Artificial intelligence promotes pollutant and carbon emission reduction by enhancing the overall green energy efficiency in cities, where the core logic of this intermediary pathway lies in incorporating energy consumption and pollutant emissions into a comprehensive evaluation framework. This systematically reflects how energy inputs influence economic output and alter environmental costs through a collaborative optimization framework. Improvements in overall green energy efficiency signal the achievement of dual goals: enhancing energy utilization efficiency while reducing pollution emissions, all while ensuring the continuation of economic activities (8, 25). The empowering effect of artificial intelligence on this efficiency is distinctly directed. On one hand, real-time monitoring systems equipped in industrial robots can accurately collect energy consumption and emission data from various locations, leveraging algorithms to identify high-energy-consuming segments at the regional level and promoting the flow of energy resources towards efficient utilization sectors (26, 27). On the other hand, machine learning algorithms can dynamically optimize production sequences and adjust production parameters, thereby eliminating redundant energy consumption and enhancing the overall green energy efficiency of cities.

Meanwhile, a critical question arises: How does enhancing overall green energy efficiency specifically promote pollutant and carbon emission reduction? At the urban level, energy-intensive regions often exhibit significantly higher carbon and pollutant emissions per unit of GDP due to low energy utilization efficiency compared to more efficient regions. When overall green energy efficiency improves, cities can achieve greater economic value with the same energy input, or consume less energy for the same economic output; the former reduces energy intensity per unit of GDP, while the latter directly decreases carbon and pollutant emissions from fossil fuel combustion, collectively driving down the intensity of urban pollution emissions (28). In addition, the enhancement of overall green energy efficiency can alleviate the rigid constraints from energy consumption to pollution emissions, allowing cities to pursue economic growth objectives without relying on energy-intensive and high-emission production modes. This results in a synergistic advancement of economic development and environmental governance. Therefore, combining the empowering role of artificial intelligence development on overall green energy efficiency with the driving effect of this efficiency on pollutant and carbon emission reduction.

We propose Hypothesis 2: The development of artificial intelligence can promote urban pollutant and carbon emission reduction through the enhancement of overall green energy efficiency.

Industrial structure serves as an important intermediary pathway through which artificial intelligence drives urban pollutant and carbon emission reduction. The core logic lies in the optimization direction of the industrial structure shifting from labor-intensive to technology-intensive, and from a focus on the secondary sector to a collaborative emphasis on the tertiary sector. An increase in the proportion of the tertiary sector can effectively reduce regional pollution emission intensity. Artificial intelligence facilitates the adjustment of industrial structure through the following pathways: First, industrial robots replace jobs in energy-intensive manufacturing, prompting labor to shift towards the low-pollution tertiary sector, thereby directly increasing the share of the tertiary sector. Second, digital transformation optimizes manufacturing production efficiency, phases out outdated energy-intensive capacities, and generates demand for productive services, further promoting the adjustment of the industrial structure towards high technology and low emissions (29).

When the industrial structure is optimized, the low energy consumption characteristics of the tertiary sector will directly lower overall pollution levels in the city. At the same time, the increased share of high-tech manufacturing within the secondary sector also reduces industrial pollution emissions through the optimization of production processes, collectively lowering the intensity of urban pollution emissions. Moreover, the adjustment of industrial structure can alleviate the conflict between economic growth and pollution emissions, allowing cities to pursue economic growth objectives without the need for high-emission production modes (30).

Therefore, we propose Hypothesis 3: The development of artificial intelligence can optimize the industrial structure by promoting the increased share of the tertiary sector, thereby indirectly facilitating urban pollutant and carbon emission reduction.

Green technology innovation is the third key intermediary pathway through which artificial intelligence drives urban pollutant and carbon emission reduction, primarily manifested in the upgrades of pollution control technologies and breakthroughs in low-carbon production technologies, as indicated by the volume of green invention patent applications. Such technologies can reduce pollution emission intensity at the source and enhance treatment efficiency, thereby promoting regional environmental governance (8, 31). The driving pathways of artificial intelligence have two facets. First, the energy consumption and emission data accumulated by industrial robots provide empirical evidence for the research and development of green technologies, assisting research institutions in identifying technological bottlenecks (32). Second, machine learning algorithms accelerate the research and development process, shortening the timeframe for green technologies from laboratory experimentation to industrialization, while also broadening the application scenarios of these technologies in energy-intensive industries.

How does green technology innovation facilitate pollutant and carbon emission reduction? High-pollution areas are often constrained by traditional technologies, making it challenging to effectively reduce emissions. In contrast, green technology innovation can overcome these barriers. An enhancement in green technology innovation capacity is often accompanied by an increase in the number of green invention patents and accelerated technology transfer. On one hand, advanced pollution control technologies significantly enhance treatment efficiency and reduce end-of-pipe emissions; on the other hand, low-carbon production technologies can reduce fossil fuel consumption at the source, thereby decreasing carbon emissions. Together, these two factors contribute to the dual reduction of urban pollutant and carbon emission reduction. Furthermore, green technology innovation also elevates the overall technological level of enterprises within the region, thereby amplifying the impact of pollutant and carbon emission reduction (33).

In summary, we propose Hypothesis 4: The development of artificial intelligence promotes green technology innovation by driving an increase in the number of green invention patent applications, thereby indirectly facilitating urban pollutant and carbon emission reduction.

This paper aims to investigate the impact of artificial intelligence development on urban pollutant and carbon emission reduction. Currently, most studies on policy evaluation primarily employ the Difference-in-Differences method, Synthetic Control method, and Regression Discontinuity Design. However, these traditional causal inference models have several limitations in application. For instance, the Difference-in-Differences method requires that the treatment and control groups exhibit the same trend of change, meaning that the study samples must satisfy the parallel trends assumption (8, 34). Although the Synthetic Control method avoids the challenge of finding treatment and control groups with the same trend by constructing a control group, it is primarily applicable in “one-to-many” scenarios. The Regression Discontinuity Design not only has high data requirements but also can only estimate the local average treatment effect near the cutoff point. To address the shortcomings of traditional models, some scholars have combined machine learning with causal inference, with Double Machine Learning serving as a typical representation of this approach (35). Compared to traditional policy evaluation models, Double Machine Learning has unique advantages in variable selection and model estimation: it uses regularization algorithms to automatically select high-dimensional control variables, addressing multicollinearity and the curse of dimensionality that plague traditional models, thus improving predictive accuracy (36, 37). On the other hand, Double Machine Learning models can handle nonlinear relationships between variables, thus preventing errors in model specification. Based on this, this paper employs a Double Machine Learning model to assess the impact of artificial intelligence development on urban pollutant and carbon emission reduction. Specifically, the partially linear Double Machine Learning model is as follows:

In this framework, i represents the city, and t denotes the year. E I it indicates the reverse level of urban pollutant and carbon emission reduction in the city, while A I it serves as the explanatory variable representing the development of artificial intelligence. The parameter θ 0 denotes the policy effect, which reflects the influence of artificial intelligence development on the promotion of pollutant and carbon emission reduction at the urban level. Additionally, X it represents potential high-dimensional control variables, and g ( X it ) illustrates the potential nonlinear relationship between these high-dimensional control variables and urban pollutant and carbon emission reduction. Since this relationship is largely unknown, we estimate its form using machine learning algorithms as g ̂ ( X it ) The term U it represents the error term, with a conditional mean of zero. When employing g ̂ ( X it ) , it is necessary to introduce the Neyman orthogonality concept to construct the following matrix:

From Formula 1, the estimated value of the policy effect θ ̂ 0 can be derived as follows:

Here, n represents the total sample size. Based on the estimated value of the policy effect, the estimation bias of this policy effect can be further examined:

Here, ( 1 n ∑ i ∈ I , t ∈ T A I it 2 ) − 1 1 n ∑ i ∈ I , t ∈ T A I it U it follows a normal distribution with a mean of zero. However, for ( 1 n ∑ i ∈ I , t ∈ T A I it 2 ) − 1 1 n ∑ i ∈ I , t ∈ T A I it [ g ( X it ) − g ̂ ( X it ) ] , it is necessary to introduce a regularization term when estimating the specific functional form of g ( X it ) using machine learning algorithms. While this helps avoid excessive variance in g ̂ ( X it ) , it also results in a lack of unbiasedness. Consequently, θ 0 struggles to converge to θ ̂ 0 .

To accelerate the convergence of θ 0 towards θ ̂ 0 , this paper constructs an auxiliary regression:

Here, m ( X it ) represents the regression function of the policy variable on the high-dimensional control variables, and its specific form m ̂ ( X it ) also needs to be estimated using machine learning algorithms. V it denotes the error term, with a conditional mean of zero.

The specific procedure is as follows: First, we estimate m ̂ ( X it ) using machine learning algorithms and obtain the residuals denoted as V ̂ it = A I it − m ̂ ( X it ) . Next, we estimate g ̂ ( X it ) using machine learning algorithms, transforming the main regression into E I it − g ̂ ( X it ) = θ 0 A I it + U it . Finally, we treat V ̂ it as an instrumental variable for A I it in regression, allowing for the acquisition of unbiased coefficient estimates.

Through the aforementioned steps, this paper can still obtain an unbiased estimate of treatment effects using Double Machine Learning, even in the context of unknown functional forms of covariates. Moreover, the impact of artificial intelligence development on urban pollutant and carbon emission reduction exhibits heterogeneous effects across different geographical locations and cities with varying characteristics. To account for these heterogeneous characteristics in the model estimation, this paper establishes a more general interaction model (38):

The estimation methods for the relevant parameters are consistent with those of the partially linear model.

This paper utilizes panel data from 282 prefecture-level and above cities in China covering the years 2009 to 2023, with the data sources outlined as follows: Firstly, the core variable data primarily originates from publicly available authoritative statistical yearbooks such as the “China Urban Statistical Yearbook,” “China Environmental Statistical Yearbook,” “China Science and Technology Statistical Yearbook,” and “China Energy Statistical Yearbook.” The data on industrial robot installation density is supplemented with industrial development reports and relevant white papers published by the industrial and information departments of each city. The CO₂ emission data required for the pollutant and carbon emission reduction index is sourced from relevant databases of the Global Carbon Project and carbon emission inventories published by provincial ecological and environmental departments. The PM₂.₅ concentration data is compiled using information from the China Air Quality Online Monitoring Analysis Platform, in conjunction with publicly available data from urban environmental monitoring stations. Input–output data necessary for measuring green total factor energy efficiency is derived from the aforementioned statistical yearbooks and municipal statistical bulletins. Secondly, the tool variable data for “the minimum distance between prefecture-level cities and backbone optical cable cities” is obtained from the national geographic information public service platform released by the National Administration of Surveying, Mapping and Geo-information, which provides geographic coordinate information for the cities. The list of backbone optical cable cities is determined based on the “National Backbone Optical Cable Network Construction Planning” released by the Ministry of Industry and Information Technology and related reports in the telecommunications industry. Geographic Information System (GIS) is then utilized to calculate the minimum distance between each city and the backbone optical cable cities, while cross-validation is performed using publicly available network layout information from telecommunications operators. Thirdly, the data on green invention patent applications is sourced from the Patent Information System of the National Intellectual Property Administration. Data on the proportion of the tertiary sector and other aspects of industrial structure, as well as financial development level data, are extracted from the “China Urban Statistical Yearbook” and “China Financial Statistical Yearbook,” ensuring the accuracy and authority of the data (Table 1).

In light of the significant regional heterogeneity among prefecture-level cities in China regarding the foundation of the artificial intelligence industry, energy consumption structure, and environmental governance capacity, the municipalities directly under the central government (Beijing, Shanghai, Tianjin, and Chongqing), as national-level central cities, exhibit a far higher installation density of industrial robots and investment in green technology research and development compared to ordinary prefecture-level cities. Furthermore, they possess independent policy authority in areas such as pollutant and carbon emission reduction pilot programs and fiscal resource allocation, which creates distinct stratification differences in core variable characteristics and policy environments compared to other prefecture-level cities. Including these municipalities in the sample could potentially lead to biased estimation results.

Based on this, this paper excludes the aforementioned four municipalities in the empirical analysis, retaining the remaining 278 prefecture-level cities as the research sample to avoid sample selection bias caused by special administrative levels, thereby enhancing the representativeness of regression results for ordinary prefecture-level cities and the accuracy of core causal effect estimates. The results are presented in Table 3 (1). The regression results indicate that after excluding the municipalities directly under the central government, the impact coefficient of artificial intelligence development on urban pollutant and carbon emission reduction exhibits some variation; however, its positive effect remains significant at the 1% level. This indicates that the baseline regression results exhibit strong robustness.

Considering that outliers in the regression sample may cause instability in the regression coefficients, resulting in deviations from the true relationships, this paper thus applies 1 and 5% winsorization to all variables, with the regression results presented in Columns (2) and (3) of Table 3. Column (2) shows the regression results after 1% winsorization, while Column (3) presents the results after 5% winsorization. It can be observed that after the removal of outliers, although the regression coefficients exhibit a slight decrease, they remain significant, further corroborating the robustness of the conclusions presented in this paper.

To eliminate potential biases introduced by the specification of the double machine learning model, this paper conducts robustness checks on the baseline regression results from multiple perspectives: (1) adjusting the division ratio of the training and testing sets from the original 1:4 to 1:2 and 1:7, examining the impact of the division method on the results; (2) replacing the original random forest prediction algorithm with lasso regression, gradient boosting, and support vector machines to compare the effects of different machine learning algorithms on the research conclusions.

Columns (1) to (5) of Table 4 present the regression results after redefining the double machine learning model. The results indicate that regardless of adjustments made to the sample split ratio or the selected machine learning algorithm, the effect of artificial intelligence development on urban pollutant and carbon emission reduction remains significant. However, the estimated values of the policy effects are influenced. These findings indicate that the research conclusions exhibit robustness under different model specifications, thereby further validating the reliability of the baseline regression results.

The distribution of artificial intelligence development levels also exhibits a degree of non-randomness, which is often influenced by various resource endowment factors, such as the thickness of the regional industrial base, the clustering degree of the manufacturing industry, the intensity of technological research and development investments, and the completeness of high-end equipment supply chains. Cities with a strong manufacturing base are more likely to attract industrial robots, and such cities also have significantly better foundations for pollutant and carbon emission reduction compared to other cities. This may introduce endogeneity issues due to “selection bias,” thereby interfering with the precise identification of the causal effects between artificial intelligence and pollutant and carbon emission reduction.

Based on this, this paper draws on the instrumental variable design approach of double machine learning proposed by Chernozhukov et al. (40), and constructs a partially linear instrumental variable model using double machine learning, aiming to achieve an unbiased estimate of the causal effect of artificial intelligence on pollutant and carbon emission reduction by isolating the endogenous part of the core explanatory variable.

Here, I V it represents the instrumental variable. In this study, the minimum distance (unit: km) between prefectural-level cities and optical cable backbone cities is employed as the instrumental variable corresponding to the core explanatory variable, namely the installation density of industrial robots. On one hand, the minimum distance is jointly determined by the geographical locations of the cities and the layout of the national optical cable backbone network, which serves as a geographic locational characteristic variable. The location of optical cable backbone cities is primarily based on the overall planning of the national communication network and has no direct correlation with the pollutant and carbon emission reduction levels of individual prefectural-level cities or their demand for industrial robot applications. Furthermore, the geographical distance is not affected by endogenously reverse effects from urban economic activities, environmental governance policies, or similar factors, thereby strictly meeting the exogeneity assumption of instrumental variables.

On the other hand, optical cable backbone cities serve as core hubs for regional digital information transmission, and the distance to these cities directly affects the accessibility of digital infrastructure in prefectural-level cities: the closer the distance, the more developed the digital infrastructure, which is conducive to the networking operation, data interaction, and intelligent application of industrial robots, thereby significantly affecting the installation density of industrial robots. Meanwhile, this geographical distance itself does not directly affect the pollutant and carbon emission reduction index, but rather indirectly impacts the pollutant and carbon emission reduction processes by influencing the installation density of industrial robots, fully satisfying the relevance and exclusion assumptions of instrumental variables.

Column (6) of Table 4 indicates that the estimated coefficient of the core explanatory variable, artificial intelligence development, remains significantly negative for the dependent variable, the pollutant and carbon emission reduction index, which is consistent with the direction of the baseline regression results, thus validating the effectiveness of the instrumental variable model estimates.

Artificial intelligence influences urban pollutant and carbon emission reduction by optimizing energy utilization efficiency, thereby reducing pollution and carbon emissions per unit of output and ensuring the synergistic optimization of economic output, energy savings, and pollution reduction. Specifically, industrial robots collect energy consumption and emission data in real time, dynamically adjust production parameters, and plan energy consumption timings in advance, thereby reducing energy waste and redundancy, improving efficiency, and subsequently promoting urban pollutant and carbon emission reduction.

At the same time, green total energy efficiency is measured using the SBM-DEA model, with fixed asset investment, the number of employees, and total energy consumption as input variables, and regional GDP as the expected output variable. Industrial sulfur dioxide emissions, industrial nitrogen oxides emissions, and carbon dioxide emissions serve as non-desired output variables. This approach utilizes a non-radial and non-angular measurement method to avoid the shortcomings of traditional efficiency measurement that overlook non-desired outputs.

Column (1) of Table 5 shows the impact of artificial intelligence development on green total energy efficiency. It can be observed that the development of artificial intelligence significantly promotes the improvement of green total energy efficiency, and this effect remains significant at the 1% level after controlling for variables such as the level of economic development, financial development, and urbanization rate. This indicates that the development of artificial intelligence indirectly drives the achievement of pollutant and carbon emission reduction goals through this mechanism, thereby validating the mediating role of green total energy efficiency in the relationship between artificial intelligence development and pollutant and carbon emission reduction, thus confirming the hypothesis.

One of the core mechanisms through which the development of artificial intelligence impacts urban pollutant and carbon emission reduction is by promoting industrial structure optimization and reducing the proportion of high-energy-consuming industries. Industrial structure optimization primarily manifests in the transition from the secondary sector to the tertiary sector, as well as the upgrading of high-energy-consuming industries to low-energy-consuming industries. This transformation effectively reduces pollutant emissions and carbon emissions.

In this process, artificial intelligence achieves the goals of pollutant and carbon emission reduction through three aspects of empowerment: First, industrial robots replace low-end, high-energy-consuming jobs, prompting a shift of labor towards the low-energy-consuming tertiary sector, thereby reducing overall unit energy consumption and pollutant emissions. Second, intelligent production technologies drive structural upgrading within the secondary sector by phasing out high-energy-consuming outdated capacities and developing low-energy-consuming industries, thus achieving reductions in carbon emissions. Finally, digital platforms optimize the allocation of industrial resources, reducing dependency on high-energy-consuming industries and guiding resources to concentrate in low-emission areas, effectively lowering overall pollution emissions and carbon footprints.

Column (2) of Table 5 shows the impact of artificial intelligence development on industrial structure. It can be observed that the improvement in artificial intelligence levels significantly promotes the increase in the share of the tertiary sector in total output, and this effect remains significant at the 1% level after controlling for variables such as economic development level, financial development level, and urbanization rate. With the optimization of industrial structure, the overall energy consumption intensity in the region also decreases. The development of artificial intelligence indirectly drives urban pollutant and carbon emission reduction through this mechanism, thereby validating the mediating role of industrial structure in the relationship between artificial intelligence development and urban pollutant and carbon emission reduction, thus confirming the hypothesis.

Another core mechanism through which artificial intelligence promotes pollutant and carbon emission reduction is by facilitating green technology innovation, thereby reducing pollution and carbon emissions per unit of output. Green technology innovation can reduce pollutant emission intensity at the source, enhancing pollutant treatment efficiency through advanced governance technologies and leveraging low-carbon technologies to decrease carbon emissions in production processes, ultimately achieving the synergistic advancement of pollutant and carbon emission reduction.

The development of artificial intelligence drives green technology innovation and assists in emission reductions through three aspects of technological empowerment. First, data accumulated from production and emissions by industrial robots provides empirical evidence for green technology research and development, enabling precise identification of directions for technological breakthroughs. Second, machine learning algorithms accelerate the research and development process, shortening the time it takes for green technologies to transition from laboratories to industrialization. Finally, practical applications of green technologies are promoted through features such as intelligent monitoring and automated control, broadening the applicability of technologies in high-energy-consuming industries, thereby enhancing the effectiveness of green technology innovation and advancing urban pollutant and carbon emission reduction.

Column (3) of Table 5 presents the impact of artificial intelligence development on green technology innovation. It can be observed that advancements in artificial intelligence significantly promote an increase in the number of green invention patent applications, and this effect remains significant at the 1% level even after controlling for a series of foundational variables. As the capability for green technology innovation improves, regional industrial pollution treatment efficiency and low-carbon production levels increase concurrently. Through this mechanism, artificial intelligence indirectly drives the achievement of pollutant and carbon emission reduction goals, thereby validating the mediating role of green technology innovation in the relationship between artificial intelligence development and pollutant and carbon emission reduction, thus confirming the hypothesis.