Does Artificial Intelligence Promote Firms’ Green Technological Innovation?

Abstract

Green technological innovation represents one of the critical driving forces for addressing environmental issues and advancing the sustainable development process. As a key driver of the new round of technological transformation, artificial intelligence is bound to exert significant impacts on firms’ green technological innovation. In this study, green technology innovation is divided into clean production and pollution control technology innovation according to the production link. A double fixed-effects model was used to test the impact of AI using data from Chinese listed companies from 2006 to 2020. The research findings are as follows: First, artificial intelligence has a significant contribution to green technology innovation in different segments. Second, mechanism analysis reveals that artificial intelligence enhances green technological innovation by improving human capital caliber and firm efficiency. Third, heterogeneity analysis shows that the greater the intensity of environmental regulation a firm faces, the greater the incentive for the firm to use AI for green technology innovation; its effect on pollution control technological innovation is more significant for firms in high-pollution industries; and its effect on clean production technological innovation is more prominent for enterprises in low-pollution industries.

1. Introduction

Protecting the ecological environment while promoting economic development represents a critical pathway for every country to achieve sustainable development [1]. Thus, a substantial body of scholars have explored the issue of coordinated development between economic growth and environmental protection from various perspectives, so as to facilitate green development. Green development hinges on green technological innovation, and enterprises, as the most critical micro-level entities in the economy; effectively, leveraging their role as the primary drivers of green technological innovation constitutes a key pathway to accelerate the process of green sustainability [2]. However, green technological innovation not only requires substantial financial resources but also involves high uncertainty regarding returns. For enterprises, comprehensive consideration of difficulties and returns is necessary to alter existing technologies and production processes, thus leading to insufficient motivation [3].

Meanwhile, with the rapid advancement of big data and algorithms, the development of artificial intelligence has ushered in a new era. The new technologies embedded in artificial intelligence are characterized by high efficiency, greenness, and sustainability [4]. This has not only triggered the Fourth Industrial Revolution but has also become a new driving force for the world economy to achieve green and sustainable growth at the present stage. Many countries worldwide, including the United States, EU member states, and China, have incorporated AI into their national strategic agendas. At present, the Chinese government has positioned the application of AI technology as a key strategic means to enhance regional green technological innovation capabilities and ignite social innovation. The 2025 Chinese Government Work Report reiterated the need to sustain and deepen the advancement of intelligentization and digitization and launch “AI+” initiatives. As the concept of green sustainable development becomes deeply integrated into China’s high-quality economic development, whether AI can boost enterprises’ green technological innovation has become a pressing practical question that urgently needs to be answered.

Therefore, this paper primarily analyzes the issue from the following perspectives: What is the impact of China’s current AI development on green technological innovation? What are the underlying mechanisms driving this impact? Does the effect vary under different contextual conditions? Analyzing these questions helps us to understand the environmental externalities arising from the global dissemination of AI technologies and makes us more able to provide beneficial references and insights for the sustainable development of emerging economies.

The innovations of this study are as follows:

Firstly, the existing studies on the impact of AI on green technological innovation predominantly adopt an aggregate perspective of green technological innovation, thereby conflating two distinct processes: clean production technology innovation and pollution control technology innovation. By situating the analysis within the production chain processes of enterprises, the paper subdivides green technological innovation into clean production and pollution control technological innovation. It investigates firms’ choices in AI application across different innovation links, providing a new perspective for stimulating enterprises to shift from passive pollution control to proactive innovation.

Secondly, from the dual dimensions of human capital optimization and corporate efficiency enhancement, this paper deeply analyzes the logical mechanism through which artificial intelligence influences green technology innovation and depicts the differentiated pathways by which AI promotes green technology innovation via these two dimensions.

Lastly, by analyzing heterogeneity from the perspectives of industry characteristics, the stringency of environmental regulations and firm ownership, this study provides a policy-making basis and scientific recommendations for formulating differentiated policies to leverage AI for green technological innovation.

2. Literature Review

2.1. The Impact of Artificial Intelligence

Existing research on artificial intelligence has shown that AI applications can exert significant impacts on multiple dimensions, including employment [5,6], income [7], productivity [8,9], and micro-level corporate environmental investment [10]. AI itself represents a form of technological innovation capable of developing new products and restructuring production processes [11]. When combined with other information technologies, it can effectively reduce enterprises’ production costs [12,13] and incentivize improvements in their innovation capabilities [14]. In contrast to the limited role of traditional restrictive environmental regulations in promoting enterprises’ green technological innovation, can non-regulatory approaches like AI more effectively drive such innovation? Indeed, AI can achieve precise control over enterprise production processes through deep learning with integrated data, enabling energy conservation and emission reduction via more intensive production methods [15,16]. Therefore, the key to achieving green technological innovation lies in implementing clean technologies that control pollution throughout the entire industrial production process. A small body of research at the national and industry levels has found that AI influences green technological innovation by reducing labor costs, improving energy efficiency, and alleviating financing constraints [17,18,19].

2.2. The Links and Types of Green Technological Innovation

Although a few investigations have examined the advantages of artificial intelligence (AI) in green technological innovation activities, they have not clearly explained which specific links of enterprises’ green technological innovation these AI-driven advantages act upon. Most of the existing literature investigates green technological innovation in the aggregate, effectively conflating two independent processes: production and pollution control. In reality, facing pressure from environmental policies, enterprises typically have four strategic choices: first, addressing pollution at the source of the production chain through direct green technological innovation to reduce upstream contamination [20,21]; second, substituting polluting inputs with cleaner alternatives to minimize pollution generated by production factors [22]; third, at the final stage of the production chain, addressing the problem through pollution control technology innovation or direct equipment procurement [23]; fourth, integrating green technological innovation into the entire production chain via management model innovation and product innovation [24].

It is not difficult to see that the influence of AI varies significantly depending on the specific links of the production chain where enterprises apply it. Therefore, the role of artificial intelligence in green technology innovation cannot be generalized, and a deeper exploration of its impacts across different types and links is necessary. To clarify how AI affects green technological innovation, this paper is based on the occurring links and types, categorizing green technological innovation into clean production technological innovation and pollution control technological innovation. Using data from Chinese listed companies during 2006–2020, we conduct an empirical investigation. This research not only helps to dissect the multidimensional mechanisms through which AI drives green technological innovation and its systemic impacts on global sustainable development but also provides valuable references for emerging economies to overcome technical bottlenecks, resource constraints, and other challenges in their green transformation.

3. Theoretical Analysis and Research Hypotheses

3.1. The Impact of Artificial Intelligence on Enterprises’ Green Technological Innovation

Green technological innovation activities often exhibit characteristics of being long-term, complex, and highly uncertain [25]. Constrained by the high thresholds of technology and capital requirements, enterprises often show little willingness to promote their own green technology innovation. Artificial intelligence (AI), with its unique attributes of high technological content and low environmental costs, plays a non-negligible role in promoting green technological innovation. Although the initial motivation for AI applications was not explicitly intended to promote green technology innovation, for enterprises, the adoption of AI can enhance resource utilization, reduce energy consumption [21,26], and exert positive impacts on the environment, thereby driving green technology innovation [27]. The specific manifestations are as follows:

The first is environmental monitoring and precise control. Artificial intelligence (AI) technologies break down data information barriers, enabling environmental protection authorities to grasp truthful and accurate pollution emission information through technical monitoring and intelligent tracking. This facilitates precise control of pollutants and carbon emissions [10,28]. Simultaneously, this aids relevant authorities in formulating environmental policies tailored to enterprises and enhancing enterprises’ proactive willingness to engage in green innovation and reduce pollution.

The second is a reduction in trial-and-error costs in green technological innovation. AI can provide simulated experimental platforms, where technologies such as data analysis and digital simulation help enterprises to simulate and predict innovation projects [29], thereby reducing the trial-and-error costs of green technological innovation.

The third is resource optimization and energy-saving emission reduction. The labor substitution effect brought by AI enables enterprises to allocate high-quality human resources to R&D innovation, environmental technology improvement, and other links [30]. Meanwhile, AI effectively reduces corporate costs and minimizes energy consumption and carbon emissions by optimizing production, manufacturing, sales, and logistics processes [31,32].

Based on this, the following hypothesis H1 is proposed.

Hypothesis 1.

Artificial intelligence is conducive to promoting enterprises’ green technological innovation.

3.2. Analysis of the Influence Mechanism of Artificial Intelligence on Green Technology Innovation

Artificial intelligence promotes green technology innovation by improving the quality of human capital. The adoption of AI technology exerts dual effects on the labor market, namely labor substitution and new job creation, which in turn trigger changes in the labor structure [33]. Compared with conventional technological innovation, the realization of green technological innovation relies more heavily on high-quality human capital due to its inherent specificity [34]. First, the skill-biased theory points out that the application of information technology increases the demand for workers’ critical thinking abilities and high-skill levels [35], thereby enhancing the caliber of human capital. High-caliber human capital stimulates the potential for green technological innovation cooperation among enterprises through knowledge spillover and sharing channels, expanding the intellectual foundation for corporate green technological innovation and thus improving green technological innovation capabilities. Second, AI gives rise to technologically talented people such as engineers, designers, and data analysts who are adapted to intelligent technologies [36]. These technologically talented people possess strong technical absorption capabilities and can quickly master and apply cutting-edge technologies in their fields [37]. They can not only use AI to achieve complementary innovation but also engage in original innovation, accelerating the diffusion of green innovative technologies [38].

Secondly, AI applications promote green technological innovation by enhancing corporate efficiency. First, AI automates and intelligentizes enterprise procurement, production, and sales processes, enabling effective coordination of material flow, information flow, and capital flow [39]. This improves resource utilization efficiency and production efficiency. Second, AI leverages massive data and algorithms for predictive modeling and optimization, reducing search costs and facilitating efficient supply–demand matching [40], thereby supporting green production practices. Finally, AI strengthens communication with upstream and downstream stakeholders, enhancing resource acquisition and integration capabilities to improve overall operational efficiency [41]. This not only boosts corporate output and revenue but also attracts investor capital, increasing the funding available for green technological innovation.

Based on the theoretical analysis, we propose hypotheses H2a and H2b:

Hypothesis 2a.

Artificial intelligence promotes enterprises’ green technological innovation by improving the caliber of human capital.

Hypothesis 2b.

Artificial intelligence promotes enterprises’ green technological innovation by enhancing corporate efficiency.

3.3. The Stages and Types Through Which Artificial Intelligence Promotes the Occurrence of Green Technological Innovation

At the production source and throughout the production process, artificial intelligence (AI) fosters clean production innovation among enterprises. As a critical digital technology resource for expanding product and service frontiers in corporate operations, AI equips firms with advanced technologies [42], optimizing production workflows and enhancing product competitiveness. On one hand, AI applications facilitate the dynamic optimization of marginal production costs. The long-term viability of clean production technologies hinges on improvements in resource utilization efficiency, which AI achieves through real-time data processing to enable the precise allocation of production factors. Concurrently, AI provides enterprises with advanced technical infrastructure for clean production, including refining/purification systems and monitoring/protection technologies [43]. By minimizing energy intensity and enhancing value-added outputs, AI elevates energy productivity, drives the clean conversion of traditional energy sources, and scales up clean energy applications [44], thereby upgrading corporate clean production performance. Second, AI enables the intelligent amortization of upfront R&D costs. The core bottlenecks to clean production innovation—high fixed costs and uncertainty—are mitigated by AI through algorithmic simulation and data-driven optimization, reducing trial-and-error expenses. Digital twin technology, for instance, allows for virtual simulation of production process reconfiguration [45], eliminating the need for physical prototyping in the traditional “trial-and-error iteration” model. This empowers enterprises to actively improve production processes throughout the entire production cycle, achieve technological innovation in cleaner production [46].

In the context of end-of-pipe governance, artificial intelligence (AI) promotes corporate pollution treatment technology innovation. As a means of end-of-pipe governance, pollution treatment technology innovation directly responds to regulatory mandates on pollutant emission concentrations and total volumes. This represents a survival-driven investment for enterprises to avoid administrative sanctions and maintain production licenses. For heavily polluting firms in particular, facing stricter environmental regulations, their technology innovation preferences are the result of a combination of institutional constraints and cost rationality. Due to lower profit margins, heavy-pollution industries tend to favor pollution control innovations with predictable short-term costs and high revenue certainty. On one hand, AI enhances pollutant treatment precision, mitigating secondary pollution caused by low manual operation accuracy [47] and improving pollution control efficiency. On the other hand, enterprises use intelligent sensing systems to achieve information collection, monitoring, intelligent analysis, and granular management in end-of-pipe governance processes [48]. This expands end-of-pipe governance from traditional pollutant treatment to integrated energy conservation and pollution reduction [49], driving green transformation in terminal governance and enabling pollution control technology innovation.

Based on the theoretical analysis, we propose hypotheses H3 and H4:

Hypothesis 3.

Artificial intelligence applications promote cleaner production technology innovation through energy-saving technology and process innovation.

Hypothesis 4.

AI applications facilitate pollution control technological innovation through end-of-pipe treatment-strengthening mechanisms.

The research framework and hypotheses of this paper are shown in Figure 1 below.

4. Empirical Research

4.1. Model Design

Following the research of Acemoglu and Restrepo (2020) [5], this paper constructs the following benchmark regression model and employs a panel double fixed-effects analysis to test the impact of artificial intelligence on green technological innovation.

$${\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_2{\mathrm{x}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}{+\mathsf{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(1)

Coefficient ${\mathsf{\beta }}_1$ measures the impact of artificial intelligence on green technological innovation; ${\mathsf{\beta }}_2$ represents the impact of control variables; $\mathrm{i}$ denotes the enterprise; $\mathrm{t}$ denotes the year; the explained variable ${\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_{\mathrm{i}\mathrm{t}}$ indicates the firm’s green technology innovation capability; ${\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{t}}$ is the explanatory variable, representing artificial intelligence applications; ${\mathrm{x}}_{\mathrm{i}\mathrm{t}}$ represents all control variables in the study; ${\mu }_i$ represents the fixed effects of sample enterprises, controlling for firm-level time-invariant related factors; ${\mathsf{\gamma }}_{\mathrm{t}}$ represents the time-fixed effects of the research sample to control for time-trending factors; ${\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$ is a random error term.

4.2. Description of Related Variables

4.2.1. Explanatory Variables

The core explanatory variable in this paper is artificial intelligence (${\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{e}\mathrm{t}}$). Currently, there is no scholarly consensus on a uniform measurement standard for the level of AI application. AI encompasses various forms, including robotics, image recognition, and multimodal models. Among these, industrial robots—as the highly integrated outcome of modern industry and smart manufacturing, serve not only as the primary manifestation of AI in industrial production but also as a critical driver of industrial intelligence and automation. A substantial body of the existing literature has leveraged industrial robot applications to explore the microeconomic effects of AI. Acemoglu and Restrepo (2020) [5] and Fan et al. (2021) [37] are classic articles in the literature in this field. Drawing on their research, this paper employs the robot penetration rate at the enterprise level as a proxy for artificial intelligence applications.

The IFR database has released data on the use of industrial robots across different countries and industries. However, metrics measured at the industry level cannot truly reflect the situation at the enterprise level. Acemoglu and Restrepo [5] constructed a regional-level industrial robot penetration index using a general equilibrium model to measure AI adoption at the regional level in the United States. Wang and Dong [50] adapted this method to measure industrial robot penetration in China at the enterprise level. This is currently a relatively representative method in the literature for measuring the application of industrial robots at the micro-enterprise level in China. The following two steps are the measurement methods:

Step 1: Calculate the installation density of industrial robots in different industries.

$${\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{y}}_{\mathrm{e}\mathrm{t}}^{\mathrm{C}\mathrm{H}}={\displaystyle \frac{{\mathrm{r}\mathrm{o}\mathrm{b}\mathrm{o}\mathrm{t}}_{\mathrm{e}\mathrm{t}}^{\mathrm{C}\mathrm{H}}}{{\mathrm{l}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}}_{\mathrm{e}\mathrm{t}=2010}^{\mathrm{C}\mathrm{H}}}}$$

(2)

where ${\mathrm{r}\mathrm{o}\mathrm{b}\mathrm{o}\mathrm{t}}_{\mathrm{e}\mathrm{t}}^{\mathrm{C}\mathrm{H}}$ represents the stock of industrial robots in the Chinese industry in 2010, and ${\mathrm{l}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}}_{\mathrm{e}\mathrm{t}=2010}^{\mathrm{C}\mathrm{H}}$ represents the number of people employed in the Chinese industry in 2010, with 2010 as the base period. ${\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{y}}_{\mathrm{e}\mathrm{t}}^{\mathrm{C}\mathrm{H}}$ represents the installed density of industrial robots in the Chinese industry in 2010.

Step 2: Construct the industrial robot penetration indicator at the enterprise level.

$${\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{e}\mathrm{t}}={\displaystyle \frac{{\mathrm{p}\mathrm{w}\mathrm{p}}_{\mathrm{i}\mathrm{e}\mathrm{t}=2011}}{\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{u}{\mathrm{p}\mathrm{w}\mathrm{p}}_{\mathrm{i}\mathrm{e}\mathrm{t}=2011}}}\times {\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{y}}_{\mathrm{e}\mathrm{t}}^{\mathrm{C}\mathrm{H}}$$

(3)

where ${\mathrm{p}\mathrm{w}\mathrm{p}}_{\mathrm{i}\mathrm{e}\mathrm{t}=2011}$ represents the proportion of employees owned by the production department of enterprise i in China’s e-industry in 2011, with 2011 as the base period; $\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{u}{\mathrm{p}\mathrm{w}\mathrm{p}}_{\mathrm{i}\mathrm{e}\mathrm{t}=2011}$ represents the median of the sum of the number of employees in the production department of enterprises in China in 2011 as a proportion of the industry; and ${\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{e}\mathrm{t}}$ represents the penetration degree of industrial robots of enterprise i in China’s e-industry in year t, which is an explanatory variable of this paper, and is taken as logarithmic treatment.

4.2.2. Explained Variables

Slightly different from the indicators used by previous scholars, the explanatory variable is green technology innovation capacity ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$), which is selected to be measured by aggregate green technology innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}$), cleaner production technology innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{c}\mathrm{e}$), and pollution control technology innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{d}\mathrm{e}$). In the processing method, in order to eliminate the “right-skewed” characteristics of green technology patents, we refer to the method of the existing research [51]; the value is processed by first adding one to it and then taking the logarithm.

Aggregate green technological innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}$): Measured by adding one to the logarithm of the number of green patent applications filed by enterprises. The reason for choosing this indicator is that green patents can quantify the output of enterprises’ green technology innovation activities well. The number of patent applications rather than the number of patents granted is chosen because the latter has a time lag and is affected by the subjective factors of patent agencies and politics. A comprehensive comparison reveals that it is more reasonable to use the number of patent applications to measure green technology innovation capacity.

Cleaner production technology innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{c}\mathrm{e}$): Energy efficiency technology innovation or resource allocation through energy efficiency technology innovation. Refer to Zhang et al. (2022) [52]. It is measured by the number of alternative energy inventions licensed plus one taking the logarithm ($\mathrm{n}\mathrm{y}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$) and the number of energy-saving inventions licensed plus one taking the logarithm ($\mathrm{j}\mathrm{n}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$).

Pollution control technology innovation ($\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{d}\mathrm{e}$): Pollution control technology innovation is a technological innovation for the removal of end-of-pipe pollutants to reduce pollutant emissions. Refer to Zhang et al. (2022) [52]. It is measured by the number of inventions licensed for waste management plus one taking the logarithm ($\mathrm{f}\mathrm{w}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$) and the number of inventions licensed for transportation plus one taking the logarithm ($\mathrm{y}\mathrm{s}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$).

4.2.3. Control Variables

(1) Age of the firm ($\mathrm{l}\mathrm{n}\mathrm{a}\mathrm{g}\mathrm{e}$): This is measured by subtracting the year the firm went public from the year of the statistic and then taking the logarithm of the resulting value. (2) Return on corporate assets ($\mathrm{r}\mathrm{o}\mathrm{a}$): This is measured by the ratio of a firm’s net profit to its total assets. (3) Capital intensity ($\mathrm{c}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a}\mathrm{l}$): Generally, capital-intensive firms are more innovative, as measured by the total capital of the firm as a share of operating revenues. (4) Degree of industry competition ($\mathrm{h}\mathrm{h}\mathrm{i}$): This is represented by the Herfindahl index. It should be noted that the Herfindahl index is an inverse indicator, and the smaller its value, the more competitive the industry is, and the more it can promote enterprises to carry out green technological innovation [53]. (5) Level of economic development ($\mathrm{l}\mathrm{n}\mathrm{p}\mathrm{g}\mathrm{d}\mathrm{p}$): We calculate real GDP per capita in real terms and then log the result. (6) Financial health of the firm ($\mathrm{l}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{n}$): This is measured by the total liabilities to total assets of the firm and taking the logarithm of this ratio. Usually, the healthier the financial position of a company, the easier it is to provide more money for technological innovation. (7) R&D investment intensity ($\mathrm{l}\mathrm{n}\mathrm{r}\mathrm{d}$): Firms’ R&D expenditures as a percentage of year-end total assets are used to measure the ratio and the logarithm of the ratio is taken.

4.3. Sample Curation and Data Origin

The research sample of this paper is companies listed in Shanghai and Shenzhen A-shares from 2006 to 2020, and the data used in the robot penetration calculation come from the IFR. Since Chinese robots started to accumulate significant data records in 2006, the year 2006 is taken as the starting year for the study of this paper. The reason why this paper selects listed companies as research samples is that listed companies are larger and more influential, and are representative of the industry, so they can produce a demonstration effect in green technology innovation. The green transformation brought by this demonstration effect strongly promotes the sustainable development of society.

The data source for green patents is the China Research Data Service Platform (CNRDS); economic characterization data are taken from the Cathay Pacific Database (CSMAR); relevant data at the regional level are taken from the China Urban Statistical Yearbook; and the number of people employed in the industry and province and the total number of people employed nationwide are from the China Labor Statistical Yearbook. In addition, the following treatments are applied to the raw data in this paper: (1) The samples with ST, *ST, and too many missing values are excluded. (2) The samples of financial listed firms are excluded. (3) In order to avoid the effect of extreme values in the sample on the results, this study shrinks the continuous variables in the sample by 1% and 99%.

Table 1. Descriptive statistics of variables.

Variable Name	Variable Expression		Obs	Mean	Std.	Min	Max
$\mathrm{Aggregate}\mathrm{green}\mathrm{technology}\mathrm{innovation}$	$\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}$		25,256	0.335	0.743	0	3.611
$\mathrm{Cleaner}\mathrm{production}\mathrm{technology}\mathrm{innovation}$	$\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{c}\mathrm{e}$	$\mathrm{n}\mathrm{y}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$	25,256	0.078	0.257	0	2.761
		$\mathrm{j}\mathrm{n}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$	25,256	0.102	0.312	0	2.164
$\mathrm{Innovations}\mathrm{in}\mathrm{pollution}\mathrm{control}\mathrm{technology}$	$\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\_\mathrm{d}\mathrm{e}$	$\mathrm{f}\mathrm{w}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$	25,256	0.086	0.279	0	2.632
		$\mathrm{y}\mathrm{s}\mathrm{l}\mathrm{n}\mathrm{n}\mathrm{o}$	25,256	0.057	0.199	0	1.159
$\mathrm{Artificial}\mathrm{intelligence}\mathrm{application}$	$\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}$		25,256	2.619	0.952	−2.843	7.582
$\mathrm{Age}\mathrm{of}\mathrm{the}\mathrm{enterprise}$	$\mathrm{l}\mathrm{n}\mathrm{a}\mathrm{g}\mathrm{e}$		25,256	2.099	0.865	0	3.259
$\mathrm{Return}\mathrm{on}\mathrm{total}\mathrm{enterprise}\mathrm{assets}$	$\mathrm{r}\mathrm{o}\mathrm{a}$		25,256	0.039	0.056	−0.211	0.191
$\mathrm{Capital}\mathrm{intensity}$	$\mathrm{c}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a}\mathrm{l}$		25,256	0.612	0.429	0.035	2.562
$\mathrm{Degree}\mathrm{of}\mathrm{competition}\mathrm{in}\mathrm{the}\mathrm{industry}$	$\mathrm{h}\mathrm{h}\mathrm{i}$		25,256	0.098	0.121	0.017	1
$\mathrm{Level}\mathrm{of}\mathrm{economic}\mathrm{development}$	$\mathrm{l}\mathrm{n}\mathrm{p}\mathrm{g}\mathrm{d}\mathrm{p}$		25,256	2.126	0.541	0.650	3.013
Financial health of the enterprise	$\mathrm{l}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{n}$		25,256	0.433	0.209	0.049	0.887
$\mathrm{R}\&\mathrm{D}\mathrm{investment}\mathrm{intensity}$	$\mathrm{l}\mathrm{n}\mathrm{r}\mathrm{d}$		25,256	9.953	10.680	0	12.612

demonstrates the descriptive statistics of the sample variables. It can be seen that there is a large gap in the results of the AI application indicator, where the maximum value is 7.582 and the minimum value is −2.843. This indicates that there is a large difference in the application of AI by different companies. From comparative observation of the results of different types of green technology innovation indicators, it can be found that there is obvious heterogeneity in the links and types of green technology innovation chosen by enterprises.

5. Empirical Results and Analysis

5.1. Benchmark Regression Results and Analysis

Table 2. Benchmark regression results for artificial intelligence influencing green technology innovation.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.058 ***	0.041 ***	0.045 ***	0.038 **	0.026 *
	(0.012)	(0.007)	(0.006)	(0.015)	(0.014)
lnage	0.035	0.046	0.031	0.019	0.019
	(0.091)	(0.072)	(0.072)	(0.067)	(0.068)
roa	0.118 *	0.086	0.084	0.083	0.079
	(0.069)	(0.055)	(0.055)	(0.053)	(0.052)
capital	0.145	0.081	0.028	0.156	0.157
	(0.151)	(0.118)	(0.118)	(0.114)	(0.113)
hhi	−0.133 ***	−0.138 ***	−0.109 ***	−0.078 ***	−0.081 ***
	(0.036)	(0.036)	(0.029)	(0.026)	(0.027)
lnpgdp	0.013	0.015	0.019	0.009	0.004
	(0.019)	(0.019)	(0.014)	(0.014)	(0.015)
lnfinan	0.087 ***	0.062 ***	0.051 ***	0.044 ***	0.043 ***
	(0.011)	(0.008)	(0.008)	(0.007)	(0.008)
lnrd	0.061 **	0.045 *	0.053 **	0.045 **	0.043 *
	(0.029)	(0.024)	(0.023)	(0.022)	(0.023)
C	−0.335	−0.243	−0.134	−0.282	−0.172
	(0.612)	(0.431)	(0.291)	(0.611)	(1.221)
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
Observations	25,256	25,256	25,256	25,256	25,256
R-squared	0.672	0.655	0.643	0.635	0.635

Notation: Robust standard errors in ( ). ***, **, and * indicate passing significance tests at 1%, 5%, and 10% significance levels, respectively. Y for represents the control.

shows the results of the benchmark regression of AI affecting green technology innovation. Controlling for individual and time effects, column (1) shows that industrial robot applications significantly contribute to green technology innovation. For every 1% increase in artificial intelligence penetration intensity, the aggregate green technology innovation capability increases by 0.058%, and the hypothesis H1 is supported.

According to the link where technological innovation occurs, green technological innovation is categorized into clean production and pollution control technological innovation for empirical testing. Columns (2) and (3) show the impact of artificial intelligence application on cleaner production technology innovation, and artificial intelligence application can significantly improve the cleaner production technology innovation ability of enterprises. Artificial intelligence can prompt enterprises to improve the production process, optimize the production process, and at the same time improve the efficiency of energy use, and carry out cleaner production technology innovation from the source and production end. The research hypothesis H3 of this paper is confirmed. Columns (4) and (5) show the impact of artificial intelligence application on pollution control technology innovation. The fact that artificial intelligence application significantly enhances the firms’ innovation in pollution control technology confirms the research hypothesis H4 of this paper. Compared with cleaner production technology innovation, although artificial intelligence application promotes pollution control technology innovation, the coefficient is smaller and less significant. This suggests that when artificial intelligence application promotes enterprise green technology innovation, the main links occur in the source and production process, and the driving effect on pollution control technology innovation is not as direct and significant as that of clean production technology innovation. The possible reason is that the end of the treatment of pollution emissions needs to pay a high cost; in the absence of regulatory policy pressure, the enterprise initiative is weak due to insufficient incentives.

5.2. Robustness Tests

5.2.1. Endogeneity Issues

Robot application and green technology innovation may have the potential risk of endogeneity of mutual causality, and there may be endogeneity problems such as other omitted variables that cannot be observed, which make the estimated coefficients of the core explanatory variables biased. Therefore, this paper refers to Sheng and Bu (2022) [54], based on model (3), replacing the number of subindustry installations of Chinese robots with the number of subindustry installations of U.S. robots to obtain the U.S. robot penetration, which is then logarithmically used as an instrumental variable. First, the choice of instrumental variables needs to satisfy the relevance, and the high level of industrial robot application in the United States has a demonstration effect on the application of robots in China, which can satisfy the relevance requirement. Second, the choice of instrumental variables also needs to satisfy exogeneity; the application of industrial robots in the U.S. does not have an impact on China’s green technological innovation, and is able to fulfill the condition of exogeneity.

The results of the instrumental variables approach used to test the endogeneity of AI affecting green technology innovation are displayed in columns (1)–(6) of

Table 3. Results of instrumental variable estimation.

	First Stage	Second Stage	Second Stage		Second Stage		DID
Variables	lnrobit	ginnov_total	ginnov_ce		ginnov_de		ginnov_ total
			nylnno	jnlnno	fwlnno	yslnno
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
IVlnexposure	1.315 ***
	(0.011)
lnexposure		0.061 ***	0.042 ***	0.048 ***	0.041 **	0.028 *
		(0.010)	(0.007)	(0.006)	(0.017)	(0.015)
$\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\times \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{t}2015$							0.068 ***
							(0.014)
control	Y	Y	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y	Y
Observations	25,256	25,256	25,256	25,256	25,256	25,256	25,256

Notation: Robust standard errors in ( ). ***, **, and * indicate passing significance tests at 1%, 5%, and 10% significance levels, respectively. Y for represents the control.

. The regression results from the first stage show that the penetration of industrial robots in the U.S. is significantly and positively correlated with the level of robot adoption in China. The regression results in the second stage show that industrial robot application has a significant promotion effect on green technology innovation in different segments of enterprises. This result is the same as the previous benchmark regression results, indicating that the benchmark regression in the previous part of the article is robust. Compared to the coefficients in the baseline regression, the estimated coefficients of $\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}$ increase across sessions. This reflects a downward bias in the level of robot adoption in China under the endogeneity test, and an underestimation of the impact of robot adoption on green technology innovation.

As a rapidly developing strategic emerging industry, the introduction of relevant support policies for artificial intelligence (AI) may have far-reaching impacts on its development. This paper further conducts an exogenous shock test to address endogeneity issues.

In 2015, the State Council issued the Made in China 2025 plan, which elevated intelligent and green development in the industrial sector to the national strategic level, marking a milestone for industrial green development. Therefore, this paper selects 2015 as the policy-shock timing to construct a difference-in-difference (DID) model. To avoid biases arising from the subjective assignment of treatment and control groups, a continuous variable representing AI adoption is used as a proxy for the model’s grouping variable. $\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\times \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{t}2015$ denotes the policy shock variable, and $\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{t}2015$ denotes the timing variable, while the timing variable takes the value of 1 for years after 2015 and 0 otherwise.

Column (7) of Table 3 shows the results of the test when exogenous shocks are used as instrumental variables and the coefficient on $\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\times \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{t}2015$ is significantly positive at the 1% level, indicating that enterprises subjected to the policy shock from the above guidelines exhibit more pronounced effects in green technological innovation. The hypothesis H1 was further tested.

5.2.2. Replacement of AI Application Metrics

Some of the related studies use robot installations for measuring robot penetration. Therefore, unlike the robot stock data used in the baseline regression, this section uses robot installations for remeasurement. The new robot permeability value ($\mathrm{n}\_\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}$) was calculated according to Equation (2) and regressed again. The regression results in

Table 4. Estimated results of replacing the core explanatory variables.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
n_lnexposure	0.052 ***	0.038 ***	0.040 ***	0.034 **	0.027 *
	(0.010)	(0.007)	(0.006)	(0.014)	(0.014)
control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.674	0.653	0.642	0.637	0.639
Observations	25,256	25,256	25,256	25,256	25,256

Notation: Robust standard errors in ( ). ***, **, and * indicate passing significance tests at 1%, 5%, and 10% significance levels, respectively. Y for represents the control.

exhibit that AI still significantly contributes to green technology innovation, and the estimated coefficients as well as the significance are not much different from the baseline regression. This indicates that the results of the previous regression analysis are robust and that the impact of industrial robotics applications on green technology innovation does not change depending on the measure.

5.2.3. Exclude Sample Enterprises That Have Never Applied for Patents

Considering that some companies have never applied for patents during the study period, it may have an impact on the regression results. Therefore, this paper re-examines the test after removing the companies with 0 green patent applications. The regression results in

Table 5. Excluding never-patented firms’ and systems’ GMM estimation results.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	ginnov_total	ginnov_ce		ginnov_de		ginnov_total
		nylnno	jnlnno	fwlnno	yslnno
L.ginnov_total						0.065
						(0.036)
lnexposure	0.061 ***	0.039 ***	0.047 ***	0.037 **	0.022 **	0.072 ***
	(0.013)	(0.007)	(0.008)	(0.015)	(0.009)	(0.016)
control	Y	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
Observations	19,437	19,437	19,437	19,437	19,437	25,256
AR(1)-P						0.000
AR(2)-P						0.234
Hansen						0.372

Notation: Robust standard errors in ( ). *** and ** indicate passing significance tests at 1% and 5% significance levels, respectively. Y for represents the control. The original hypothesis of the AR(1) and AR(2) test is that there is no autocorrelation. The original hypothesis of the Hansen test is that the selection of instrumental variables is valid.

exhibit that the impact of AI applications on firms’ green technology innovation remains significantly positive, and the significance of robot application on pollution control technology innovation is significantly stronger, indicating that robot application enhances the enthusiasm of corporate green technology innovation. Excluding the sample firms that never filed a patent, the regression results do not differ much from the previous benchmark regression, which further supports the accuracy of the results of this paper.

5.2.4. System GMM Estimation

Although in the previous analysis this paper used an instrumental variables approach to overcome endogeneity, it neglected the effect of temporal dynamics. Due to the cumulative nature of green technology innovation, current period green innovation in panel data may depend on lagged values due to technological inertia. The GMM estimation method, which incorporates the lagged dependent variable as an explanatory variable in the model, is able to avoid estimation bias in the fixed-effects dynamic panel due to the correlation of the lagged term with the error term. It also enables a more accurate test of the temporal dynamics of the role of AI in green technology innovation. Therefore, the system GMM is used in this part for further testing.

As can be seen from the regression results in column (6) of Table 5, the impact of AI on green technology innovation remains significantly positive and the absolute value of the impact coefficient has increased. A possible reason for this is that green technology innovation is cumulative, and fixed-effects models that do not incorporate lagged innovation variables will ignore “technological path dependence”. GMM can better reflect firms’ technology path dependence by introducing a lag term for green technology innovation. When enterprises apply AI at the initial stage, they will continue to invest in it based on the innovation foundation of the previous stage, thus forming a positive cycle in which the previous innovation promotes the acceleration of the current innovation.

5.2.5. Considering the Lag in Artificial Intelligence Applications Affecting Green Technology Innovation

Green technology innovation from input to output often requires a certain period, the firm’s green technology innovation inputs often do not have an immediate effect, and there may be a certain lag. For the purpose of avoiding the lag of green innovation to the benchmark regression results of interference, in this paper, both explanatory and control variables are taken lagged by one period and then tested. The regression results in

Table 6. Estimation results considering the lag of green technology innovation.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lag_lnexposure	0.043 ***	0.037 ***	0.039 ***	0.023 *	0.021 *
	(0.011)	(0.008)	(0.008)	(0.012)	(0.012)
lag_control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.663	0.652	0.661	0.635	0.629
Observations	22,183	22,183	22,183	22,183	22,183

Notation: Robust standard errors in ( ). *** and * indicate passing significance tests at 1% and 10% significance levels, respectively. Y for represents the control.

exhibit this, showing that the significance of robots lagged by one period on firms’ green technology innovation is unchanged, but the estimated coefficients have declined from those in the baseline regression. This indicates to some extent that, over time, the promotion effect of AI application on green technology innovation still exists, but its marginal effect shows a decreasing trend. The possible reasons for this are as follows: First, with the advancement of science and technology, the longer the artificial intelligence application, the less cutting edge its scientific and technological connotations, thus making its marginal effect on green technology innovation less effective. On the other hand, due to the wear and tear of machine parts, natural corrosion, and other reasons, the accuracy of artificial intelligence will show a trend of attenuation with the fading of time, thus causing the marginal effect of robots on green technology innovation (GTI) to decline. For this reason, dynamic updating of artificial intelligence is an important way of maintaining its high marginal contribution to GTI in enterprises.

5.3. Mechanism Testing

The previous test results show that AI significantly contributes to green technology innovation. However, the previous analysis has only portrayed the overall picture of “artificial intelligence–enterprise green technology innovation”, and has not yet studied the mechanism of the “black box”. What channels are used to realize this promotion effect? The previous theoretical analysis suggests that artificial intelligence application may affect green technology innovation by improving human capital quality and enterprise efficiency, and this part adopts the mediating effect method to test this.

For the purpose of exploring the role mechanism of AI application affecting enterprises’ green technological innovation, drawing on Jiang (2022) [55], based on Equation (1), Equations (3) and (4) are further constructed to sequentially test whether artificial intelligence application promotes green technological innovation by enhancing human capital quality and enterprise efficiency.

$${\mathrm{M}}_{\mathrm{i}\mathrm{t}}={\mathsf{\theta }}_0+{\mathsf{\theta }}_1{\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\theta }}_2{\mathrm{x}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}{+\mathsf{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(4)

$${\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\delta }}_0+{\mathsf{\delta }}_1{\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\delta }}_2{\mathrm{M}}_{\mathrm{i}\mathrm{t}}{+\mathsf{\delta }}_3{\mathrm{x}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}{+\mathsf{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(5)

where $\mathrm{M}$ is a mediating variable indicating human capital quality ($\mathrm{h}\mathrm{l}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{r}$) and firm efficiency ($\mathrm{t}\mathrm{f}\mathrm{p}$).

Human capital quality ($\mathrm{h}\mathrm{l}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{r}$): This is measured by the percentage of total personnel with a master’s degree or higher. Firm efficiency ($\mathrm{t}\mathrm{f}\mathrm{p}$): According to the literature of James and Amil (2010) [56], the firm’s TFP calculated by method A is used for representation. The remaining variables remain consistent with the earlier definitions.

5.3.1. With Human Capital Quality as the Mediator

The results of the test mediated by human capital are presented in

Table 7. Estimation results of the mediation mechanism with human capital as the mediator.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	hlabor	ginnov_total	ginnov_ce		ginnov_de
			nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.137 ***	0.053 ***	0.035 ***	0.039 ***	0.025	0.021 *
	(0.015)	(0.011)	(0.009)	(0.010)	(0.018)	(0.012)
hlabor		0.107 ***	0.101 ***	0.112 ***	0.098	0.092
		(0.015)	(0.009)	(0.011)	(0.122)	(0.122)
control	Y	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
R-squared	0.832	0.791	0.811	0.826	0.731	0.726
Observations	14,795	14,795	14,795	14,795	14,795	14,795

Notation: Robust standard errors in ( ). *** and * indicate passing significance tests at 1% and 10% significance levels, respectively. Y for represents the control.

. The results of the test in column (1) show that the effect of AI on the quality of human capital is positive and significant at the 1% level, and that each 1% increase in the level of AI adoption promotes a 0.137% increase in the proportion of high-quality employees in the firm. In column (2), both the coefficients for robot penetration and human capital quality are significantly positive, and the coefficient for robot penetration is smaller than that in the baseline regression, suggesting that human capital quality plays a partial mediating role in the impact of artificial intelligence on green technological innovation—validating the hypothesis H2a. The results in columns (3)–(6) indicate that robot applications more strongly promote clean production technology innovation by enhancing human capital quality. Possible reasons include the following: First, clean production technology innovation requires high-quality human capital with interdisciplinary knowledge and innovative thinking, whereas pollution control technology innovation—focused on end-of-pipe treatment—relies more on technical skills in specific domains and has a lower dependence on high-quality human capital. Second, high-quality human capital faces higher opportunity costs (e.g., forgoing breakthrough research in cutting-edge fields) when engaging in pollution control technology innovation, thus preferring clean production technology areas with higher knowledge barriers and stronger scarcity.

5.3.2. With Enterprise Efficiency as the Mediator

The results of the test mediated by enterprise efficiency are presented in

Table 8. Estimation results of the mediation mechanism with enterprise efficiency as the mediator.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	tfp	ginnov_total	ginnov_ce		ginnov_de
			nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.152 ***	0.054 ***	0.036 ***	0.041 ***	0.028 **	0.019 *
	(0.013)	(0.012)	(0.009)	(0.012)	(0.012)	(0.011)
tfp		0.125 ***	0.113 ***	0.112 ***	0.095 **	0.098 *
		(0.016)	(0.011)	(0.011)	(0.012)	(0.036)
control	Y	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
R-squared	0.895	0.826	0.821	0.837	0.782	0.779
Observations	14,795	14,795	14,795	14,795	14,795	14,795

Notation: Robust standard errors in ( ). ***, **, and * indicate passing significance tests at 1%, 5%, and 10% significance levels, respectively. Y for represents the control.

. The results of the test in column (1) show that the effect of AI on firm efficiency is positive and significant at the 1% level, and that each 1% increase in the level of AI application promotes firm efficiency by 0.152%. In column (2), both the coefficients for robot penetration and enterprise efficiency are significantly positive, and the coefficient for robot penetration is smaller than that in the baseline regression, suggesting that enterprise efficiency plays a partial mediating role in the impact of artificial intelligence on green technological innovation—validating the hypothesis H2b. The results of the tests in columns (3)–(6) exhibit that robot applications promote both clean production technology innovation and pollution control technology innovation by enhancing enterprise efficiency. The improvement in overall enterprise efficiency implies the optimization of production methods, organizational management, and resource allocation capabilities, which can effectively reduce factor costs and increase corporate revenues. This enables enterprises to allocate more funds to R&D activities, thereby improving their green technological innovation capabilities.

5.3.3. Rechecking the Mediating Effects

The stepwise regression method used in the previous test may have the problem of insufficient validity, for this reason, this part applies the bootstrap method to further test the mediating effect, and

Table 9. Bootstrap test results.

Dependent Variable	Dependent Variable	Effect Category	Effect Size	Standard Error	95% Confidence Interval
					Upper Limit	Lower Limit
lnexposure	hlabor	Indirect Effect	0.0016	0.0003	0.0010	0.0023
		Direct Effect	0.0080	0.0018	0.0044	0.0116
lnexposure	tfp	Indirect Effect	0.0010	0.0002	0.0014	0.0006
		Direct Effect	0.0060	0.0012	0.0084	0.0036

exhibits the results of further analysis. Setting the number of resampling iterations to 1000, the confidence intervals for both the indirect effects and direct effects of human capital quality and enterprise efficiency do not include zero. This suggests that human capital quality and firm efficiency play a partly mediating role in AI-driven green technology innovation, with the hypotheses H2a and H2b being validated again.

6. Heterogeneity Tests

6.1. Industry Heterogeneity

In order to further explore whether artificial intelligence application makes a difference in the green technology innovation of enterprises in different industries, this paper conducts group regression after categorizing the enterprises. Referring to the “Guidelines for Environmental Information Disclosure of Listed Companies” published by the PRC Ministry of the Environment in 2010, and combining them with the standard of the “Guidelines for Industry Classification of Listed Companies” revised by the China Securities Regulatory Commission in 2012, 16 types of industries, such as coal, metallurgy, chemicals, petrochemicals, building materials, and paper, are categorized as high-pollution industries, and the rest are categorized as low-pollution industries.

Table 10. Estimated results of industry heterogeneity: heavily polluting industries.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.043 ***	0.032 *	0.025	0.039 ***	0.031 **
	(0.012)	(0.019)	(0.021)	(0.012)	(0.015)
control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.696	0.687	0.693	0.745	0.745
Observations	8691	8691	8691	8691	8691

Notation: Robust standard errors in ( ). ***, **, and * indicate passing significance tests at 1%, 5%, and 10% significance levels, respectively. Y for represents the control.

exhibits the test results for the heavily polluting industries. The results show that robot applications significantly promote pollution control technology innovation among enterprises in heavily polluting industries. This may be due to two reasons: First, heavily polluting enterprises are often traditional industrial firms with severe product homogeneity and lagging green production technology. Facing mandatory regulatory policies, end-of-pipe treatment measures are less difficult to implement, yield quick results, and involve low trial-and-error costs. Second, mature heavily polluting industries may be in the maturity or decline phase of their lifecycle, with technological trajectories leaning toward incremental improvements—i.e., upgrading existing production processes rather than pursuing disruptive process innovations. This is because the technological paths in mature industries are relatively fixed, asset specificity is high, changing production processes entails enormous costs, and the benefits of clean technology innovation are difficult to quickly translate into market share. As a result, firms prefer to improve pollution control technologies within existing frameworks.

Table 11. Estimated results of industry heterogeneity: low-pollution industries.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.038 ***	0.029 ***	0.034 ***	0.019 *	0.011
	(0.011)	(0.009)	(0.011)	(0.011)	(0.012)
control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.679	0.645	0.649	0.621	0.621
Observations	16,565	16,565	16,565	16,565	16,565

Notation: Robust standard errors in ( ). *** and * indicate passing significance tests at 1% and 10% significance levels, respectively. Y for represents the control.

exhibits the test results for the low-pollution industries. The results indicate that robot applications significantly promote clean production technology innovation in low-pollution industries. Possible explanations include the following: First, low-pollution industries inherently have lower emissions and face less regulatory pressure, making the marginal benefits of end-of-pipe governance relatively low. Although clean production technologies may require high upfront investment, they can save raw materials and reduce the waste of energy in production, and bring about an improvement in the productivity of the enterprise in the long run, aligning with the profit models of low-pollution industries. Second, low-pollution industries are more likely to be in the growth or emerging stages, with unfixed technological trajectories and minimal sunk costs, making it easier to adopt new production processes and technologies for source reduction. Additionally, low-pollution industries may rely more on technological innovation to establish market barriers, thus proactively investing in clean production technologies to form differentiated competition.

6.2. Heterogeneity in the Level of Environmental Regulation

Traditional economics posits that environmental regulations increase corporate costs and weaken price competitiveness. Porter, however, argues that this view overlooks the innovation compensation effect. Porter’s hypothesis states that environmental regulations compel firms to re-examine production processes, offset compliance costs through green technological innovation, and enhance competitiveness. For different regions in China, there are large differences in industrial structure and economic development, and the intensity of government environmental regulations for firms in different regions varies. Thus we ask, does this difference in the intensity of environmental regulation affect the power of AI in green technological innovation? Thinking about and testing this question can precisely verify whether the Porter hypothesis just mentioned holds true in China. Drawing on Javorcik and Wei (2003) [57], industrial sulfur dioxide, industrial wastewater, and industrial soot emissions are summed, and the summed value as a share of GDP measures the intensity of environmental regulation. Based on whether the value of each prefecture-level city in the sample time frame is greater than the corresponding median, the enterprises are categorized into two groups, those subject to stringent environmental regulation and those not subject to stringent environmental regulation, and grouped into regressions.

The results are presented in

Table 12. Estimated results of environmental regulation heterogeneity: stringent environmental regulation.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.044 ***	0.027 **	0.029 **	0.035 **	0.031 **
	(0.012)	(0.011)	(0.013)	(0.015)	(0.015)
control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.693	0.665	0.663	0.637	0.637
Observations	20,165	20,165	20,165	20,165	20,165

Notation: Robust standard errors in ( ). *** and ** indicate passing significance tests at 1% and 5% significance levels, respectively. Y for represents the control.

and

Table 13. Estimated results of environmental regulation heterogeneity: lax environmental regulation.

	(1)	(2)	(3)	(4)	(5)
Variables	ginnov_total	ginnov_ce		ginnov_de
		nylnno	jnlnno	fwlnno	yslnno
lnexposure	0.013 *	0.008	0.011 *	0.009	0.010
	(0.007)	(0.022)	(0.006)	(0.017)	(0.016)
control	Y	Y	Y	Y	Y
Firm FE	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y
R-squared	0.632	0.579	0.579	0.535	0.535
Observations	4072	4072	4072	4072	4072

Notation: Robust standard errors in ( ). * indicate passing significance tests at 10% significance levels, respectively. Y for represents the control.

. It can be seen that robot applications have a significant effect on green technological innovation for enterprises in regions with strong environmental regulation constraints, while this effect is not evident for enterprises in weak environmental regulation regions. A possible explanation is that under low-level environmental regulations, the “crowding-out effect” of corporate costs dominates: expenses for pollution control and emission reduction divert funds originally allocated to green technological innovation, leading firms to proactively reduce investments in such innovation and consequently lowering their green technological innovation levels. Meanwhile, in weak-regulation regions, investments in green technological innovation may be stranded by policy changes, creating sunk-cost risks that inhibit the adoption of long-cycle technologies like artificial intelligence.

Regions with strong environmental regulations often undergo dynamic policy adjustments, such as annual increases in carbon tax rates and iterative updates to environmental standards, exposing enterprises to higher intertemporal compliance risks. In such cases, pollution control costs surge far beyond the expenses required for green technological innovation, rendering traditional production methods and equipment incapable of meeting strict environmental constraints. Driven by profit maximization, firms seek smarter energy-saving and emission-reduction technologies—such as introducing robotic intelligent equipment—to abandon pollution-related technologies while promoting green innovation. The enterprise “compensation effect” gradually outweighs the “crowding-out effect”, validating the Porter hypothesis in the Chinese context: strengthening environmental regulation levels motivates more firms to abandon high-energy technologies and prioritize low-energy, low-emission, and innovation-driven approaches.

6.3. Heterogeneity in Enterprise Ownership

The impact of AI on a company’s green technology innovation may vary depending on the company’s ownership. In order to test whether this difference exists, this paper divides firms according to the nature of their ownership into two categories: state-owned firms and non-state-owned firms. Grouped regression tests are conducted to examine whether there are differences in the impact of AI on green technological innovation under different ownership structures.

Table 14. Estimation results of enterprise ownership heterogeneity.

	non-SOEs	SOEs
Variables	ginnov_total	ginnov_total
lnexposure	0.012	0.036 ***
	(0.015)	(0.012)
control	Y	Y
Firm FE	Y	Y
Year FE	Y	Y
R-squared	0.679	0.512
Observations	8839	16,417

Notation: Robust standard errors in ( ). *** indicate passing significance tests at 1% significance levels, respectively. Y for represents the control.

demonstrates the results of the test that AI application has a significant contribution to green technology innovation in state-owned companies, but does not have a significant effect on non-state-owned companies. Possible explanations include the following: On the one hand, SOEs must closely align with national development strategies during their growth. Under China’s policy of accelerating the promotion of intelligent manufacturing, SOEs can more keenly perceive the policy orientation of AI and demonstrate higher enthusiasm for intelligent and green transformation. On the other hand, SOEs have more complete green management talent teams and R&D systems for green technological innovation, coupled with strong financing capabilities and high innovation levels, while non-state-owned firms are constrained by funding, technology, and talent shortages. The adoption of AI may increase production costs, which is detrimental to their green technological innovation.

7. Conclusions and Policy Recommendations

7.1. Conclusions

This study categorizes green technology innovation into clean production technology innovation and pollution control technology innovation. On the basis of the theoretical analysis, the role of AI in green technology innovation was tested using a double fixed-effects model with data from Chinese listed companies from 2006 to 2020, and the mechanism of action and heterogeneity were further explored.

The following conclusions are obtained: (1) AI significantly contributes to firms’ green technology innovations, with cleaner production technologies playing a more significant role in terms of innovation in the production chain. A series of robustness tests support this conclusion. (2) AI promotes green technological innovation in firms by improving the quality of human capital and enterprise efficiency. (3) Heterogeneity analysis found that the greater the intensity of environmental regulation a firm faces, the greater the incentive for the firm to use AI for green technology innovation; the green innovation effect of AI in firms in high-pollution industries is mainly manifested in the innovation of pollution control technology; firms in low-pollution industries tend to carry out technological innovation of clean production; and the application of AI significantly promotes the green technology innovation of state-owned firms.

7.2. Policy Recommendations

The research in this paper not only supports the role of artificial intelligence in promoting green technological innovation, but also provides theoretical support for the government to formulate relevant measures to promote green technological innovation:

(1) We must further strengthen basic research on artificial intelligence and vigorously support and develop related industries; undertake the classification of policies, broaden the application of artificial intelligence in enterprise production scenarios, and effectively extend, replenish, and strengthen the chain of the artificial intelligence industry; pool advantageous research institutes, enterprises, human capital, and funds in key core areas of AI to rapidly enhance the local supply capacity of AI; and reduce the burden of the replacement costs of AI applications on businesses through government subsidies. From the perspective of sustainability, government subsidies tilted towards green AI technologies can guide enterprises to choose environmentally friendly innovation paths, synchronize industrialization and decarbonization, and fit the synergistic goals of SDG9 (promoting sustainable industrialization) and SDG13 (climate action).

(2) Localities should introduce locally oriented environmental regulatory initiatives to force companies to adopt more advanced robots. The government should use R&D subsidies and R&D expense deductions for enterprises in low-pollution industries to promote technological innovation in cleaner production. The government should adopt control-oriented tools for enterprises in heavily polluting industries to promote technological innovation in pollution control; at the same time, it also needs to collaborate in adopting market incentive tools, such as financial subsidies and tax exemptions, to guide enterprises to turn to technological innovation in cleaner production, promote synergistic innovation between front-end cleaner production and back-end pollution control. This combination of the two realizes the dual control of “production and emission”, which embodies the concept of “establishing a whole chain environmental management system from source to end” in SDG12.

(3) Enterprises should strengthen the training of people with composite talents who understand both AI technology and green innovation; construct a perfect talent training system and form a talent echelon resource with continuous strength; increase the training of people with digital talents to improve workers’ knowledge of new technologies and environmental awareness; attract outstanding talent through national and regional talent programs to fulfill the demand for high-caliber human capital in green technological innovation; and support the application of artificial intelligence with talent team building, and inject new kinetic energy for green technology innovation. The above measures, through the closed loop of “innovation of education content, improvement of training system, and optimization of human resources structure”, not only directly implement the specific goal of “quality education and lifelong learning” in SDG4, but also indirectly promote the realization of SDG9 (innovation and industrialization) through the support of human resources. This approach also indirectly promotes the realization of SDG9 (innovation and industrialization) through the support of talented people, reflecting the synergistic and systematic nature of the SDGs.

8. Discussion and Future Research

8.1. Discussion

With the development of a new generation of information technology and the rapid development of intelligent manufacturing, the application of artificial intelligence is quietly changing the enterprise production mode. Thus we ask, can artificial intelligence drive enterprise green technology innovation? Around this issue, this paper conducts a research investigation. We compared the findings of this study with the existing literature and the consistency of the findings was as follows: First, AI application has a facilitating effect on enterprise innovation, similar to the findings of the established related literature [11,12,14]. Next, AI application has a positive environmental effect and can drive green technology innovation, similar to the findings of the established related literature [10,17,18,19].

The differences between the studies in this paper are as follows: On the one hand, this study finds differences in AI promoting green technology innovation based on the difference in production links, and the green technology innovation effect of AI is more significant for the production link. On the other hand, this paper distinguishes and tests the heterogeneity of AI’s effect on green technology innovation and finds that firms in high-pollution industries tend to favor pollution control technology innovation; firms in low-pollution industries tend to favor cleaner production technology innovation; and as firms face a greater intensity of environmental regulation, there is a greater incentive to use AI to facilitate green technology innovation.

8.2. Future Research

The research in this paper not only contributes to the understanding of the positive environmental externalities of AI technologies, but also provides useful references for green and sustainable development in emerging economies. Despite the detailed justification of this study, some aspects still deserve further exploration:

Firstly, this paper has some limitations in the measurement of artificial intelligence. Although this study draws on the mainstream literature’s approach of using robot penetration in industrial firms to measure corporate AI applications, modern AI also includes generative modeling and digital platforms, and robotics is not sufficient to cover the development of AI. Therefore, future research could consider integrating diverse AI indicators to more comprehensively analyze and test the impact of AI on green technology innovation.

Secondly, this study is based on a sample of listed companies, but the specificity of listed companies cannot easily represent the whole. Taking only listed firms as the sample may overestimate the “universal effect” of AI on green innovation and neglect the lagging application of AI by SMEs due to financial and technological barriers. In the future, the study can include non-listed companies and supplement micro-level information through questionnaires and industrial census data, comparing the green innovation effects of artificial intelligence on firms of different sizes and attributes.

Thirdly, the object and sample of this paper is Chinese firms, and the generalizability of the results may be limited by China’s unique policy environment. China’s “dual-carbon” goal and state-led innovation model have shaped the underlying logic of green technological innovation, which is significantly different from that of market-driven economies. Future research can take the “institutional environment” as the core regulating variable, and through cross-country, cross-regional, and cross-industry comparative studies reveal the differentiated paths of AI green innovation under different policy-market combinations, so as to build a more generalized theoretical framework.

Finally, this paper examines the positive environmental externalities of AI, but along with the amplification of the positive effects of AI embedded in green innovation, its negative impacts have begun to come to the fore. There are potential risks that artificial intelligence technology may pose to environmental governance. This risk comes from the technical defects of AI itself on the one hand, and the profound impact of AI on government decision-making mechanisms, information dissemination mechanisms, and social organizational structures on the other. Along with the accelerated and complex evolution of AI, such risks may spread to various aspects of technology, ethics, and the rule of law, exacerbating the public’s distrust of the technology itself and even of government departments. In the future, further research can be conducted on how to utilize the technological dividends of AI while building a green innovation form that can more effectively prevent technological risks.