Is There a Conflict between Automation and Environment? Implications of Artificial Intelligence for Carbon Emissions in China

Abstract

While artificial intelligence (AI) has had a great impact on the global economy, it has also brought new hope and opportunities for environmental protection. In this context, the authors of this paper collected balanced panel data for 30 Chinese provinces during 2006–2019 and studied the impact of AI development on local carbon emissions by using a two-way fixed-effect model. The results show that AI has significantly lowered carbon emissions. Using a series of robustness tests and instrumental variable (IV) analysis, it was found that the results are still reliable. Furthermore, mechanism analysis revealed that AI mainly reduces carbon emissions by improving energy structure and technological innovation. The lower the dependence on fossil energy, the higher technological innovation becomes, and the better the carbon reduction effect of AI. In addition, the regional heterogeneity test detected that the emission reduction effect of AI is best in the East, followed by the West, and not significant in the Central region. Therefore, to fully exploit the positive effects of AI on carbon emissions, this paper suggests accelerating intelligent transformation, formulating differentiated AI development strategies, promoting the green transformation of energy usage, and strengthening local human capital accumulation.

1. Introduction

With the rapid industrialization and urbanization around the world, greenhouse gas emissions, such as carbon dioxide, have increased rapidly, which has led to extreme climate disasters, such as global warming, glacier melting, and sea-level rise, posing a serious threat to human life and the global ecosystem [1,2]. Therefore, reducing carbon emissions has emerged as a major priority for nations worldwide. Since the reform and opening up in 1978, China’s economy, a sizable developing nation, has experienced extraordinary growth as a result of institutional transformation and a series of institutional innovations [3]. However, China’s extensive growth, which is characterized by high energy consumption, high pollution, and high emissions, has led to an increasingly prominent environmental pollution issue, especially the large amount of carbon emissions, which has seriously affected the quality of China’s economic growth (as shown in Figure 1). According to the World Energy Statistics Yearbook 2020, China became the world’s largest carbon emitter in 2019. Hence, reducing carbon emissions is the key to coordinating economic growth and environmental protection [4]. Thus, the Chinese government proposed to actively and steadily promote carbon neutrality and carbon peaks in 2022. Obviously, China is actively seeking the optimal emission reduction path, striving to fulfill its emission reduction commitments, and contributing Chinese wisdom and Chinese solutions to global emission reduction. In this context, what factors affect carbon emissions? How can an emission reduction path that adapts to China’s economic structure be built? Answering these questions is of great significance to addressing global climate change and achieving high-quality development of China’s economy [5,6].

To achieve carbon emission reduction under the premise of sustained economic development, the key lies in the continuous advancement of technology [7]. In recent years, AI, as a vital technology driving global technological revolution and growth, has been developing rapidly in China, with more and more robots being widely used in industrial production activities [8]. According to the International Federation of Robotics (IFR), China ranked first in the world, with 139,859 robots installed in 2019. Existing studies show that AI has profound impacts on economic growth, labor productivity, labor market, income distribution, market structure, industrial organization, and legal rules and legal order, but the effect of AI on environmental pollution, especially its carbon emission reduction effect, is still a research gap and an urgent problem to be solved [9]. In the era of big data, AI, as a disruptive frontier technology, has a strong ability to acquire environmental information, expanding the spatial and temporal scope of environmental protection and opening up new paths for lowering carbon emissions. Theoretically, the link between AI and carbon emissions is complicated. On the one hand, AI development helps to accelerate production adjustment, reduce fossil energy consumption, improve energy use performance, and lower carbon emissions. On the other hand, the rebound effect of its rapid expansion in industrial sectors, driven by the rapid growth of robotics investments, has brought about concern about the potential growth of carbon emissions caused by AI [10,11]. So, as a developing country, what is the impact of AI development on China’s carbon emissions? What is the internal mechanism of its effect? Does it show spatial heterogeneity? To explore these questions, it is of great importance to assess the effect of AI on environmental pollution, give full play to the technological dividend brought by AI development in the process of lowering carbon emissions, accelerate the realization of the double carbon goal in the future, and thus, promote the sustainable and healthy development of the Chinese economy.

The authors of this paper mainly analyzed the transmission mechanism of AI on carbon emissions based on a two-sector monopolistic competitive model, collected panel data from 30 provinces in China from 2006 to 2019, and empirically investigated, using a two-way fixed effect model, the effect of AI on local carbon emissions. Compared to the existing literature, the novelties of this study mainly have four aspects. Firstly, this paper constructs a two-sector monopolistic competition model, including householders and enterprises, and uses mathematical methods to theoretically explain the impact mechanism of AI development on carbon emissions. Secondly, using the Chinese industry robot usage data issued by the International Federation of Robotics (IFR) and matching it with seven major industries in China, we construct a novel index that can reflect the level of AI development in each region, namely the density of industrial robots installed in each province. Thirdly, by using a two-way fixed effects model, we confirm that there is a negative link between AI application and carbon emissions, indicating that AI development can effectively promote regional carbon emission reduction. After a series of robustness tests and IV analysis, the results still hold. Finally, in terms of impact mechanisms, this study verifies that AI can affect carbon emissions through energy structure and technological innovation. In conclusion, the above findings not only enrich the existing research but also provide decision-making references for Chinese governments to seek the optimal carbon reduction path in the context of the information era.

The rest of this paper is organized as follows. The literature on the nexus between AI and carbon emissions is reviewed in Section 2. Next, based on a two-sector monopolistic competition model, the impact mechanisms of AI on carbon emissions are discussed. The econometric model and data used for empirical analysis are presented in Section 4. The empirical results are presented in Section 5. Finally, we conclude the study.

2. Literature Review

When it comes to the nexus between AI and carbon emissions, the literature related to this study can be divided into three streams: the factors influencing carbon emissions, the economic effects of AI, and the impact of AI on environmental pollution. An overview of the recent research relating to these topics is presented as follows.

2.1. Factors Influencing Carbon Emissions

Due to the serious negative impact of carbon emissions on global ecology and human life, scholars are broadly discussing what factors affect carbon emissions [6,12]. Specifically, the factors influencing carbon emissions can be further divided into three categories, namely institutional factors, economic factors, and population factors. In terms of institutional factors, the improvement of the carbon emissions trading system can greatly promote energy conservation in a country or region and improve the allocation efficiency of energy, thus reducing carbon emissions [13,14]. For example, Neves et al. [15] deliberated the impact of environmental regulatory policies on carbon emissions based on cross-country data for 17 European Union countries from the period of 1995–2017. The results showed that environmental regulation can reduce carbon emissions by attracting high-quality investments. This conclusion was also verified by Croci et al. [16]. Based on the difference-in-difference (DID) model and the prefecture-level data for 113 cities in China during 2004–2015, Zhang et al. [17] detected that the carbon trading policy adopted in 2013 has effectively reduced urban carbon emissions by about 16.2%. By using a quasi-natural experiment for 30 provinces in China from 2004 to 2019, Guo et al. [18] confirmed the effectiveness of the carbon trading policy to lower regional carbon emissions. These results were also further confirmed by Yang and Wang [19]. In addition, Guo et al. [20] argued that a carbon tax is an effective means of curbing carbon emissions and preventing global climate change, and this result was also echoed by Jia and Lin [21]. In terms of economic factors, there is no agreement on the link between carbon emissions and economic development. Using the data of BRIC countries from the period of 1971–2005, Pao and Tsai [22] found that there was an inverted U-shaped nexus between economic growth and carbon emissions, indicating that as real output continues to increase, carbon emissions will first increase, but will eventually decrease. Based on the panel data of 27 European Union countries from 1988 to 2009, Lee and Brahmasrene [23] studied the nexus of economic growth, tourism, and carbon emissions. The results showed that economic growth significantly promoted carbon emissions, but tourism did not. By using Chinese provincial panel data during the period of 1995–2017, Li et al. [4] investigated the impacts of macroeconomic factors on carbon emissions. The results showed that green investments reduced carbon emissions, but financial development and energy investments increased carbon emissions. Using Chinese provincial panel data, Xu and Li [5] explored how fiscal decentralization affected carbon emissions and confirmed that fiscal decentralization promoted regional carbon emissions through the spatial spillover effect. In terms of demographic factors, scholars have mainly focused on the impacts of population aging, population agglomeration, and population flow on carbon emissions [24]. For example, by using Chinese provincial panel data from the period of 1997–2012, Zhang and Tan [25] studied the link between population aging and carbon emissions. Their results demonstrated that there was a significant positive correlation between population aging and carbon emissions in China. Using the panel data of 207 large cities in China from 2005 to 2018, Yi et al. [26] explored the impact of population agglomeration on urban carbon emissions. They confirmed that population agglomeration significantly raised urban carbon emissions. This conclusion is highly similar to those from the study of Yi et al. [27]. Based on a panel fixed-effect model for 30 provinces in China from 2005 to 2018, Wu et al. [28] empirically tested the impact of population flow on China’s carbon emissions. The results showed that China’s population flow can reduce the growth of carbon emissions. In conclusion, although the academic community has carefully examined the factors influencing carbon emissions, there is still a scarcity in the literature on how to achieve carbon reduction from an AI development perspective. As revolutionary information technology in the era of big data, AI development can bring about a profound impact on economic growth and environmental protection in a country or region. Unfortunately, there is still a research gap to be filled in the theoretical and empirical research on the impact of AI development on regional carbon emissions.

2.2. The Economic Effects of Artificial Intelligence

Artificial intelligence (AI), as a key indicator of a new technological revolution in the big data era, has yielded a profound impact on the global economy [29,30,31]. Accelerating AI development led by industrial robot applications has become an important means to promote high-quality economic development. Therefore, exploring the economic effects of AI development has increasingly become a focus in academia [32]. At present, it can be found that studies on AI development in the field of economics have mainly focused on the impact of robot usage on employment. Specifically, some scholars argued that industrial robot use would reduce production costs for enterprises and encourage them to increase capital investment to replace the labor force, thereby reducing employment [33]. For example, based on the industrial robot inventory data in the United States from 1990 to 2007, Acemoglu and Restrepo [34] explored the nexus between robot usage and employment. They confirmed that robot usage reduced employment rates and workers’ wages. Using the panel data for 42 countries from the period of 2001–2017, Jung and Lim [8] discovered that the use of industrial robots could suppress employment growth, indicating that there was a labor-substituting effect of industrial robots. Based on panel data for 30 provinces in China during 2003–2017, Ma et al. [35] detected that industrial robot use reduced middle-skill employment in the manufacturing industry. On the contrary, other scholars believe that industrial robot use could drive enterprises to shift their production mode towards automation and increase the labor demand for non-automated production [36,37]. For example, based on industry robot data for 17 countries from 1993 to 2007, Graetz and Michaels [38] detected that an increase in robot usage density could promote an increase in labor productivity and wages. Using the China Customs Database during 2000–2012, Zhang et al. [39] empirically found that industrial robot use could promote firm-level employment by 31.65%. In addition, some scholars believe that there is a non-linear link between AI and employment [40]. For instance, by using a panel smooth transition regression model and data for 40 countries from 2000 to 2019, Nguyen and Vo [41] explored the impact of AI on employment. The results exhibited that the non-linear relationship between AI and unemployment depended on inflation. Meanwhile, some scholars have explored the economic impacts of AI development from the perspectives of economic growth, income distribution, industrial upgrading, and technology innovation [42]. For example, based on the new data on AI job recruitment for 343 cities in the United States, Makridis and Mishra [43] showed that AI could significantly promote economic growth through the agglomeration of modern service industries. Using the panel data of 265 prefecture-level cities in China during 2005–2019, Qian et al. [44] confirmed that AI development could promote local green economic growth, but it had a siphon effect on neighboring green economic growth. The results were highly in accordance with the studies of Fan and Liu [45] and Zhao et al. [46]. Based on the survey data of Chinese private enterprises from 2008 to 2016, Wang et al. [47] examined how AI affected income distribution in private enterprises. The findings showed that AI development promoted the transfer of income from labor to capital by reducing labor demand and improving capital productivity. Lin et al. [48] believed that the application of industrial robots could promote the flow of production factors between industrial sectors and improve the allocation efficiency of production factors, thereby driving industrial upgrading. This result was further echoed by Zou and Xiong [49]. Using the panel data of Chinese 14 manufacturing sectors during 2008–2017, Liu et al. [50] argued that AI development significantly promoted technology innovation by improving knowledge absorptive capacities, accelerating technology spillover, and increasing R&D investment. To sum up, the previous literature has deeply recognized the impact of AI on labor employment, economic growth, income distribution, industrial upgrading, and technological innovation in economics. However, scholars have not given sufficient attention to the impact of AI on environmental pollution, especially carbon emissions. Environment protection and high-quality development are essentially coexisting and mutually beneficial. Therefore, strengthening the research on the environmental effects of AI development, especially for developing countries, is of great significance for achieving ecological environment protection and reducing carbon emissions.

2.3. The Impacts of Artificial Intelligence on Environmental Monitoring and Governance

In the field of natural science, many scholars have noticed the role of AI technology in environmental monitoring and pollution control [51]. In recent years, with the widespread application of AI technology in environmental governance, which aims to solve environmental problems and improve environmental quality, requiring the active participation and cooperation of governments, enterprises, and citizens, informatization and intelligence have become mainstream trends in environmental protection, which will have a profound effect on reducing environmental pollution [9,52,53]. Some scholars have even applied AI to the supervision and control systems of waste treatment. For example, Sengorur et al. [54] applied AI technology to an automatic odor-sensing system for waste treatment companies. Liu et al. [55] confirmed that AI effectively improved the efficiency of pollution control by optimizing the production conditions of enterprises and promoting the updating and upgrading of sewage treatment equipment. Furthermore, for environmental decision-making assistance, Chen et al. [56] developed an environmental decision support system by integrating different AI technologies. Subsequently, more and more scholars have focused on how AI may support auxiliary decision-making for environmental policy design. However, a few scholars have pointed out that AI could have adverse effects on environmental protection. For instance, Sharifi et al. [57] believe that the shortcomings of AI technology and its profound impact on government decision-making mechanisms, information dissemination mechanisms, and social organizational structures have brought potential risks and huge costs to environmental governance. Sepulveda [58] further pointed out that the inherent algorithmic flaws in AI technology can deviate from design objectives, increase uncontrollable risks, and lead to significant errors and judgments by environmental decision-makers. Zhang et al. [59] believe that lowering the threshold of AI environmental technology may bring risks of subject opposition, as well as the risk of losing control of public opinion due to changes in the dissemination mode of ecological environment information. Thus, it is not difficult to observe that there is no consensus in the literature on the impact of AI on environmental pollution. With increasingly prominent environmental issues, a few scholars have begun to empirically analyze the impact of AI development on environmental pollution [60]. For example, based on panel data for 30 provinces in China during the period of 2010–2020, Wang and Li [61] explored the relationship between industrial intelligence and environmental pollution. They revealed that local industrial intelligence could improve local environmental pollution. Using the Chinese prefecture-level panel data from 2013 to 2018, Yu et al. [62] confirmed that industrial robot use significantly reduced urban air pollution by improving energy use efficiency and promoting green technological innovation, which remained robust after addressing endogeneity. This result was further verified by Li and Tian [63]. In short, AI technology can not only improve energy efficiency and reduce environmental pollution, but it may also have a negative impact on environmental protection and even the environment itself. Environmental governance is a long-term and complex process that requires the joint efforts of governments, enterprises, and citizens. However, the existing literature is mostly limited to the description and logical deduction of AI situations, lacking strict data support and argumentation. Conducting a quantitative evaluation of the carbon emission reduction effects of AI technology is an important direction for future research.

3. Theoretical Analysis

This section aims to theoretically explain the impact mechanisms and effects of AI on carbon emissions. Referring to Acemoglu and Restrepo [34], it is assumed that in a monopolistic competitive market composed of differentiated products, enterprise uses labor and robots for production, and the consumer’s utility level is determined by homogeneous products, differentiated products, and carbon emissions.

3.1. Consumer Demand

Assume that there is a typical consumer in the goods market. The utility level is affected by the consumption of homogeneous and differentiated products as well as carbon emissions. In this case, the consumer’s demand preference for the differentiated product ω satisfies the constant elasticity of substitution (C.E.S.), and an increase in carbon emissions per unit product will reduce the utility level. Then, the total utility function can be expressed as follows:

$$U=q_0{}^{1-\mu }{\left[{\displaystyle \underset{\omega \in \Omega }{\int }{\left(\lambda {(\omega )}^{-1}q(\omega )\right)}^{(\sigma -1)/\sigma }d\omega }\right]}^{\mu \sigma /(\sigma -1)}$$

(1)

where Ω is the set of differentiated products, q 0 is the number of homogeneous products, q(ω) is the quantity of product ω, and λ(ω) is the carbon emissions per unit of product ω, which is negatively correlated with the utility level. μ is the spending proportion on differentiated products, and 0 ≤ μ ≤ 1. σ is the substitution elasticity for several differentiated products, and σ > 1. For simplicity, it is assumed that the homogeneous product is the numeraire. By solving the utility maximization problem, the overall consumer price index for differentiated products can be obtained as follows:

$$P={\left[{\displaystyle \underset{i\in \Omega }{\int }{\left[p(\omega )\lambda (\omega )\right]}^{1-\sigma }d\omega }\right]}^{\frac{1}{1-\sigma }}$$

(2)

where P is the overall price level, and p(ω) is the price of the differentiated product ω.

3.2. Production Behavior of Enterprises

Referring to Acemoglu and Restrepo [34], it is assumed that each firm produces a differentiated product using both robots and labor, that the productivity of robots φ(m) is equal, and that the labor productivity φ_i(l) is heterogeneous across firms and obeys the Pareto distribution, that is, 1 − G_i(φ) = φ^−κ. The factor price of labor is w and the factor price of the robots is r. These two production factors are complete substitutes. Therefore, whether a firm uses robots or labor to produce depends on the ratio of the marginal cost of the two.

Specifically, a firm’s production costs consist of marginal costs, pollution abatement fixed costs, and general fixed costs. It is assumed that in order to reduce carbon emissions, enterprises will not only increase their marginal costs but also invest fixed costs to maintain technological innovation related to pollution reduction. Following Kugler and Verhoogen [64], the fixed cost function of pollution abatement for firms is assumed to be f(λ_i) = 1/(λ_i^αα), with α measuring the degree of investment in pollution abatement, indicating that the stronger the abatement effort, the higher the cost of technological innovation, and α > σ − 1. Following Flach and Unger [65], it is assumed that enterprises will invest more in production factors to reduce the carbon emissions per unit of product to achieve pollution abatement process control in the production process, leading to an increase in the marginal cost. In addition, we also considered the general fixed cost invested by firms in marketing, advertising, and other activities. Finally, the firm’s total cost function can be set as follows:

$$T{C}_i=\{\begin{array}{cc}{q}_{i}w/{\lambda }_i{}^{\theta }{\phi }_i(l)+f({\lambda }_i)+fe,& \left({\phi }_i(l)/w\right)>\left(\phi (m)/r\right)\\ q_{i}r/{\lambda }_i{}^{\theta }\phi (m)+f({\lambda }_i)+fe,& \left({\phi }_i(l)/w\right)\le \left(\phi (m)/r\right)\end{array}$$

(3)

where φ is the firm’s factor productivity, θ(0 < θ < 1) is the elasticity of reducing the impact of carbon emissions per unit product on the marginal costs, and fe is the general fixed cost of the firm’s inputs. Under monopolistic competition, in order to increase product demand, firms will choose the optimal price p(φ_i) with unit product carbon emissions λ_i to maximize profits. Therefore, the profit maximization problem is described as follows:

$$\mathrm{Max} \pi \left({\phi }_i\right)=p({\phi }_i)q_i-{q_i{\upsilon }_i/{\lambda }_i{}^{\theta }{\phi }_i-\left({\lambda }_i{}^{\alpha }\alpha \right)}^{-1}-fe$$

(4)

where υ_i is the cost of the factors used by firms to determine the preferred mode of production, such as the cost of labor factor w or robotic factor r. Assuming that firms set prices according to the fixed cost markup in the monopolistic competition, the optimal output and the optimal carbon emission per unit output can be obtained as follows by solving the above profit maximization problem.

$$q_i={\left[{(\mu A)}^{\alpha -\theta }{(1-\theta )}^{(1-\theta )\sigma -1}{\left(\sigma /(\sigma -1)\right)}^{(\sigma -\alpha )\sigma }{\upsilon }_i{}^{(\sigma -1)-\alpha \sigma }{\phi }_i{}^{\alpha \sigma -(\sigma -1)}\right]}^{\frac{1}{\left(\alpha -(\sigma -1)(1-\theta )\right)}}$$

(5)

$${\lambda }_i={\left[(\mu A)(1-\theta ){\left(\sigma /(\sigma -1)\right)}^{-\sigma }{\upsilon }_i{}^{1-\sigma }{\phi }_i{}^{\sigma -1}\right]}^{\frac{1}{\left((\sigma -1)(1-\theta )-\alpha \right)}}$$

(6)

where A represents the total spending on consumption products, and q_i and λ_i denote the optimal output and the optimal carbon emissions per unit of output, respectively.

3.3. Impact Mechanism of AI Development on Carbon Emissions

Since labor productivity obeys Pareto distribution, under monopolistic competition, a firm’s critical productivity and critical marginal cost in long-run equilibrium are φ*(l) and w/φ^*(l), respectively. In order to assess the effects of robot application on a firm’s carbon emissions, this study discusses two scenarios according to the setting form of a firm’s total cost function. If r/φ(m) > w/φ*(l), meaning that the marginal cost of a robot adopted by the firm, r/φ(m), is higher than the maximum marginal cost for its survival, w/φ*(l), the firm cannot survive even if it replaces labor with robots, and thus, there is no automation in the market. Conversely, if r/φ(m) < w/φ*(l), firms whose marginal cost is higher than the critical marginal cost w/φ*(l) will replace labor with robots in production. Similarly, firms whose marginal cost is between the marginal cost of robots r/φ(m) and the critical marginal cost w/φ*(l) will also use robots in production, indicating that the higher the productivity of robots is, the more firms will adopt robots, and the more widespread AI application will be. Thus, to explore the mechanism of AI on carbon emissions, this study only discusses the situation where the marginal cost of robots is lower than the critical marginal cost. Specifically, it is assumed that AI development (Robot) is measured by the proportion of the number of robots adopted by firms in production, and its expression is as follows:

$$Robot={\displaystyle \underset{1}{\overset{\phi (m)}{\int }}g({\phi }_i)d{\phi }_i}/{\displaystyle \underset{1}{\overset{+\infty }{\int }}g({\phi }_i)d{\phi }_i}=1-{\left(\phi (m)\right)}^{-\kappa }$$

(7)

where φ(m) is the productivity of robots, and κ is the Pareto distribution parameter. The total quantity of carbon emissions is defined as the sum of carbon emissions of heterogeneous firms, and the carbon emission for each firm is the product of carbon emissions per unit product and the firm’s output. Therefore, the total carbon emissions are as follows:

$$Carbon=\left(1-{\left(\phi \left(m\right)\right)}^{-\kappa }\right)N\times C{\left(\phi \left(m\right)\right)}^{\frac{\alpha \sigma -2(\sigma -1)}{\alpha (1-\xi )}}+\frac{\kappa {\left(\phi \left(m\right)\right)}^{\frac{\alpha \sigma -2(\sigma -1)}{\alpha (1-\xi )}}}{C\left(\kappa -\left(\frac{\alpha \sigma -2(\sigma -1)}{\alpha (1-\xi )}\right)\right)}$$

(8)

where ξ = (σ − 1)(1 − θ)/α, C = (μA)^{α − θ − 1}(1 − θ)^{(1 − θ)σ − 2}(σ/(σ − 1))^{(θ – α + 1)σ}, and N is the number of heterogeneous firms. Assuming σ < (2/(2 − α)), and φ(m) > (1 + ακ(1 − ξ)/(2(σ − 1) − ασ))^(1/k), we can obtain the partial derivative of carbon emissions from AI application by using Equations (7) and (8), i.e., ӘCarbon/ӘRobot < 0, indicating that AI development is negatively correlated with carbon emissions. Intuitively, AI development reduces the marginal cost of production and promotes the total factor productivity, which in turn lowers the factor inputs required for carbon reduction, thus ultimately reducing carbon emissions in the economy. Therefore, we can obtain the following research hypotheses:

Hypothesis 1 (H1).

AI development can reduce carbon emissions.

The main reason for H1 is that AI development can reduce carbon emissions by replacing labor input, promoting energy transformation, and improving pollution control efficiency. To be specific, AI development has an impact on carbon emissions through the following two mechanisms.

The first mechanism is the technological innovation effect. Green innovation has provided a new impetus for achieving China’s ecological civilization and high-quality development. Enterprises can utilize green technology innovation to optimize resource allocation, thereby reducing the dependence on fossil fuels in production and ultimately reducing carbon emissions. However, technology innovation has the characteristics of high capital investment, high market risk, and a long research period, and enterprises lacking innovation resources may not turn to green innovation paths led by green technology. AI development helps enterprises to innovate their production mode, promote the automated collection and analysis of market demand information, strengthen their response capabilities, improve the matching efficiency between the product and the market, and significantly promote technology innovation while alleviating resource misallocation. Consequently, the second hypothesis can be summarized as follows:

Hypothesis 2 (H2).

AI development can significantly reduce carbon emissions by enhancing technological innovation (TI). The higher the level of TI, the better the emission reduction effect of AI development.

The second mechanism is the energy structure effect. AI development can improve energy efficiency and promote energy transformation, thus reducing carbon emissions. Specifically, China’s economic growth is mostly driven by energy consumption, which is also a significant contributor to carbon emissions. At present, the utilization efficiency of traditional fossil fuels, namely coal and oil, is relatively low in China. There is an amount of energy waste and loss during energy extraction, processing, conversion, delivery, and use at the terminal, leading to increasingly prominent carbon emissions. AI development has provided new ideas for solving the problem of low energy utilization efficiency. On the one hand, it improves energy productivity by reducing fossil energy use and increasing added value, thereby reducing carbon emissions. On the other hand, it helps to accelerate the exploitation and use of new renewable energy, which can promote energy transformation, thus reducing carbon emissions. Therefore, the third hypothesis can be proposed as follows in regard to the analysis just mentioned:

Hypothesis 3 (H3).

AI development can effectively reduce carbon emissions by promoting the transformation of energy structure (ES). The lower the dependence on traditional fossil fuels, the better the emission reduction effect of AI development.

4. Methodology, Variables, and Data Sources

4.1. The Setting of Preliminary Econometric Model

Using the theoretical analysis presented above, this study further empirically investigated the impacts of AI on carbon emissions. Referring to Acemoglu and Resrepo [34], a two-way fixed effect panel econometric model in this study was constructed for empirical analysis, and its structure can be defined as follows:

CE_it = α + β·Robot_it + γ·X_it + μ_i + δ_t + ε_it

(9)

where the subscripts i and t denote the province and year, respectively. The explained variable CE_it reflects the carbon emissions of province i in year t. The key explanatory variable Robot_it reflects the development level of AI, which is measured by the installation density of industrial robots in each province. The variable X_it is a series of control variables, including economic development (PGDP), the level of industrialization (IS), environmental regulation (ER), population aging (PA), and economic openness (FDI). The regression coefficients are expressed by the symbols α, β, and γ in the model. In particular, β reflects the impact of AI development on carbon emissions. Specifically, if β < 0, it means that AI can reduce carbon emissions, and conversely, when β > 0, it indicates that AI will promote carbon emissions. In addition, μ_i denotes the province’s fixed effects that do not vary over time to alleviate estimation bias caused by factors such as resource factor endowments and institutions in each province, δ_t denotes the year’s fixed effects that do not vary with individuals, and ε_it is a random error term.

4.2. Variables Selection

(1) Explained variable: carbon emissions (CE). Since the Chinese government has not yet released official data on carbon emissions at the national, regional, or industry level, the authors of this study needed to calculate the carbon emissions for each province. According to the 2006 Guidelines for National Greenhouse Gas Inventories issued by the Intergovernmental Panel on Climate Change (IPCC) of the United Nations and the common practice in academia [66,67], we used the local carbon emissions from fossil energy consumption and cement production to calculate the level of carbon emissions in each region. To be specific, local fossil energy can be divided into seven types, which include coal, diesel, gasoline, kerosene, fuel oil, natural gas, and coke. The formula for calculating regional carbon emissions can be expressed as follows:

$$CE={\displaystyle \sum _{k=1}^7{Q}_k\times C{F}_k}\times C{C}_k\times CO{F}_k\times \frac{44}{12}+QC\times E{F}_{cement}$$

(10)

where CE reflects the level of carbon emissions for each province (as shown in Figure 2), k denotes the specific type of fossil energy, Q_k denotes the consumption of k-th fossil energy, CF_k is the average low calorific value of the k-th fossil energy, CC_k is the carbon content of the k-th fossil energy, COF_k denotes the carbon oxidation factor of the k-th fossil energy, and 44/12 is the mass ratio of carbon dioxide molecules to elemental carbon. QC stands for the quantity of cement production, and EF_cement is the carbon emission factor of cement production.

(2) Core explanatory variable: artificial intelligence (AI). The use of industrial robots is the most intuitive indicator of the level of AI development. Therefore, this paper uses the density of industrial robots installed to reflect AI development. Due to the fact that the International Federation of Robotics (IFR) provides data at the national industry level and the industry classification standards are not consistent with those in China, referring to Qian et al. [44], we matched the industrial robots data released by IFR for China with the seven major industries in the China Labor Statistics Yearbook, including the agricultural industry, mining industry, manufacturing industry, electricity, water and gas supply industry, construction industry, education industry, and other industries, to obtain the density of industrial robots installed in each region. To be specific, the formula for measuring the density of industrial robots installed in each region can be expressed as follows:

$$Robo{t}_{it}={\displaystyle \sum _{j=1}^7\frac{L_{ijt}}{L_{it}}}\times \frac{R_{jt}}{L_{jt}}$$

(11)

where Robot_it is the density of industrial robots installed in region i during period t. L_ijt is the employment number in period t, industry j, and region i. L_it denotes the number of people employed in period t and region i. L_ijt/L_it represents the ratio of people employed in period t, industry j, and region i. R_jt reflects the stock of industrial robots in period t and industry j. L_jt is the national employment number in industry j and period t. R_jt/L_jt denotes the density of industrial robots installed at the national level.

(3) Control variables. In addition to AI development, there are many other factors that affect carbon emissions. In order to alleviate the estimation bias caused by the omitted variables, a series of control variables related to carbon emissions were added to the model, which included economic development (PGDP), industrial structure (IS), environmental regulation (ER), population aging (PA), and the degree of openness (FDI). To be specific, the level of economic development (PGDP) was measured by the per capita GDP in each region, which illustrates that local economic development leads to an increase in energy consumption, thus promoting local carbon emissions. Industrial structure (IS) was measured by the ratio of the output of secondary industries to GDP in each region, which indicates that the higher the level of industrialization, the higher the energy consumption and carbon emissions. Environmental regulation (ER) was measured by the ratio of industrial pollution control investments to GDP in each province, which reflects the impact of local environmental control on regional carbon emissions. Population aging (PA) was measured by the proportion of elderly people aged 65 and above in each region, which reflects the environmental effect of population age structure. The degree of openness (FDI) was measured by the per capita amount of foreign investment utilized in each region, which can help to argue the impact of foreign investment policies on carbon emissions.

4.3. Data Sources and Description

Given the integrity of data and study needs, balanced panel data, which were collected from 30 Chinese provinces during 2006–2019 (excluding Tibet, Hong Kong, Macau, and Taiwan), were used in the empirical analysis to explore the impact of AI on regional carbon emissions. The data for each variable were mainly obtained from the China Statistical Yearbook, China Science and Technology Statistical Yearbook, China Energy Statistical Yearbook, China Environmental Statistical Yearbook, China Population and Labor Statistical Yearbook, EPS database, and the International Federation of Robotics (IFR) for all years. For the missing data of a few variables in some years, this paper uses the linear interpolation method to fill the gap. Using the local consumer price index (CPI) in 2006, all nominal value-based variables in this study were deflated to exclude the impact of price fluctuation. Meanwhile, all variables were treated by using the natural logarithm (ln) to eliminate estimation bias caused by heteroskedasticity. The results of the descriptive statistics for each variable listed in the model are shown in

Table 1. Results of description statistics of all variables.

Symbol	Definition	Obs	Min	Max	Mean	Std. Dev.
lnCE	Per capita carbon emissions	420	0.564	3.361	1.766	0.506
lnRobot	Development level of artificial intelligence	420	2.657	11.875	7.378	1.759
lnPGDP	Per capita GDP	420	8.657	11.694	10.320	0.549
lnIS	Ratio of secondary industry output	420	2.782	4.078	3.789	0.226
lnER	Degree of environmental regulation	420	−1.577	4.597	2.333	0.841
lnPA	Rate of population aging	420	1.699	2.789	2.268	0.210
lnFDI	Foreign direct investment	420	1.615	9.121	6.318	1.341

Note: ln indicates that each variable is taken as a natural logarithm, hereinafter the same.

.

5. Empirical Analysis

5.1. Panel Unit Root Test and Multicollinearity Test

Usually, empirical analysis on balanced panel data faces the spurious regression problem, which can easily lead to estimation bias. Therefore, it is common practice to perform a stationarity test. To be specific, five panel unit root tests in this study were adopted to ensure that all variables were stationary, including the LLC test, IPS test, Breitung test, ADF-Fisher test, and PP-Fisher test. Columns 1–5 in

Table 2. Results of panel unit root tests and VIF test.

Variable	LLC	IPS	Breitung	ADF-Fisher	PP-Fisher	VIF
lnCE	−6.413 ***	−4.719 ***	−1.587 *	148.443 ***	83.342 **	—
lnRobot	−1.833 **	−2.830 ***	−4.826 ***	161.877 ***	182.988 ***	3.02
lnPGDP	−3.041 ***	−7.831 ***	−2.100 **	86.734 **	120.608 ***	3.59
lnIS	−3.733 ***	−1.438 *	−2.263 **	146.368 ***	82.740 **	1.41
lnER	−6.726 ***	−4.273 ***	−1.918 **	130.179 ***	79.176 **	1.71
lnPA	−8.990 ***	−1.815 **	−1.825 **	155.918 ***	77.469 *	1.76
lnFDI	−5.242 ***	−1.918 **	−1.892 **	126.234 ***	101.267 ***	2.39

Note: ***, **, and * represent significance at 1%, 5%, and 10% levels, respectively, hereinafter the same.

report the results of the stationarity tests. The results show that all variables passed the five stationarity tests, proving that all variables in the model were stationary and suitable for empirical analysis. In addition, to identify the issue of multicollinearity, the variance inflation factor (VIF) test was adopted (as shown in Column 6 of Table 2). From the VIF test results, the VIF values of each variable were less than the critical value of 10. Among them, the minimum and maximum values were 1.41 and 3.59 respectively, illustrating that there was no severe multicollinearity among the variables listed in the model.

5.2. Baseline Regression Analysis

To investigate the effect of AI on regional carbon emissions, the authors of this paper used five estimation methods to gradually perform empirical analysis based on the panel econometric model (9). Specifically, first, the pooled ordinary least squares (Pooled OLS) method was used for the full sample, and the findings are listed in Column 1 of

Table 3. Regression results of impact of AI on carbon emissions for whole sample.

Variable	Pooled OLS	FGLS	RE	FE	Two-Way FE
	(1)	(2)	(3)	(4)	(5)
lnRobot	−0.067 ***	−0.031 ***	−0.044 ***	−0.055 ***	−0.172 ***
	(−3.94)	(−2.89)	(−2.77)	(−3.21)	(−6.61)
lnPGDP	0.933 ***	0.838 ***	0.717 ***	0.728 ***	0.372 ***
	(15.72)	(21.43)	(12.45)	(11.87)	(4.80)
lnIS	0.409 ***	0.378 ***	0.172 **	0.083	0.350 ***
	(4.55)	(5.90)	(1.98)	(0.88)	(3.04)
lnER	0.283 ***	0.201 ***	0.059 ***	0.046 ***	0.054 ***
	(10.59)	(11.47)	(4.42)	(3.71)	(4.07)
lnPA	0.011	−0.043	0.013	0.014	0.182 **
	(0.10)	(−0.59)	(0.16)	(0.18)	(2.06)
lnFDI	−0.147 ***	−0.152 ***	−0.061 ***	−0.041 ***	−0.029 **
	(−7.45)	(−10.41)	(−4.10)	(−2.82)	(−2.08)
_Cons	−8.671 ***	−7.536 ***	−5.743 ***	−5.538 ***	−3.047 ***
	(−12.27)	(−17.12)	(−10.17)	(−10.13)	(−4.90)
R2	0.517	—	0.598	0.602	0.668
F/Wald statistic	75.83	854.33	522.64	97.14	39.38
Hausman				63.33
Obs	420	420	420	420	420

Note: t value listed in parentheses. *** and ** denote significance at the 1% and 5% levels.

. Second, to eliminate heteroskedasticity as much as possible, we used the feasible generalized least squares (FGLS) method to estimate the model, and the estimated results can be seen in Column 2 of Table 3. Third, a random effects (RE) method was adopted for the full sample, and the findings are listed in Column 3 of Table 3. Fourth, a fixed-effects (FE) method that only controls an individual effect was used for the full samples, and the findings are listed in Column 4 of Table 3. Finally, we used the two-way fixed-effect (Two-way FE) method, which controls both the individual effect and the time effect, to estimate the model, and the regression results can be seen in Column 5 of Table 3.

First, we analyzed the impact of AI development on regional carbon emissions. From the estimated findings listed above, it can be seen that AI has a significant inhibitory effect on carbon emissions, which tentatively verifies Hypothesis 1 (H1) mentioned in this study. Specifically, according to the coefficients listed in Column 5, there is a significant negative link between AI and carbon emissions; namely, each percentage point rise in AI is associated with a 0.172 percentage point fall in carbon emissions, meaning that AI development promotes the development of information technology and improves the utilization efficiency of energy, especially the widespread use of clean energy, thus contributing to carbon emission reduction. This result is highly similar to the study of Ma et al. [68], who argued that clean technology innovation can reduce carbon emissions.

Next, we further analyzed the effects of each control variable listed in Table 3 on regional carbon emissions. For economic development (lnPGDP), the estimated coefficient was significantly positive in all models. Taking Column 5 as an example, the coefficient of economic development was 0.372 at the 1% significance level. This result indicates that as economic output increases rapidly, local carbon emissions will significantly rise. For industrialization (lnIS), the estimated coefficient was also significantly positive in most models, illustrating a significant positive link between industrialization and regional carbon emissions. For environmental regulation (lnER), the estimated coefficient was significantly positive in all models. In Column 5, the estimated coefficient of environmental regulation is equal to 0.054 at the significance level of 1%, demonstrating that environmental regulation is positively correlated with carbon emissions. In terms of population aging (lnPA), except for the two-way FE model, the estimated coefficients were not significant in other models. For foreign direct investment (lnFDI), the estimated coefficient was significantly negative in all models, indicating that FDI can effectively reduce carbon emissions. The feasible explanation is that FDI brings advanced production technologies and management experience, which can reduce environmental control costs, and thus, promotes carbon emission reduction.

In summary, the above results show that AI development and FDI can reduce carbon emissions, while economic development, industrialization, and environmental regulation can promote carbon emissions, and the deterioration impact of population aging on local carbon emissions is not significant on the whole.

5.3. Robustness Analysis

To ensure the robustness and reliability of the findings listed in baseline regression and to avoid estimation bias, the authors of this study used four methods to test the robustness. The first approach was to adjust the measurement of the explained variable. Considering the impact of China’s household registration system on local carbon emissions, the authors of this paper further treated the carbon emissions based on the number of registered provincial households as the population and then substituted them into the empirical model for analysis. The second approach was to change the key explanatory variable. Industrial robots are integrators of AI technology. The scale effect generated by industrial robots from their introduction and installation to their application has a certain time lag. Thus, we used the density of industrial robots installed with a one-period lag to re-estimate the model. The third approach was to adjust the time interval of the study samples. To eliminate estimation bias caused by sample extremes, the authors of this study reconstructed the empirical samples and performed regression estimation by removing the data from 2006 and 2019. The last approach was to modify the estimation method for the model. In order to alleviate endogeneity, the model was estimated by using the system generalized method of moments (System GMM), which introduced the lagged terms of the variables as instrumental variables. The results of the four robustness tests are displayed in

Table 4. Robustness test results of impact of AI on carbon emissions.

Variable	Adjusted Explained Variable	Adjusted Explanatory Variable	Adjusted Sample Interval	Adjusted Estimation Method
	(1)	(2)	(3)	(4)
L.lnCE				0.816 ***
				(18.69)
lnRobot	−0.150 ***	−0.166 ***	−0.147 ***	−0.052 ***
	(−6.09)	(−6.59)	(−5.68)	(−3.45)
lnPGDP	0.324 ***	0.335 ***	0.565 ***	0.077 **
	(4.39)	(4.26)	(6.13)	(2.11)
lnIS	0.282 ***	0.264 **	0.103	0.271 ***
	(2.58)	(2.29)	(0.87)	(3.18)
lnER	0.049 ***	0.039 ***	0.038 ***	0.014 **
	(3.90)	(2.97)	(2.91)	(2.33)
lnPA	0.026	0.159 *	0.134	0.145 ***
	(0.31)	(1.81)	(1.53)	(5.54)
lnFDI	−0.028 **	−0.025 *	−0.030 **	−0.028 **
	(−2.10)	(−1.79)	(−2.12)	(−2.57)
_Cons	−2.052 ***	−2.273 ***	−3.818 ***	−1.155 ***
	(−3.47)	(−3.46)	(−5.09)	(−3.05)
Province FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.687	0.610	0.645	—
F statistic	42.87	29.76	33.40	—
AR(1)				p < 0.000
AR(2)				p = 0.243
Sargan				p = 0.664
Obs	420	390	360	390

Note: t value listed in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

.

According to the results in Table 4, the robustness and reliability of the baseline regression are further confirmed. To be specific, Column (1) listed in Table 4 presents the findings when changing the measurement of the explained variable. The regression coefficient of AI on carbon emissions was −0.150 at the level of 1%, indicating that AI effectively lowers regional carbon emissions. Column (2) of Table 4 reports the estimation results when adjusting the indicator of the key explanatory variable. The coefficient of AI on carbon emissions was still robust at the significance level of 1%, which illustrates that AI development can promote local carbon emission reduction. Column (3) of Table 4 reports the estimation results when adjusting the time interval for the study samples. The coefficient and significance of AI development are highly similar to the results in Column (1) of Table 4, indicating that AI development is significantly and negatively related to carbon emissions. Column (4) of Table 4 reports the estimation results when modifying the estimation method, namely System GMM. Among them, the correlation tests for residual series show that there was only a first-order serial correlation in the model, and the validity of all instrumental variables used in the estimation was confirmed by the Sargan test. In addition, the coefficient of AI on carbon emissions was also negative at the level of 1%, proving that AI development helps to reduce carbon emissions.

5.4. Instrumental Variables Regression Analysis

Considering the possibilities of two-way causality and missing variables between AI and carbon emissions, this study further adopted the instrumental variable (IV) method to alleviate the endogeneity issues. Specifically, we selected the interaction term between the number of Internet users in 1998 and the nationwide number of computers produced in 2005–2018 as IV in the estimation. There are two reasons for this practice. First, choosing the number of Internet users as IV met the principle of relevance. Because AI is developed on the basis of the popularization of computer and Internet technology, areas with a large number of Internet users in history are likely to be areas with rapid AI development. Second, compared to the development of AI, the number of Internet users will not have a direct impact on carbon emissions. Specifically, the impact of the historical number of Internet users on carbon emissions is minimal. Since the sample belonged to the panel data, it was necessary to convert the number of Internet users in 1998 into a panel indicator. Therefore, we constructed the interaction term between the number of Internet users and the nationwide number of computers produced in the previous year as IV in the estimation.

Table 5. IV regression results of impact of AI on carbon emissions.

Variable	(1)	(2)	(3)	(4)
	Stage One	Stage Two	Stage One	Stage Two
	lnRobot	lnCE	lnRobot	lnCE
IV_lnRobot	0.367 ***		0.245 ***
	(6.70)		(4.75)
lnRobot		−0.154 *		−0.715 ***
		(−1.75)		(−3.22)
Control Variable	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes
Year FE	No	No	Yes	Yes
Kleibergen-Paap rkLM		10.53		9.87
Cragg-Donald Wald F		150.82		75.47
F statistic		16.46		3.83
Obs	420	420	420	420

Note: t value listed in parentheses. *** and * denote significance at 1% and 10% levels, respectively.

reports the two-stage IV estimation (2SLS) results of AI on carbon emissions. Specifically, Columns (1) to (2) show the 2SLS results that only control for province FE, while Columns (3) to (4) show the 2SLS results when controlling for both province FE and year FE. From the regression results, it can be seen that there was a significant positive link between the interaction term (IV_lnRobot) and AI in the first stage, which can meet the correlation requirement. Meanwhile, the estimation results not only pass the weak IV test but also reject the underidentification test, proving that the IV estimation is valid. In addition, AI had a significant negative impact on carbon emissions in the second-stage regression, demonstrating that the emission reduction effect of AI is still robust by using the panel IV estimation.

5.5. Regional Heterogeneity Analysis

The findings of preliminary regression show a significant negative link between AI and carbon emissions across the whole sample. However, China has a vast territory with large differences in geographical location, resource endowment, and economic policies among different provinces. The problem of spatial imbalance in economic development is also becoming increasingly prominent, thus resulting in East–West and North–South differences in space for environmental pollution (as shown in Figure 2). In this context, we determined if there is a significant spatial heterogeneity effect of AI development on carbon emissions among different regions. Therefore, we first split the full sample into three sub-samples, which included the Eastern area, the Central area, and the Western area, in line with the criteria of the National Bureau of Statistics (NBS) in China. To be specific, the East and West contain 11 provinces respectively, while the Center contains 8 provinces. Then, the whole sample was further divided into two sub-samples, namely the Northern and Southern regions, and each region contains 15 provinces. The results of the regional heterogeneity test are reported in

Table 6. Regional heterogeneity results of impact of AI on carbon emissions.

Variable	East	Center	West	North	South
	(1)	(2)	(3)	(4)	(5)
lnRobot	−0.151 ***	−0.033	−0.101 ***	−0.320 ***	0.051 *
	(−3.24)	(−0.68)	(−2.77)	(−6.97)	(1.94)
lnPGDP	0.709 ***	0.167	−0.541 ***	0.566 ***	0.692 ***
	(6.12)	(1.19)	(−4.56)	(4.17)	(7.46)
lnIS	−0.022	0.066	0.721 ***	0.479 ***	0.119
	(−0.12)	(0.39)	(3.56)	(2.79)	(0.91)
lnER	0.055 ***	0.073 ***	0.023	0.078 ***	0.021
	(3.31)	(3.29)	(0.99)	(3.71)	(1.61)
lnPA	0.236 **	−0.283	−0.316 *	−0.003	−0.260 ***
	(2.26)	(−1.56)	(−1.80)	(−0.02)	(−2.72)
lnFDI	−0.120 ***	−0.056 *	0.005	−0.021	0.038 *
	(−4.27)	(−1.90)	(0.28)	(−1.10)	(1.91)
_Cons	−4.555 ***	0.414	4.604 ***	−4.290 ***	−5.878 ***
	(−4.35)	(0.33)	(4.38)	(−3.74)	(−8.24)
Province FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
R2	0.545	0.812	0.871	0.718	0.792
F statistic	7.82	19.44	44.18	23.63	35.32
Obs	154	112	154	210	210

Note: t value listed in parentheses. ***, **, and * denote significance at the levels of 1%, 5%, and 10%, respectively.

.

In terms of the East, Center, and West, the results listed in Table 6 confirm a notable spatial heterogeneity in how AI affects carbon emissions. To be specific, the findings for the Eastern region are reported in Column 1 of Table 6. Among them, the coefficient of AI on carbon emissions was −0.151 at the level of 1%, proving that AI helps to promote carbon emission reduction in the Eastern region. Column 2 listed in Table 6 presents the regression results for the middle area. Among them, the impact of AI on carbon emissions was not significant, illustrating that there is unreasonable development of AI technology in the middle, and the energy utilization efficiency and pollution treatment capacity need to be further improved, resulting in its inhibitory effect on carbon emissions not being fully demonstrated. In Column 3 of Table 6, the estimation results of the Western region are presented. Notably, the impact of AI development on carbon emissions was still negative at the significance level of 1%, but the scale was smaller than it in the East, meaning that the effect of AI on carbon emission reduction in the Western region is significantly lower than that in the Eastern region.

For the North and South, the geographic heterogeneity in the effect of AI on carbon emissions is further shown in the results listed in Table 6. To be specific, the findings for the Northern region are reported in Column 4 of Table 6. Among them, the coefficient of AI on carbon emissions was −0.320 at the 1% significance level, implying that AI development helps to promote carbon emission reduction in the Northern region. The findings for the Southern region are reported in Column 5 of Table 6. Notably, the AI regression coefficient was equal to 0.051 at the 10% significance level, exhibiting a strongly positive link between AI development and carbon emissions. In other words, AI can promote carbon emissions in the Southern region. The possible explanation is that AI theoretically can enhance firms’ green innovation capacity and reduces carbon emissions, but AI development itself also requires a large amount of energy consumption, especially electricity, thereby exacerbating the carbon emission crisis. In view of the huge difference in energy structure between the South and North in China, the South mainly relies on hydroelectric power generation with strong seasonality. To alleviate the huge pressure from energy consumption caused by AI development, it is inevitable that fossil fuel consumption will increase in the South, which will lead to a sharp rise in carbon emissions.

5.6. Transmission Mechanism Analysis

As mentioned above, the benchmark regression and a series of robustness tests have confirmed that AI development can promote carbon emission reduction. However, what is the channel by which AI affects carbon emissions? To thoroughly answer this question, this study further analyzed the impact mechanism of AI on carbon emissions. According to the theoretical analysis mentioned in this study, AI development mainly affects carbon emissions through two channels, namely the energy structure effect and the technological innovation effect. To further argue the existence of these two transmission channels, the authors of this paper conducted a mechanism analysis from two aspects. Firstly, we confirmed the existence of an energy structure channel. Specifically, we selected the ratio of coal consumption to energy consumption as the proxy indicator for energy structure and constructed an interaction term between energy structure and AI, which was added to the empirical econometric model (9) to test the existence of the energy structure channel. Meanwhile, based on the median of the national energy structure in 2019, the full sample was divided into two regions, namely a high-coal-consuming region and a low-coal-consuming region, to further examine the extent to which AI affects carbon emissions through the energy structure channel. Secondly, this paper confirms the existence of a technological innovation channel. To be specific, the authors selected the number of patents authorized in each province as the proxy variable for technological innovation and constructed an interaction term between technological innovation and AI for empirical analysis, which was added to the panel econometric model (9) to test the existence of the technological innovation channel. Moreover, based on the median of national technological innovation in 2019, the full sample was classified into two subsamples, namely a high technological innovation region and a low technological innovation region, to further show the extent to which AI affects carbon emissions through the technological innovation channel. The findings of the mechanism test are presented in

Table 7. Mechanism regression results of impact of AI on carbon emissions.

Variable			Energy Structure		Technological Innovation
	(1)	(2)	(3)	(4)	(5)	(6)
			High	Low	High	Low
lnRobot	−0.280 ***	0.089 **	−0.153 **	−0.174 **	−0.198 ***	−0.080
	(−12.92)	(2.14)	(−1.98)	(−2.25)	(−2.68)	(−1.11)
lnRobot × lnES	0.037 ***
	(15.15)
lnRobot × lnTI		−0.018 ***
		(−7.67)
_Cons	−2.408 ***	−3.651 ***	−3.936 **	−3.020 *	−3.975 *	−2.861 *
	(−4.90)	(−6.26)	(−2.35)	(−1.76)	(−1.75)	(−1.89)
Control variable	Yes	Yes	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes
R2	0.795	0.714	0.723	0.649	0.619	0.689
F statistic	71.93	46.19	38.19	20.09	18.28	33.42
Obs	420	420	210	210	210	210

Note: t value listed in parentheses. ***, **, and * denote significance at the level of 1%, 5%, and 10%, respectively.

.

According to the results listed in Column 1 of Table 7, the interaction term between the energy structure and AI (lnRobot × lnES) was positive at the 1% significance level. A feasible explanation is that as coal usage rises, AI development will consume a large amount of energy, which will exacerbate regional energy consumption tension and eventually lead to higher carbon emissions. From the results in Column 2 of Table 7, the interaction term between technological innovation and AI (lnRobot × lnTI) was negative at the 1% significance level. This means that as technological innovation improves, AI will greatly promote production efficiency and energy utilization efficiency for enterprises and improve their pollution control abilities, ultimately lowering carbon emissions. The findings listed in Columns 3 and 4 of Table 7 reflect the impacts of AI on carbon emissions in samples with high and low coal consumption, respectively. The results show that the smaller the ratio of coal energy consumption, the more obvious the transformation effect of the local energy structure brought by AI development, and the higher the effect of carbon emission reduction. Meanwhile, the findings listed in Columns 5 and 6 of Table 7 exhibit the effects of AI on carbon emissions in the samples with high and low levels of technological innovation, respectively. The findings reveal that the higher the level of technological innovation, the more effects there are of energy conservation and environmental control caused by AI development, and the higher the effect of carbon emission reduction. Thus, the findings provide strong empirical evidence for Hypothesis 2 (H2) and Hypothesis 3 (H3).

In summary, the above mechanism analysis results indicate that AI development can affect regional carbon emissions through mechanisms such as energy structure and technology innovation, i.e., the carbon reduction effect of AI development is higher in regions with a low share of coal consumption and high technology innovation.

6. Conclusions and Policy Recommendations

While the advent of AI technology has had a great impact on the global economy and human life, it has also brought new hope and opportunities for environmental protection. In this context, it is worth noting whether AI development can promote regional carbon emission reduction. Based on Chinese robot use data released by the IFR, the authors of this paper constructed balanced panel data for 30 provinces in China from 2006 to 2019 and empirically tested the effect of AI on carbon emissions by using a two-way fixed-effect model. The results show that AI development has significantly lowered carbon emissions in China. To further identify the impact of AI on carbon emissions, the authors also conducted a series of robustness tests and IV analysis by replacing the dependent variable, replacing the key explanatory variable, adjusting the sample intervals, and considering endogeneity, confirming that the baseline results are still robust. Furthermore, mechanism analysis revealed that AI development mainly reduces carbon emissions by promoting energy structure transformation and technology innovation. The lower the dependence on fossil energy and the higher the technological innovation, then the better the carbon reduction effect of AI development in a region. In addition, the heterogeneity analysis showed that the reduction effect of AI development on carbon emissions is highest in the East, followed by the West, and not significant in the central. Meanwhile, AI development has a higher carbon reduction effect in the North than in the South. In general, the above results on the carbon reduction effect of AI are not only related to the transformation direction of industrial development in the context of a digital economy but also have notable practical benefits for China’s pursuit of a carbon peak and carbon neutrality. Therefore, the following policy recommendations can be proposed:

Firstly, the Chinese government should play a better role in promoting the development of the digital economy, actively expanding the breadth and depth of AI applications, and stimulating enterprises to achieve intelligent transformation. On the one hand, it is necessary to accelerate the construction of a collaborative innovation system for robot use and drive innovation for the supply of industrial robot major core technology to meet the demand for the intelligent transformation of enterprises so as to foster a good ecological environment for promoting the digitalization and intelligence of Chinese industries while helping the rapid growth of the AI industry in China. On the other hand, it is necessary to rely on the development dividend of the digital economy to promote the deep integration of AI and related industries and use industrial robots as a starting point to reverse the development modes of high-energy-consuming, high-polluting, and high-emission industries in various regions. This will also achieve effective synergistic supervision of carbon emissions between the government and enterprises, and thus, help regions achieve low-carbon development while reducing carbon emissions.

Secondly, the government should effectively formulate differentiated development policies to narrow the gap in AI development among regions. In view of the regional heterogeneity of the impact of AI development on carbon emissions, and therefore, in the process of further promoting AI applications, a gradual implementation strategy should be followed, focusing on promoting the carbon emission reduction of AI applications in the Eastern and Northern regions. At the same time, local governments should solve the problem of unequal development among regions caused by the differences in resource endowments, strive to eliminate the regional digital gap, coordinate and promote the development of new infrastructure in rural and poor areas, dissolve regional digital barriers, and focus on building a nationwide industry data information platform to provide basic support for AI application and empowering real industries at the national level.

Thirdly, the government should actively promote the development of energy digitalization and accelerate the clean transformation of energy structure. On the one hand, local governments should actively promote the development of clean energy based on a digital economy, vigorously carry out the development and utilization of clean energy data, continuously improve the efficient linkage of digital information, stimulate the market vitality of clean energy, and build a clean energy digital management platform as soon as possible to realize the linkage between the upstream and downstream industry chains of clean energy, continuously expand the application scenarios of clean energy, timely meet new market demands, and accelerate the replacement of traditional energy. On the other hand, the Chinese government should accelerate the digital development of the energy sector, realize accurate measurements and prediction with information technology, improve energy usage efficiency, realize efficient energy dispatch and utilization, vigorously promote the construction of energy digital control centers in high energy-consuming industries, continuously improve the modernization of the energy supervision system and supervision capacity, and comprehensively improve energy use efficiency.

Finally, by using AI technology, the government should focus on gathering innovation resources and strengthening innovation capabilities to gradually realize the shift from factor-driven growth to innovation-driven growth. Specifically, with the continuous development of the digital economy, the Chinese government should use digital technologies to achieve a low-carbon transformation of production. In terms of how to achieve the goal of carbon peak and carbon neutrality, the government needs to adjust the structure of fiscal expenditure and direct the flow of funds to basic research fields, such as energy conservation and emission reduction, by increasing the expenditure on science and technology, and thus, driving the formation of the innovation preferences of local enterprises to provide a good foundation for green and low-carbon transformation. In addition, the government should enhance the accumulation of innovative human capital, improve its suitability with AI technology, and use technological and policy means to squeeze out innovative human capital from high-energy-consuming and high-polluting industries or state-owned companies, meaning that high-quality human capital can be released more in the emerging industries using green technology as a key breakthrough for independent innovation.

Although this paper explores the mechanisms and effects of AI on carbon emissions and draws rich conclusions, there are still two shortcomings. Firstly, due to the lack of data on the application of AI technology at the enterprise level, this study failed to examine the impact of AI technology on carbon emissions at a micro-level. Secondly, the authors only selected data at the provincial level in China for empirical analysis, and the conclusions may only apply to China and not to other countries. In the future, we will begin to construct the micro-data of AI technology at the enterprise level, and then explore the carbon reduction mechanism of AI technology at the micro-level. At the same time, we will use cross-country panel data to analyze the carbon emission reduction effect of AI technology through the energy structure, economic development, and industrial structure in different sample countries.