level administrative region ${AI}_{i\_pro}$, and the greater the value is, the higher the degree of AI development in the region is.

4.2.3. Control Variables

Based on existing research, this paper integrates a range of control variables within the model, including the state of economic prosperity ${GDP}_{it}$, measured by the per capita $\mathrm{G}\mathrm{D}\mathrm{P}$ of prefecture-level cities; population size ${POPU}_{it}$, quantified by the count of resident population within the region; the level of openness to the external world ${FDI}_{it}$, assessed through the proportion of actual foreign capital utilized to $\mathrm{G}\mathrm{D}\mathrm{P}$; the level of industrialization ${IND}_{it}$, calculated as the percentage of the secondary sector to $\mathrm{G}\mathrm{D}\mathrm{P}$; the average wage ${URBAN}_{it}$, quantified by the average wage of active employees; and the logarithm of the energy mix ${LNENERGY}_{it}$ measured by the ratio of total energy consumption of 100 tons of standard coal to gross regional product of CNY 10,000.

4.2.4. Data Sources

This paper utilizes panel data from 285 prefecture-level cities in China, spanning from 2011 to 2022, and the data are mainly obtained from the China Carbon Accounting Databases (CEADs), China Urban Statistical Yearbook, China Energy Statistical Yearbook, China Industrial Statistical Yearbook, China Statistical Yearbook of Population and Employment, VCPE database, patent database, trademark database, patent database, software copyright database, etc. The keyword frequency of the governmental work report and the word frequency data of listed AI companies are obtained based on the keywords using big data crawler technology. In this paper, the missing data are filled in by an interpolation method, and simultaneously, to mitigate the impact of heteroscedasticity and magnitude, the panel data are counted; the statistical summaries of the variables are presented in the

Table 2. Descriptive statistics of variables.

	(1)	(2)	(3)	(4)	(5)
Variables	N	Mean	sd	Min	Max
${\mathrm{C}}_{\mathrm{i}\mathrm{t}}$	3420	1.695	0.948	0.145	1.895
${\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}\_\mathrm{d}\mathrm{i}\mathrm{s}}$	3420	0.119	0.283	0.051	0.172
${\mathrm{P}\mathrm{O}\mathrm{P}\mathrm{U}}_{\mathrm{i}\mathrm{t}}$	3420	5.897	0.703	2.970	8.136
${\mathrm{S}\mathrm{A}\mathrm{L}\mathrm{A}\mathrm{R}}_{\mathrm{i}\mathrm{t}}$	3420	0.110	0.371	0.851	1.231
${\mathrm{F}\mathrm{D}\mathrm{I}}_{\mathrm{i}\mathrm{t}}$	3420	0.176	0.282	0.000	2.491
${\mathrm{I}\mathrm{N}\mathrm{D}}_{\mathrm{i}\mathrm{t}}$	3420	0.429	0.144	0.446	0.844
${\mathrm{G}\mathrm{D}\mathrm{P}}_{\mathrm{i}\mathrm{t}}$	3420	5.595	3.181	0.742	25.986
${\mathrm{L}\mathrm{N}\mathrm{E}\mathrm{N}\mathrm{E}\mathrm{R}\mathrm{G}\mathrm{Y}}_{\mathrm{i}\mathrm{t}}$	3420	−5.113	.52	−7.889	−3.186
Number of cities	285	285	285	285	285

.

4.3. Empirical Tests

4.3.1. Multicollinearity Test

In this study, the variance inflation factor (VIF) is first utilized for testing whether there is multicollinearity among the variables, and the outcomes of this test are displayed in

Table 3. Multiple covariance test.

Variable	VIF	1/VIF
${AI}_{it\_dis}$	1.16	0.862
${lnenergy}_{it}$	1.87	0.535
${GDP}_{it}$	1.56	0.641
${SALAR}_{it}$	1.76	0.568
${FDI}_{it}$	1.12	0.893
${POPU}_{it}$	1.11	0.901
${IND}_{it}$	1.32	0.758
${Ene}_{it}$	1.78	0.562
${Fac}_{it}$	1.26	0.794
${Ins}_{it}$	1.28	0.782
Mean	VIF	1.42

. Table 3 reports the test results of the variance inflation factor; the average value of the variance inflation factor is 1.31, the peak value is 1.76, the minimum value is 1.11, and the variance inflation factor of all the variables has a value of less than the boundary value of 10. Therefore, all the variables included in this paper pass the variance inflation factor (VIF) test, and there is no significant multicollinearity among the factors.

4.3.2. Benchmark Regression

Table 4. Benchmark regression.

Variables	(1) FE	(2) RE	(3) GLS
${AI}_{it\_dis}$	0.025 *	0.027 *	0.183 ***
	(1.72)	(1.78)	(6.20)
${AI}_{it\_dis}$2	−0.002 **	−0.002 **	−0.008 ***
	(−2.18)	(−2.20)	(−5.53)
Constant	16.378 ***	15.154 ***	9.904 ***
	(45.79)	(44.75)	(25.97)
Control	Yes	Yes	Yes
Observations	3420	3420	3420
R-squared	0.258	0.248	0.410
Number of cities	285	285	285

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

presents the outcomes of the baseline regression tests. In particular, Column (1) displays the regression outcomes for the fixed effect (FE), column (2) displays the regression results of the random effect (RE), and column (3) shows the regression outcomes of generalized least squares (GLS). In light of the Hausman test results, from the data presented, it is evident that the Hausman test p = 0.000 < 0.01; therefore, drawing on the analysis, this paper opts for the fixed-effect model for the benchmark regression test. On this basis, to remove the problems of heteroskedasticity and serial correlation within the model, this paper adopts the generalized least squares (GLS) method, which transforms the original model and thus eliminates heteroskedasticity and autocorrelation, to estimate the model. Referring to the test results in Column (3), although the coefficient of ${AI}_{it\_dis}$ displays a statistically significant positive connection, which indicates that there is a positive effect of AI on CE, the coefficient for the quadratic term of ${AI}_{it\_dis}$ shows a significant negative effect, which indicates the presence of a statistically significant nonlinear relationship between AI and CE, and it is an inverted “U”-type relationship.

4.3.3. Heterogeneity Test

①

The heterogeneity of urban resource endowments

Subgroup 1: Resource-based cities, with the extraction and utilization of natural resources like minerals and forests as their leading industry, have led to increasing resource constraints in the development process and are therefore in urgent need of restructuring and modernization. According to the National Development and Reform Commission (NDRC) list of National Sustainable Development Plan for Resource-based Cities, cities are distinguished into resource-based cities and non-resource-based cities to test the heterogeneity of the influence of AI on CE.

The results are presented in

Table 5. Heterogeneity test of urban resource endowments.

	(1)	(2)
Variables	Resource-Based City	Non-Resource-Based Cities
${AI}_{it\_dis}$	0.046 **	−0.051
	(2.20)	(−1.56)
${AI}_{it\_dis}$2	−0.003 ***	0.003
	(−2.84)	(1.61)
Constant	16.862 ***	16.344 ***
	(15.98)	(32.22)
Control	Yes	Yes
Observations	1344	2028
R-squared	0.270	0.274
Number of cities	126	153

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

, and the estimated coefficients of the primary and secondary terms for resource-based cities are 0.046 and −0.003, which are significant at the 5 percent level, showing that the impact of AI on CE in resource-based cities exhibits an inverted U-shaped curve that promotes and then suppresses CE, whereas the impact on non-resource-based cities is not obvious. On the one hand, the economic structure of resource-based cities tends to be highly dependent on the extraction and processing of natural resources, such as minerals and energy, and these industries are generally characterized by high levels of energy expenditure and environmental degradation. The introduction of AI may initially accelerate the efficiency of resource exploitation and utilization, increasing production and energy intensity, thus driving up CE. With the further application of AI technology, these cities are gradually shifting to the stage of intelligent management and green transformation, where the efficiency of resource extraction is improved and pollution management is optimized, leading to the beginning of a decline in CE. On the other hand, the advancement of AI in non-resource cities mostly relies on high-tech industries and is less dependent on resources, so the effect of AI on CE in such cities is not significant.

②

Heterogeneity of low-carbon pilot

Subgroup 2: To test the heterogeneity of the impact of AI on CE, cities are distinguished into low-carbon pilot cities and non-low-carbon pilot cities based on the NDRC’s Circular on low-carbon pilot initiatives for provinces, regions, and cities.

Table 6. Heterogeneity test of low-carbon pilot cities.

	(1)	(2)
Variables	Low-Carbon Pilot Cities	Non-Low-Carbon Pilot Cities
${AI}_{it\_dis}$	0.089 ***	−0.013
	(3.33)	(−0.63)
${AI}_{it\_dis}$2	−0.005 ***	0.000
	(−3.25)	(0.22)
Constant	16.244 ***	16.552 ***
	(13.41)	(31.27)
Control	Yes	Yes
Observations	1224	2148
R-squared	0.309	0.249
Number of cities	102	179

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

displays the results, and the estimated coefficients of the primary and secondary terms of the low-carbon initiative cities are 0.089 and −0.005, indicating that the effect of AI on CE in the low-carbon pilot cities presents an inverted U-shaped curve of promotion followed by suppression, while the effect on non-low-carbon initiative cities is not significant. On the one hand, low-carbon pilot cities are mostly cities with higher levels of AI development, and the effect is worse at the initial stage of policy implementation; in the meantime, to consolidate the extent of AI development in the city, there is no considerable decrease in CE under the original development needs, and AI still promotes CE at this stage. With the continuous improvement and execution of the policy, low-carbon pilot cities have improved the development strategy of AI by optimizing the industrial structure, adjusting the industrial layout and other measures, reducing the original emissions, and suppressing the level of CE through energy saving and efficiency. On the other hand, the stage of development of AI in non-low-carbon pilot cities is low, and such cities mostly rely on the primary industry to develop the economy and do not focus on the advancement of the emerging science and technology industry. The influence of AI on CE in such cities is not significant.

③

Heterogeneity based on regional industrialization levels

When exploring the impact of AI on CE, there are significant differences in the industrialization levels of different regions in terms of technology application, energy consumption patterns, and CE characteristics, which makes the effect of AI on CE heterogeneous. In regions with a higher level of industrialization, AI may act more on the improvement of production efficiency and the optimization of energy use, thus producing a larger reduction effect on CE. Based on this, this paper expresses the regional industrialization level by the proportion of added value of the secondary industry to the GDP and divides it into high-level and low-level regions according to the median for separate tests. The results, as shown in

Table 7. Heterogeneity test of regional industrialization levels.

	(1)	(2)
Variables	High Level of Industrialization	Low Level of Industrialization
${AI}_{it\_dis}$	0.044 **	−0.006
	(2.57)	(−0.26)
${AI}_{it\_dis}$2	−0.002 ***	−0.000
	(−2.70)	(−0.08)
Constant	17.694 ***	16.073 ***
	(34.74)	(32.82)
Control	Yes	Yes
Observations	1710	1710
R-squared	0.345	0.225
Number of cities	106	179

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

, show that the primary and secondary terms of AI are significant and show an inverted “U”-shaped relationship at the industrialization level, while they are not significant at the low level, which may be due to the fact that regions with high levels of industrialization are usually faced with greater pressure on energy consumption and CE. The application of AI in these regions, especially in the production process [31], can significantly reduce energy consumption by optimizing resource allocation and improving energy efficiency, thus bringing more obvious emission reduction effects. The industrial structure of low-industrialized regions is relatively simple and relies more on services and light industry, which has less demand for AI technology.

4.3.4. Mediation Effect Test

To test the intrinsic mechanism of AI affecting CE, therefore, this paper draws on the research of Wen et al. [32] to set up the following model to quantitatively analyze the transmission effects of energy efficiency ${Ene}_{it}$, factor market allocation ${Fac}_{it}$, and industrial structure ${\mathrm{I}\mathrm{n}\mathrm{s}}_{\mathrm{i}\mathrm{t}}$.

$$ln{C}_{it}={\alpha }_0+{\alpha }_{1}ln{AI}_{it\_dis}+{\alpha }_{2}ln{AI}_{it\_dis}^2+X_{it}^{\prime }\gamma +{\mu }_{it}$$

(9)

$$ln{Med}_{it}={\beta }_0+{\beta }_{1}ln{AI}_{it\_dis}+X_{it}^{\prime }\gamma +{\mu }_{it}$$

(10)

$$ln{C}_{it}={\delta }_0+{\delta }_{1}ln{AI}_{it\_dis}+{\delta }_{2}ln{AI}_{it\_dis}^2+{{\delta }_{3}ln{Med}_{it}+{\delta }_{4}ln{Med}_{it}^2+X}_{it}^{\prime }\gamma +{\mu }_{it}$$

(11)

Here, ${Med}_{it},$ the mediating variables, include energy efficiency ${Ene}_{it}$, factor market allocation ${Fac}_{it}$, and industrial structure ${Ins}_{it}$, where energy efficiency ${Ene}_{it}$ is assessed by the ratio of GDP regarding total energy consumption (10,000 tons of standard coal) in prefecture-level cities, factor market allocation ${Fac}_{it}$ adopts Fan’s [33] measure of the factor marketization index, and the industrial structure ${Ins}_{it}$ employs the ratio of tertiary to secondary industry output value.

The test is divided into three steps: the first is to test the significance of the sum of the coefficients in column (3) of the benchmark regression Table 4. If the sum is significant, it indicates that Formula (9) can pass the significance test and enter the test of Formula (10). The following step is to test the significance of columns (1), (3), and (5) in

Table 8. Intermediation effects.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	${\mathit{E}\mathit{n}\mathit{e}}_{\mathit{i}\mathit{t}}$	${\mathit{C}}_{\mathit{i}\mathit{t}}$	${\mathit{F}\mathit{a}\mathit{c}}_{\mathit{i}\mathit{t}}$	${\mathit{C}}_{\mathit{i}\mathit{t}}$	${\mathit{I}\mathit{n}\mathit{s}}_{\mathit{i}\mathit{t}}$	${\mathit{C}}_{\mathit{i}\mathit{t}}$
${AI}_{it\_dis}$2		−0.006 ***		−0.005 ***		−0.002 ***
		(−4.36)		(−3.53)		(−2.72)
${AI}_{it\_dis}$	0.003 **	0.133 ***	0.001 ***	0.129 ***	0.002 *	0.033 **
	(2.68)	(4.35)	(5.98)	(4.31)	(1.65)	(2.21)
${Ene}_{it}$2		−0.050 ***
		(9.02)
${Ene}_{it}$		0.383 ***
		(3.01)
${Fac}_{it}$2				−0.047 ***
				(−7.57)
${Fac}_{it}$				0.033 ***
				(3.78)
${Ins}_{it}$2						−0.172 ***
						(−4.07)
${Ins}_{it}$						0.493 ***
						(3.88)
Constant	5.695 ***	15.076 ***	0.139 ***	9.493 ***	2.654 ***	16.610 ***
	(40.66)	(17.26)	(19.85)	(24.27)	(10.80)	(43.16)
Control	Yes	Yes	Yes	Yes	Yes	Yes
Observations	3420	3420	3420	3420	3420	3420
R-squared	0.300	0.081	0.821	0.248	0.518	0.245
Number of cities	285	285	285	285	285	285

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

. If it is significant, it indicates that there is a significant relationship between AI and the three intermediary variables. Then, in the test of Formula (11), the third step, we introduce the quadratic terms of AI and mediating variables and observe the significance of ${\mathsf{\delta }}_1,{\mathsf{\delta }}_2$,${\mathsf{\delta }}_3$,${\mathsf{\delta }}_4$ in columns (2), (4), and (6) of Table 8 to test and verify whether AI affects the CE through the three mediating variables, and the product of ${\mathsf{\beta }}_1$ and ${\mathsf{\delta }}_4$ can be used as the value for testing the strength of the mediating effect.

Further, to test that each mediating variable plays a nonlinear mediating role, this study utilizes the bootstrap method, setting the number of repeated samples to 5000 and the confidence interval to 95%. The results of the analysis are shown in

Table 9. Bootstrap test for mediation effects.

Mechanism Variables	95% Confidence Interval
	Upper Limits	Lower Limits
${Ene}_{it}$	0.00037	01191
${Fac}_{it}$	0.00327	0.01200
${Ins}_{it}$	0.00070	0.06237

, and the confidence intervals corresponding to the values of the instantaneous mediating effects of the three mediating variables do not contain 0. This indicates that the instantaneous mediating effects of energy efficiency, factor allocation, and industrial structure on the relationship between AI and CE diminish from positive to significant with the increase in AI; i.e., nonlinear mediating effects exist.

First, the mediating effect of energy efficiency ${\mathrm{E}\mathrm{n}\mathrm{e}}_{\mathrm{i}\mathrm{t}}$ is tested. The results presented in column (1) of Table 8 test the relationship between AI and energy efficiency ${Ene}_{it}$, and the calculated coefficient of AI is 0.008, which shows significance at the 5% threshold, indicating that AI and energy efficiency are strongly positively correlated; i.e., AI can promote energy efficiency. The results of column (2) test the relationship among AI, energy efficiency, and CE; the calculated coefficient of the main term of energy efficiency is 0.384, and the calculated coefficient of the quadratic term is −0.05, which is significant at 1%, and therefore, it can be concluded that AI can promote energy efficiency to improve and thus affect the CE, and the energy efficiency is an “inverted U-shaped” mediator of CE. Considering the “inverted U-shaped” mediating effect on CE, this nonlinear relationship may be because in the initial stage, while AI-driven advancements in energy efficiency can reduce energy consumption per unit of output, they may also lower the effective cost of energy, incentivizing increased production or expanded use of energy-intensive technologies. This could offset some of the expected reductions in CE, particularly in industries where energy demand is elastic. Improving energy efficiency often requires technological transformation or equipment renewal, which may bring about an increase in resource consumption in the short term, especially since the introduction of new technologies often requires additional energy inputs, and the energy consumption of these inputs may exceed the benefits of energy saving in the initial stage [34], increasing CE. When energy efficiency reaches a certain high level, technological transformation and production processes use resources more efficiently, energy waste is significantly reduced, and CE begins to show a downward trend.

Second, the mediating effect of factor market allocation is tested. The findings in column (3) of Table 8 evaluate the relationship between AI and factor market allocation${Fac}_{it}$; the estimated coefficient value of AI is 0.001, which is significant at the specified level test of 1%, indicating that AI and factor market allocation are significantly and positively correlated; i.e., AI can promote the optimization of factor market allocation. The results of column (4) test the relationship among AI, factor market allocation, and CE; the estimated value of the primary variable’s coefficient of factor market allocation is 0.033, and the estimated coefficient of the secondary term is −0.047, which meets the 1% significance level test; therefore, it can be concluded that AI can promote the optimization of the factor market allocation and thus affect the CE, and the factor market allocation plays an “inverted U-shaped” role in CE. Considering the “inverted U-shaped” mediating effect, the probable explanation for this result is that for the factor market allocation in the optimization process, the initial stage may be through the enhancement of production efficiency, to promote the concentration and intensive use of resources, and the growth of the production scale, the increase in production capacity, and the agglomeration effect of resource allocation are often accompanied by an increase in energy consumption [35], so CE will increase as well. As factor market allocation is gradually optimized and enters an advanced stage, and resource allocation becomes more refined, the efficiency of energy use during the production process of enterprises may be significantly improved, and CE will be gradually suppressed.

Third, the mediating effect of industry structure ${Ins}_{it}$ is tested. The findings In Table 8, column (5), test the relationship between AI and industrial structure${ Ins}_{it}$; the estimated coefficient value for the AI is 0.002, which passes the 10% significance level test, indicating that AI and industrial structure are significantly positively correlated; i.e., AI can promote the optimization and modernization of the industrial structure. The results in column (6) test the relationship among AI, industrial structure, and CE; the regression coefficient of the main variable of industrial structure is 0.493, and the estimated coefficient of the secondary term is −0.172, which passes the 1% significance level tests; therefore, it can be concluded that AI can facilitate the enhancement of the industrial structure and thus affect the CE, and the industrial structure plays an “inverted U-shape” role in CE. The nonlinear relationship could be attributed to the fact that the process of an industrial structure optimization, may first lead to the expansion and technological upgrading of resource-intensive and high-carbon-emission industries, so that such structural adjustment will initially foster the development of high-carbon-emission-intensive industries or the expansion of resource-consuming industries, leading to a rise in CE. As the industrial structure gradually optimizes, especially through the promotion of AI technology, the traditional high-carbon emission industries will be transformed or replaced, and the rise of new green industries and low-carbon technologies will gradually take the lead, ultimately leading to a decline in CE.

4.3.5. Robustness Test

①

Exchange of explanatory variables

To avoid errors in estimation results caused by differences in the estimation methods of the independent variables, This paper uses per capita CE as a measure of electricity consumption ${ELE}_{it}$ instead of the CE to carry out the robustness test to re-estimate the level of CE to confirm the robustness of the estimation results; column (1) of

Table 10. Robustness tests.

	(1)	(2)	(3)	(4)
Variables	${\mathit{E}\mathit{L}\mathit{E}}_{\mathit{i}\mathit{t}}$	Exclusion Sample	Subsample	GMM
${AI}_{it\_dis}$	0.008 **	0.026 *	0.153 ***	0.039 ***
	(2.29)	(1.75)	(4.55)	(36.50)
${AI}_{it\_dis}$2	−0.000 **	−0.002 **	−0.007 ***	−0.002 ***
	(−2.51)	(−2.20)	(−3.92)	(−34.90)
Constant	2.195 ***	16.337 ***	9.877 ***
	(22.69)	(45.35)	(23.92)
Control	Yes	Yes	Yes	Yes
Observations	3420	3372	2565	3135
R-squared	0.722	0.255	0.215
Number of cities	285	281	285	285
AR(1)				p value = 0.000
AR(2)				p value = 0.393
Hansen				0.413

t-statistics in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

displays the results after replacing the explanatory variables, and the calculated coefficients of the main variable and the quadratic term of the influence of AI on CE are 0.008 and −0.001, showing an inverted U-shaped relationship of promotion followed by suppression, indicating the robustness of the previous regression results.

②

Excluding developed area samples

Since a significant disparity exists in the level of AI development and CE in municipalities provincial capitals and other prefecture-level cities, it is essential to carry out the regression analysis again after excluding the above samples, and the results are displayed in Table 10, Column (2). The regression coefficients of the primary term and the secondary term after excluding the samples of municipalities and provincial capitals are 0.026 and −0.002, respectively, and they are statistically significant, which indicates that for most cities, the impact of AI on CE is a nonlinear inverted U-shaped curvilinear relationship, demonstrating the robustness of the conclusions above.

③

Subsample regression

The outbreak of the novel coronavirus epidemic in late 2019 and the blockade and restriction measures taken by China to contain the spread of the virus have led to a succession of global shutdowns, supply chain disruptions, and business closures, which have significantly lowered the level of carbon emissions in various regions in a short period of time and have also created a new demand for the development of AI technology. Therefore, when discussing the relationship between AI and carbon emissions, the novel coronavirus epidemic needs to be taken into account to avoid overestimating the actual impact of AI on carbon emissions. In order to exclude the impact of the novel coronavirus epidemic, which is an uncertainty event, this paper excludes the samples after 2020 and retains only the pre-2019 samples for the robustness test. The results are shown in Column 3 of Table 10, where the significance and direction of the regression coefficients of the primary and secondary terms of AI do not change significantly from the benchmark regression results, indicating that the benchmark conclusions of this paper are still robust after excluding the shock of the novel coronavirus epidemic.

④

Endogeneity test

With the global focus on carbon emission control, stringent carbon emission policies may prompt the AI industry to focus more on low-carbon technology innovation and promote the development of energy-efficient algorithms and hardware. In addition, carbon emission restrictions may lead to higher energy prices, forcing the AI field to seek more efficient, low-energy solutions. The existence of possible reverse causality leads to endogeneity, and to solve this problem, this paper introduces the first-order lagged terms of the explanatory variables into the baseline regression model, expands the baseline regression model into a dynamic panel model, and then estimates the parameters using a two-step system GMM method, considering the lagged first-order explanatory variables and the current explanatory variables and the quadratic terms as endogenous, and using the second-order lagged terms of the two variables as the two-step instrumental variables for the system GMM estimation. The results, as shown in Column 3 of Table 10, show that the level of AI and carbon emissions still exhibit a significant inverted “U”-shaped relationship, indicating that the benchmark regression is robust.

5. Conclusions and Implications

This research provides an empirical analysis of the effect of AI on CE with panel data from 285 prefecture-level cities in China between 2011 and 2022, and the results demonstrate that the impact of AI on CE of prefecture-level cities exhibits an “inverted U-shape” relationship; i.e., in the short term, large-scale application of AI increases regional CE, and over the medium to long horizon, it will have a stronger CE reduction effect as the application scenarios and technologies of AI continue to mature. In the short run, the large-scale application of AI will increase regional CE, and in the medium and long term, as the application scenarios and technology of AI continue to mature, it will produce a strong CE reduction effect. Furthermore, this paper examines the mediating effect of AI on CE and how AI affects regional CE by influencing energy efficiency, market factor allocation, and industrial structure. Finally, this paper discusses the heterogeneity of carbon emissions in resource-based cities, low-carbon pilot cities, and cities with different levels of industrialization, and the results show that artificial intelligence has a considerable effect on the reduction in carbon emissions in resource-based cities and low-carbon pilot cities. The findings of this research underscore the potential of AI to play a transformative role in reducing carbon emissions, not only through technical and economic advancements but also by shaping social structures. As AI technologies mature, they offer the potential to foster a more sustainable and equitable future by reducing the environmental burden of industrialization. The short-term increase in CE due to AI application also suggests an urgent need for communities and industries to prepare for the long-term benefits of these technologies through strategic investments in green infrastructure and training. Furthermore, the role of AI in optimizing industrial structures and improving energy efficiency is not just an economic consideration but also an important step towards more sustainable urban living.

Based on the findings, this paper makes the following recommendations:

First, assist enterprises in carrying out technological research and development, and shorten the cycle of intelligent transformation of enterprises. The government should continue to improve the policies related to intelligent manufacturing. In the field of technology research and development, a special support fund for intelligent manufacturing can be established, an R&D subsidy can be given to core technology research projects such as AI, industrial Internet, and digital twins, and the insurance compensation mechanism for major technical equipment can be implemented, and at the same time, the government can invest in the establishment of industrial parks to focus on providing enterprises with resources related to intelligence, including the establishment of network infrastructure in the park, data processing centers, and other core facilities of intelligence, to provide all-around support for the introduction of intelligence by enterprises, ensure that enterprises have stable and reliable technical support, and shorten the time that the process of intelligent transformation has an uplifting effect on CE intensity. The intelligent transformation process has a positive effect on CE intensity.

Secondly, promote the integrated development of “intelligence + greening”. First, based on AI technology, establish an enterprise-led collaborative innovation system, support the establishment of an open, collaborative, and efficient technology research and development platform, and promote the organic integration of the innovation ecosystem, industrial, and policy chains. Second, improve the utilization rate of production factors. Optimize the allocation of resources through intelligent technology, strengthen the precise control of core resources including energy and raw materials, for instance, by promoting the recycling of resources and the greening of the production process, and gradually realize a highly efficient, energy-saving, and environmentally friendly production model. Once again, this promotes the optimization of economic structure. Leverage the spillover effect of AI to reconstruct the business model of traditional industrial industries, improve their original technical condition and production mode, and make them develop in the direction of modernization.

Thirdly, policy and financial support should be strengthened to maximize the energy-saving and carbon-reducing potential of AI. Local governments should fully leverage the comparative advantages of their regions and develop targeted and tailored policies according to local conditions. For resource cities with a strong carbon-locking effect, the transformation of the energy consumption framework ought to be accelerated to decrease the share of enterprises in high-energy-consuming industries; for low-carbon pilot cities that play a “leading role” in CE reduction, AI technology ought to be utilized to implement the low-carbon, digital, and intelligent transformation of the modern industrial ecosystem, and continue to enhance the extent of green development of industries. Simultaneously, improve the incentive mechanism for green development, set up special funds to support green intelligent manufacturing projects, and reduce the burden of enterprise transformation.