Currently, there are four proxy variables employed to characterize CEL. The initial proxy variable is the total CES (Wang et al., 2023b; Ding et al., 2023; Tang and Yang, 2023). Some scholars have developed carbon emission databases containing total CES data (Shan et al., 2020). Furthermore, a positive correlation exists between satellite light data and total CES, prompting some scholars to derive total CES based on satellite light data (Meng et al., 2023). The second proxy variable pertains to carbon density, quantified as the ratio of total CES to GDP (Chen P. et al., 2022; Li et al., 2022; Tang and Yang, 2023), the ratio of total CES to the value added by the secondary industry (Yi et al., 2022), or the ratio of industry energy-related CES to industry sales value (Liu et al., 2022). The third proxy variable relates to per capita CES, determined by dividing total CES by the year-end population (Wang et al., 2023b; Tang and Yang, 2023). The fourth proxy variable is carbon emission performance, a measure that considers both economic development and total CES. Typically, it is assessed using a Data Envelopment Analysis (DEA) model featuring multiple input and output indicators (Zhang W. et al., 2022).

Existing research indicates that numerous factors can influence CES. For example, economic development, per capita income, population size, technological advancement, green technology innovation, openness, urbanization, industrial concentration, industrial upgrading, energy regulations, energy demand, energy consumption, energy intensity, energy prices, energy structure, energy efficiency, energy innovations, human capital, carbon taxation, financial development, transportation infrastructure, environmental regulation, marketization, green total factor productivity (GTFP), working hours, digital economy, and AI can influence CES (Chen Y. et al., 2022; Zhang X. et al., 2022; Li et al., 2022; Yi et al., 2022; Wang et al., 2023b; Ding et al., 2023; Meng et al., 2023; Tang and Yang, 2023; Yanzhe and Ullah, 2023).

Given the rapid development, extensive application, and the challenge of defining precise boundary and composition for AI, an ongoing debate persists regarding the measurement of AIDL (Damioli et al., 2021; Bianchini et al., 2023). Consequently, a consensus on the measurement of AIDL remains elusive. Presently, there are five proxy variables utilized to gauge AIDL. The first proxy variable is the frequency of AI-related terms appearing in the reports of publicly listed companies (Zhang W. et al., 2022) or in the annual government work reports (Tang and Yang, 2023). The second proxy variable is the industrial robot data sourced from the International Federation of Robotics (IFR). This encompasses metrics such as the increment of industrial robots (Li and Tian, 2023), the increment of industrial robots per worker (Chen Y. et al., 2022), the stock of industrial robots (Zhang X. et al., 2022; Liu et al., 2022; Wang et al., 2023b; Li and Tian, 2023), the stock of industrial robots per unit of GDP (Li et al., 2022), the stock of industrial robots per worker (Chen P. et al., 2022; Chen et al., 2022 Y.; Li et al., 2022; Lv et al., 2022; Vorozheykina, 2022; Yang and Shen, 2023), and the adjusted penetration of industrial robots (Acemoglu and Restrepo, 2020; Chen Y. et al., 2022). The third proxy variable relates to the number of AI patents (Damioli et al., 2021; Yang, 2022; Bianchini et al., 2023). The fourth proxy variable involves AI-related research paper counts (Li et al., 2022). The fifth proxy variable employs an AI Index (Ding et al., 2023; Maslej et al., 2023), typically exemplified by Stanford University’s AI Index (Maslej et al., 2023).

Extensive research has highlighted the dual nature of AI’s impact on CES, encompassing both reduction and increase (Chen P. et al., 2022; Kaack et al., 2022; Cowls et al., 2023). On the one hand, AI has the potential to reduce CES through various pathways. AI can reduce CES by fostering green technology innovation, improving energy efficiency, and driving industrial upgrading (Elnour et al., 2022b; Himeur et al., 2022; 2023; Ding et al., 2023). The deployment of industrial robots can diminish CES by promoting green technology innovation, optimizing the industry structure, enhancing digital infrastructure, improving GTFP, lowering energy intensity, driving technological innovation, promoting research and development investment, encouraging manual labour substitution, saving work time, and promoting green employment (Chen P. et al., 2022; Chen et al., 2022 Y.; Elnour et al., 2022a; Li et al., 2022; Meng et al., 2022; Wang et al., 2023b; 2024; Li and Tian, 2023).

On the other hand, AI can also contribute to a surge in CES through various channels. First, the operation of computationally intensive industrial robots will consume substantial energy and will generate CES (Wang et al., 2023b). Training and deploying large AI models, such as ChatGPT, can generate substantial CES (An et al., 2023). Second, in addition to being responsible for the CES generated during the operational phase, AI devices should also share responsibility for the embodied emissions resulting from other stages of its life cycle, including the raw material extraction phase, manufacturing phase, transportation phase, and hardware disposal phase (Kaack et al., 2022; Wu et al., 2022; Cowls et al., 2023). Third, the digital infrastructures supporting AI development have significantly contributed to increased CES by increasing energy consumption (Tang and Yang, 2023). Last, AI’s capacity to enhance production and consumption efficiency can result in a rebound effect, leading to increased production and consumption level and consequently, elevated CES (Kaack et al., 2022).

The dual nature of AI’s impact on CES sparks ongoing debate regarding the net effect of AI on CES. Presently, there exist three main perspectives regarding this point. The first perspective contends that the net effect of AI on CES is negative. This perspective has been substantiated by research conducted at various levels, including the city (Chen P. et al., 2022; Zhang W. et al., 2022; Wang et al., 2023b), provincial (Wang et al., 2023a; Ding et al., 2023), manufacturing industry (Liu et al., 2022; Li and Tian, 2023), and national (Chen Y. et al., 2022; Li et al., 2022) levels. Furthermore, the CRE of AI demonstrates spatial heterogeneity (Chen P. et al., 2022; Chen et al., 2022 Y.; Zhang W. et al., 2022; Li et al., 2022; Meng et al., 2022; Wang et al., 2023b; 2024; Ding et al., 2023), time heterogeneity (Liu et al., 2022), industry heterogeneity (Li et al., 2022; Liu et al., 2022; Li and Tian, 2023; Wang et al., 2024), and spatial spillover (Zhang W. et al., 2022; Ding et al., 2023) characteristics. Some scholars have also pointed out that the intensity of the CRE of AI is closely related to the scale of high-skilled labor, digital endowment, and the intensity of environmental regulation (Wang et al., 2024). The second viewpoint holds that the net effect of AI on CES is positive. This perspective has been substantiated by certain studies (Bianchini et al., 2023; Tang and Yang, 2023). For example, some scholars have suggested that the carbon increment effect of AI is weaker in regions with large green technology endowments (Bianchini et al., 2023). The third perspective considers the net effect of AI on CES to be uncertain. Some scholars have proposed that estimating the overall immediate impact of AI on CES is exceedingly challenging due to the absence of data on the deployment rate of AI, the diversity of application areas, and the lack of precise procedures to attribute emissions effect to AI usage (Kaack et al., 2022). Some scholars have proposed that the development of AI does not necessarily lead to an immediate carbon emission effect, and AI can only reduce carbon emissions in the industrial sector when the level of intelligence reaches a certain threshold (Wang et al., 2024). Some scholars have also proposed that the impact of industrial robots on CES exhibits an inverted U-shaped relationship (Liu et al., 2024).

In summary, existing research has not systematically analyzed the impact mechanisms and effects of AI on CES. There is a shortfall in revealing the nonlinear CRE of AI and precisely measuring the AIDL. The main purpose of this paper is to address these gaps by utilizing provincial panel data from China.

AI can both reduce and increase CES. On the one hand, AI can reduce CES. Firstly, AI can play a pivotal role in guiding scientists, governments, and individuals to mitigate CES (Yi et al., 2022; Al-Nefaie and Aldhyani, 2023; Hu and Man, 2023; Nassef et al., 2023; Zadmirzaei et al., 2023; Zhao et al., 2023). Secondly, AI assumes a crucial role in promoting the innovation, dissemination, and adoption of green technologies (Chen P. et al., 2022; Li et al., 2022), thereby reducing CES. Thirdly, AI, which plays a crucial role in accelerating the shift of energy supply structure and energy consumption structure from a high CES scenario to a low CES scenario, is effective in mitigating CES (Chen Y. et al., 2022; Yi et al., 2022). Fourthly, AI contributes to CES reduction by facilitating the industrial structure with high CES transfer to the industrial structure with low CES (Chen P. et al., 2022; Ding et al., 2023). Fifthly, AI can enhance energy efficiency and GTFP (Paryanto et al., 2015), thereby mitigating energy consumption and CES. Sixthly, AI contributes to CES reduction by reducing trade-related costs and enhancing openness, because enhanced openness can attract foreign enterprises with advanced green technologies and management practices. Lastly, AI plays a vital role in carbon reduction by enhancing the efficiency of carbon capture (Priya et al., 2023).

On the other hand, AI can also increase CES. Firstly, AI system, including AI models and AI devices, is carbon-intensive due to its heavy energy reliance, continuous upgrading, and widespread utilization (Strubell et al., 2019; Kaack et al., 2022; Bianchini et al., 2023; Bieser et al., 2023; Cowls et al., 2023; Jean-Quartier et al., 2023). Secondly, AI has the potential to impede the transition to a more sustainable energy structure, consequently contributing to increased CES. For instance, oil companies can utilize AI to extract and sell oil and gas more efficiently, which could hinder the energy structure transformation. Thirdly, AI may lead to a rebound effect in production and consumption, consequently resulting in increased CES (Huang et al., 2022; Kaack et al., 2022). Lastly, AI can potentially contribute to increased CES by enhancing openness and expanding the scale of trade. Drawing upon the above, this paper proposes the following hypothesis:

Hypothesis 1

Although AI can exert both positive and negative impact on CES, but the CRE of AI is greater, and the net effect is a reduction in CES.

The nonlinear CRE of AI primarily stems from two aspects. On the one hand, change in the AIDL can affect the marginal CRE of AI, thereby leading to a nonlinear CRE of AI. Data, computational infrastructure, and algorithms constitute the pivotal elements of the AI system. Unlike other inputs, data often yields increasing marginal return, thereby leading to an increasing marginal CRE of AI and a nonlinear CRE of AI. For example, as the scale of data grows, AI models trained on data can become more precise in predicting CES and can offer more possibilities to promote carbon reduction. Some scholars have observed an increasing positive marginal effect of intelligent manufacturing on industrial GTFP (Yang and Shen, 2023). Therefore, this paper proposes the following hypothesis:

Hypothesis 2

The enhancement of AIDL can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI.

On the other hand, variations in other factors that facilitate the innovation and implementation of AI technologies can also impact the marginal CRE of AI, thereby leading to a nonlinear CRE of AI. Firstly, change in the level of marketization can impact the marginal CRE of AI and lead to a nonlinear CRE of AI. Increased marketization can enhance the ability to optimize resource allocation and provide more opportunities in the market to unlock the business potential of technologies (Yi et al., 2022). Consequently, a higher degree of marketization encourages the innovation and application of AI technologies, thereby magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Secondly, change in the level of human capital can influence the marginal CRE of AI and lead to a nonlinear CRE of AI. AI comprises a complex technological ecosystem. Undoubtedly, the innovation and application of AI technologies pose substantial challenges. Thus, a higher level of human capital enables a more effective identification, innovation, absorption, and application of AI technologies, drawing upon prior relevant knowledge, thereby magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Thirdly, change in the level of digital infrastructures can influence the marginal CRE of AI and lead to a nonlinear CRE of AI. Data is pivotal in both the development and application of AI technologies. A higher level of digital infrastructures can enhance the generation, collection, storage, transmission, and analysis of valuable data. Consequently, improved digital infrastructures foster the advancement and utilization of AI technologies, thereby magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Fourthly, change in the level of economic development can influence the marginal CRE of AI and lead to a nonlinear CRE of AI. Greater economic development will amplify the capacity and demand for AI products and services. This, in turn, promotes the innovation and application of AI technologies, further magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Fifthly, change in the level of openness can influence the marginal CRE of AI and lead to a nonlinear CRE of AI. A higher level of openness translates to more opportunities for acquiring new knowledge. Consequently, greater openness facilitates the identification, innovation, absorption, and application of AI technologies, thereby magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Lastly, change in the level of government intervention can influence the marginal CRE of AI and lead to a nonlinear CRE of AI. AI is recognized as a strategic technology, prompting many countries to implement policies aimed at fostering its innovation and application. Therefore, a higher level of government intervention accelerates the pace of innovation and application of AI technologies, thereby magnifying the marginal CRE of AI and leading to a nonlinear CRE of AI. Drawing upon the above, this paper proposes the following hypothesis:

Hypothesis 3

The improvement of marketization, human capital, digital infrastructures, economic development, openness, and government intervention level can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI.

This paper will employ the following model to test whether the net effect of AI on CES is negative (Li et al., 2022): C E L i t = α 0 + α 1 A I D L i t + α 2 l n p g d p i t + α 3 g o v e r i t + α 4 l n f d i i t + α 5 l n i n d u s i t + α 6 l n d r o a d i t + α 7 u r b a n i t + α 8 l n g r e e n i t + α 9 e s t r u c i t + α 10 h i g h e r i t + α 11 g t f p i t + α 12 l n e n e r g y i t + ω i + ω t + ω i t (1) where, i represents the province, and t denotes the year. C E L i t represents the carbon emission level. A I D L i t denotes the AI development level. l n p g d p i t , g o v e r i t , l n f d i i t , l n i n d u s i t , l n d r o a d i t , u r b a n i t , l n g r e e n i t , e s t r u c i t , h i g h e r i t , g t f p i t , and l n e n e r g y i t are the control variables, and these variables represent economic development level, government intervention level, openness level, industrial scale, transportation development level, urbanization level, green technological innovation level, energy structure, industrial structure upgrading, green total factor productivity, and energy consumption level, respectively. The literature basis for selecting these control variables and the measurement methods for the CEL, the AIDL, and the control variables are presented in Table 1. α 0 is the intercept term. ω i , ω t and ω i t are the province fixed effect, the time fixed effect, and the random error term, respectively. The coefficient of AI is negative, indicating that the net effect of AI on CES is negative.

As depicted in Table 1. To demonstrate the robustness of the empirical results, this article will adopt three methods to measure the CEL in the empirical analysis.

As shown in Table 1. To demonstrate the robustness of the empirical results, this paper will also adopt three methods to measure the AIDL in the empirical analysis. First, this study will use the number of AI patents to characterize the AIDL. Several studies have chosen the quantity of AI patents (lnAI) to characterize the AIDL (Damioli et al., 2021; Yang, 2022; Bianchini et al., 2023). However, identifying AI patents is not a straightforward task, as there is no unified criterion for their identification, unlike green patents. To address this challenge, most studies begin by selecting keywords related to AI and then extract the quantity of AI patents from the patent database using a keyword-matching approach (Damioli et al., 2021; Yang, 2022; Bianchini et al., 2023). In this paper, we will also extract the quantity of AI patents from the patent database based on a keyword-matching approach. Table 2 shows the keywords related to AI, drawing from existing studies and the AI category (Damioli et al., 2021; Yang, 2022; Bianchini et al., 2023). In comparison with previous research, this paper adopts a more extensive list of AI-related keywords and further categorizes them into hardware, software, and application layers.

Second, we will use both the cumulative inventory of operational industrial robots and the growth in the quantity of industrial robots to measure the AIDL. Following the methodology of related studies (Chen Y. et al., 2022; Li et al., 2022; Yang and Shen, 2023), we can obtain the cumulative inventory of operational industrial robots (lnstock) and the growth in the quantity of industrial robots (lninstall) at the provincial level, which can be measured as follows: r o b o t p t = ∑ j   l a b o u r p j t   l a b o u r j t × r o b o t j t (2) where, p represents the province. t denotes the year. j is the type of industry. r o b o t p t denotes the cumulative inventory of operational industrial robots or the growth in the quantity of industrial robots.   l a b o u r p j t and   l a b o u r j t represent the labour force quantity. r o b o t j t is the operational stock of industrial robots or the increment of industrial robots.

This paper will employ the following model to test whether the improvement of AIDL can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI (Li et al., 2022): C E L i t = λ 0 + λ 1 A I D L i t + λ 2 A I D L i t 2 + λ 3 l n p g d p i t + λ 4 g o v e r i t + λ 5 l n f d i i t + λ 6 l n i n d u s i t + λ 7 l n d r o a d i t + λ 8 u r b a n i t + λ 9 l n g r e e n i t + λ 10 e s t r u c i t + λ 11 h i g h e r i t + λ 12 g t f p i t + λ 13 l n e n e r g y i t + ξ i + ξ t + ξ i t (3) where, A I D L i t 2 is the quadratic term of AIDL. The meanings of other variables are similar to that in Formula (1).

This paper will employ the following models to test whether the improvement of marketization, human capital, digital infrastructures, economic development, openness, and government intervention level can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI (Li et al., 2022), respectively: C E L i t = γ 0 + γ 1 A I D L i t + γ 2 l n p g d p i t + γ 3 g o v e r i t + γ 4 l n f d i i t + γ 5 l n i n d u s i t + γ 6 l n d r o a d i t + γ 7 u r b a n i t + γ 8 l n g r e e n i t + γ 9 e s t r u c i t + γ 10 h i g h e r i t + γ 11 g t f p i t + γ 12 l n e n e r g y i t + γ 13 l n m a r k e t i t + γ 14 l n m a r k e t i t × A I D L i t + τ i + τ t + τ i t (4) C E L i t = β 0 + β 1 A I D L i t + β 2 l n p g d p i t + β 3 g o v e r i t + β 4 l n f d i i t + β 5 l n i n d u s i t + β 6 l n d r o a d i t + β 7 u r b a n i t + β 8 l n g r e e n i t + β 9 e s t r u c i t + β 10 h i g h e r i t + β 11 g t f p i t + β 12 l n e n e r g y i t + β 13 h c a p i t a l i t + β 14 h c a p i t a l i t × A I D L i t + δ i + δ t + δ i t (5) C E L i t = ζ 0 + ζ 1 A I D L i t + ζ 2 l n p g d p i t + ζ 3 g o v e r i t + ζ 4 l n f d i i t + ζ 5 l n i n d u s i t + ζ 6 l n d r o a d i t + ζ 7 u r b a n i t + ζ 8 l n g r e e n i t + ζ 9 e s t r u c i t + ζ 10 h i g h e r i t + ζ 11 g t f p i t + ζ 12 l n e n e r g y i t + ζ 13 d i g i t a l i t + ζ 14 d i g i t a l i t × A I D L i t + ε i + ε t + ε i t (6) C E L i t = η 0 + η 1 A I D L i t + η 2 l n p g d p i t + η 3 g o v e r i t + η 4 l n f d i i t + η 5 l n i n d u s i t + η 6 l n d r o a d i t + η 7 u r b a n i t + η 8 l n g r e e n i t + η 9 e s t r u c i t + η 10 h i g h e r i t + η 11 g t f p i t + η 12 l n e n e r g y i t + η 13 l n p g d p i t × A I D L i t + ϵ i + ϵ t + ϵ i t (7) C E L i t = θ 0 + θ 1 A I D L i t + θ 2 l n p g d p i t + θ 3 g o v e r i t + θ 4 l n f d i i t + θ 5 l n i n d u s i t + θ 6 l n d r o a d i t + θ 7 u r b a n i t + θ 8 l n g r e e n i t + θ 9 e s t r u c i t + θ 10 h i g h e r i t + θ 11 g t f p i t + θ 12 l n e n e r g y i t + θ 13 l n f d i i t × A I D L i t + ϑ i + ϑ t + ϑ i t (8) C E L i t = μ 0 + μ 1 A I D L i t + μ 2 l n p g d p i t + μ 3 g o v e r i t + μ 4 l n f d i i t + μ 5 l n i n d u s i t + μ 6 l n d r o a d i t + μ 7 u r b a n i t + μ 8 l n g r e e n i t + μ 9 e s t r u c i t + μ 10 h i g h e r i t + μ 11 g t f p i t + μ 12 l n e n e r g y i t + μ 13 g o v e r i t × A I D L i t + ν i + ν t + ν i t (9) where, l n m a r k e t i t , h c a p i t a l i t , and d i g i t a l i t denote marketization level, human capital level, and digital infrastructure level. The literature basis for selecting above three variables and the measurement methods for above three variables are presented in Table 1. The meanings of other variables are similar to those in Formula (1). The coefficient of AI and the interaction term are both significant and share the same sign, indicating that the moderating variable can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI.

A balanced panel dataset comprising 30 Chinese provinces for the period spanning 1997–2019 has been used in this study. Tibet, Hong Kong, Macao, and Taiwan are excluded from the analysis because of data unavailability. Missing data has been imputed using the interpolation method. For all currency-measured variables, the influence of inflation has been removed by using the GDP index of each province, with the base period price level set at 1998. The natural logarithm has been applied to some variables to ensure data stability and address heteroscedasticity issues. Table 3 shows the results of descriptive statistics and the data sources of variables.

Figure 1 demonstrates the chronological evolution of the number of AI patents in China from 1997 to 2019. In Figure 1, we characterize the temporal evolution of AI patent counts in China during the study period using two indicators: the total number of AI patents per year and the average number of AI patents per province per year. Two significant observations can be made from Figure 1. First, both indicators show a growth trend, indicating that the number of AI patents in China has been increasing consistently. Second, the values of both indicators have shown an accelerated growth trend since 2012, especially after 2014, with a significantly faster growth rate. This is attributed to the breakthrough advancements made in AI technologies such as deep learning, image recognition, natural language processing, and intelligent chips during this period. For instance, in 2012, Google’s deep learning algorithm achieved a breakthrough performance in the ImageNet image recognition competition. This demonstrates that the AI patent data used in this study can effectively capture the evolution of AI technology development. It further reinforces the scientific validity of the AI patent data employed in this research.

Table 4 presents the spatial distribution of the number of AI patents across China’s provinces in certain years. From Table 4, it can be observed that the provinces with a high level of economic development and technological innovation capability, such as Beijing, Tianjin, Liaoning, Shanghai, Jiangsu, Zhejiang, Anhui, Fujian, Shandong, Guangdong, Henan, Hubei, Hunan, Shaanxi, Sichuan, and Chongqing, possess a larger number of AI patents. This indicates that AI technological innovation is closely related to the level of economic development and technological innovation capability.

The above results can be discussed from two perspectives. The first perspective is the CRE of AI. The findings in this study indicate that AI can reduce CES, further confirming existing research viewpoints (Li et al., 2022). However, in contrast to existing research, this study also reveals that using the quantity of industrial robots to measure AIDL might underestimate the CRE of AI. The second perspective is the nonlinear CRE of AI. There has been limited focus in existing research on the nonlinear CRE of AI, whereas this paper places particular emphasis on this aspect. The results in this study suggest that increasing AIDL can amplify the marginal CRE of AI and lead to a nonlinear CRE of AI. However, some scholars have proposed that the impact of industrial robots on CES exhibits an inverted U-shaped relationship (Liu et al., 2024). This result indicates that increasing AIDL will first raise the CES before eventually reducing CES. This result differs somewhat from the findings of this paper. The discrepancy may be due to the different methods used to measure AIDL in the two studies. Additionally, the results in this study suggest that even if the AIDL remains constant, changes in other factors can also amplify the marginal CRE of AI and lead to a nonlinear CRE of AI. While existing research has indicated that enhancing technological absorption capacity can strengthen the CRE of AI (Li et al., 2022), this study suggests that there are additional factors, including marketization, human capital, digital infrastructure, economic development, openness, and government intervention, can amplify the marginal CRE of AI and lead to a nonlinear CRE of AI. Another novel finding relative to existing research is human capital and digital infrastructure can play the most significant role in amplifying the CRE of AI.

The principal findings are as follows. Firstly, during the study period, the number of AI patents in China has shown a continuous growth trend. Since 2012, the growth of AI patents in China has accelerated, especially after 2014, when the number of AI patents in China entered a stage of rapid growth. Secondly, the provinces with a high level of economic development and technological innovation capability, such as Beijing, Tianjin, Liaoning, Shanghai, Jiangsu, Zhejiang, Anhui, Fujian, Shandong, Guangdong, Henan, Hubei, Hunan, Shaanxi, Sichuan, and Chongqing, possess a larger number of AI patents. Thirdly, although AI can exert both positive and negative impacts on CES, the CRE of AI is more substantial. This conclusion holds true even after considering robustness, endogeneity, and spatial heterogeneity. It is worth noting that employing the quantity of AI patents to gauge the AIDL can provide a more scientifically accurate measurement of the AI’s CRE. Fourthly, relative to the central and western regions, AI has significant achievement in reducing carbon intensity and per capita CES in the eastern region. However, there is still room for improvement in terms of reducing the total CES in the eastern region. Fifthly, the CRE of AI was greater during the period from 1997 to 2011 than during the period from 2012 to 2019. Sixthly, improving the AIDL can magnify the marginal CRE of AI and lead to a nonlinear CRE of AI. Lastly, even if the AIDL remains constant, changes in other factors such as marketization, human capital, digital infrastructure, economic development, openness, and government intervention can also amplify the marginal CRE of AI and lead to a nonlinear CRE of AI.

Based on the aforementioned conclusions, the following policy recommendations can be formulated. Firstly, facilitating AI technology innovation and leveraging AI for carbon reduction. AI can reduce CES, and improving the AIDL can magnify the marginal CRE of AI. Thus, the government should prioritize the development and utilization of AI. The enterprises should expedite the application of AI in various activities, including green energy production, the production of environmentally friendly products and services, carbon emission monitoring, carbon market trading, carbon sink management, and carbon capture technology innovation. Secondly, when assessing the CRE of AI, it is essential to utilize AI patent data and take into account the nonlinear CRE of AI. Employing the quantity of AI patents to gauge the AIDL can provide a more scientifically accurate measurement of the AI’s CRE, and the CRE of AI exhibits a nonlinear characteristic. Thus, government and research institutions should take these influences into account when assessing the CRE of AI. Lastly, optimizing the economic and social environment is crucial to fully unleash the carbon reduction potential of AI. Even if the AIDL remains constant, changes in other factors such as marketization, human capital, digital infrastructure, economic development, openness, and government intervention can also amplify the CRE of AI. Thus, the government, in the process of utilizing AI for carbon reduction, should not confine its focus solely to the development of AI but also consider the impact of other factors. For example, the government and other relevant stakeholders should refine the marketization, human capital, digital infrastructure, economic development, openness, and government intervention to amplify the CRE of AI.

The main potential challenges and practical considerations in implementing the recommended policies are follows: first, promoting AI development requires ensuring data security. However, managing and protecting data securely is a challenge. The government can introduce strict data privacy and security laws to ensure the protection of data. Second, as AI technology becomes more advanced, ethical and moral questions arise, such as the use of AI in decision-making processes that affect human lives. The government should establish ethical frameworks and guidelines for AI use. This can ensure that AI is used responsibly and does not harm human interests.

The limitations in this study are as follows. Firstly, the research conclusions are drawn based on Chinese data, and the processes of CES and AI development in China may differ from other countries. Therefore, some research findings may not be applicable in other nations. Secondly, this paper only considers the roles of some factors, including marketization, human capital, digital infrastructure, economic development, openness, and government intervention, in amplifying the CRE of AI. There may be additional factors that can amplify the CRE of AI. Lastly, this article analyzes the impact of AI on CES using traditional panel econometric models. However, the impact of AI on CES may exhibit a spatial spillover effect, which suggests that the models used in this article still requires further improvement. In the future, a spatial panel econometric model will be employed to analyze the spatial spillover effect of AI on CES.