The Impact of AI on Corporate Green Transformation: Empirical Evidence from China

Abstract

With the rapid advancement of artificial intelligence (AI), its deep integration into corporate operations has become the key driver for firms to reconfigure factor resources, boost green total factor productivity, and achieve green transformation. This analysis empirically investigates the influence of AI on corporate green transformation using panel data of China’s listed companies from 2015 to 2022. This research employs a multidimensional fixed effects linear model to analyze the relationship, finding that AI significantly enhances corporate green transformation. Mechanism analysis reveals that AI promotes green transformation by enhancing firm research and development (R&D) and firm green innovation capabilities. Heterogeneity analysis shows that the positive impact of AI on corporate green transformation is more significant in the eastern region, post-COVID−19, and in low-pollution industries. The impact is also significantly and positively moderated by the development of the non-state-owned economy and the development degree of product markets. These findings suggest that AI is a critical tool for promoting sustainable economic growth and green transformation in businesses.

1. Introduction

In 2020, OpenAI released GPT−3, an unprecedented large-scale language model that ushered in a new wave of AI. Countries including China, the United States, and India maintain an open attitude towards the development of AI, promoting AI technology research and application through policy support and incentives. As one of the key players in this AI revolution, China has recognized the strategic importance of AI. Hence, DeepSeek was launched in 2023, which has rapidly emerged as a benchmark enterprise in global AI development through its efficient technological pathways and cost-competitive strategies, establishing itself as a strategic player in the Sino–U.S. AI competitive landscape. The rapid development of science and technology, epitomized by AI, provides technical support for green transformation and is an important force in promoting green transformation [1].

Against the backdrop of the fifth technological revolution, the application of AI, machine learning (ML), and big data is changing how natural resources are managed [2]. As the most dazzling star among them, AI is gradually attracting extensive attention from both the academic community and ordinary users. AI technologies are predominantly applied in various domains, including computer vision [3,4], healthcare [5], industrial upgrading [6,7,8], the financial sector [9], human resource management [10], cultural education [11,12], and autonomous driving [13]. As a core technology of the digital economy, AI is spearheading the transformation towards sustainable economic growth [14].

The available literature generally equates the level of AI application with “automation”, using automation and task complementarities as a proxy variable [15]. Although this approach is intuitive, it has two biases. One is the sectoral coverage bias, which overemphasizes the use of AI in the manufacturing industry while neglecting its use in the service industry. The other is the technology typology bias, which fails to include non-mechanical AI forms such as machine learning. This research innovatively adopts the text mining method to quantify the depth of firm-level AI usage through the frequency of keywords and semantic embedding vectors included in annual reports. This method not only avoids the measurement bias that may be caused by automation indicators but also captures the managerial strategic prioritization of AI, thereby providing a more robust measurement basis for assessing the real effect of AI-driven green transformation.

AI not only enhances enterprises’ analytical capabilities regarding economic data but also plays a part in complex patterns of economic behavior, playing an increasingly vital role in green transformation within Chinese firms. With its powerful data processing capabilities and intelligent analytical models, AI technology can provide innovative solutions for enterprises in key operational areas, such as energy management [16], supply chain resilience [17], product lifecycle management [18], and waste management and recycling [19], thereby achieving cost control and efficiency enhancement. Previous research has shown that AI can optimize business processes, reduce operating costs, improve enterprises’ risk control capabilities, and show great potential in automating micro-tasks and enhancing research efficiency [20].

Furthermore, the existing research indicates that AI technology may also affect enterprises through the following pathways. Within the realm of employment, the application of AI technology will lead to labor substitution [21]. The creation of new job positions that emphasize the application of knowledge and skills will enhance enterprise production efficiency and cause a transformation in resource utilization methods. Within the realm of innovation capabilities, Liu et al. [22] propose that AI technology can accelerate knowledge creation and technological spillovers within firms, improve learning and absorption capabilities, and promote technological innovation. Corporate green innovation, including green technological innovation, green management innovation, and green product innovation, strengthens the effect of AI on carbon emissions reduction [23]. In the sphere of production activities, AI helps enterprises identify wasteful activities in the production process [24], acquire and apply the latest scientific and technological advancements, achieve intensified production transformation through technological improvements, and enhance sustainable development. AI can also improve environmental performance via green innovation of products and processes [25]. Some researchers also believe that advancements in AI technology will impact firm managers, encouraging greater participation in the global division of labor and international competition [26].

Indeed, green transformation has emerged as a global priority, underscored by China’s proactive stance that is encapsulated in the slogan “Clear waters and green mountains are as good as mountains of gold and silver”. Green transformation can be propelled by environmental regulation through two channels, green technology innovation and industrial structure upgrading [27,28]. Moreover, firm-level green transformation forms the micro-foundation of China’s national sustainability goals. Green total factor productivity (GTFP) incorporates energy consumption and pollution emissions into the productivity accounting framework simultaneously, which is used as the proxy variable for measuring an enterprise’s green transition [29]. Against the current backdrop of global sustainable transformation, AI will urge enterprises to implement the concepts of resource conservation and environmental protection throughout the entire production and operation process, balancing economic value with ecological value and achieving green transformation. However, AI technology also faces certain challenges while driving green transformation [30], such as technological limitations [31], operational costs, and the potential “rebound effect”. Given these considerations, does the application of AI within enterprises affect their green transformation, particularly in terms of enhancing green total factor productivity (GTFP)? Based on the above facts, we explore whether the use of AI improves corporate green transformation.

Using panel data from A-share traded enterprises in China over the period from 2015 to 2022, empirical analyses demonstrated convincingly that AI systematically enhances corporate green transformation. The findings of heterogeneity tests revealed that this enhancement is stronger in the eastern region, post-COVID−19, and in low-pollution industries. Through mechanisms analysis, we identified that firm R&D and firm green innovation capabilities are two channels through which AI facilitates corporate green transformation. Furthermore, the impact of AI on corporate green transformation is also positively moderated by the development of the non-state-owned economy and the development degree of product markets.

Theoretically and practically, our research presents three major contributions. First, we present evidence of the positive effect of AI on corporate green transformation. The findings supplement the literature regarding the impact of AI at the firm level. Second, we introduce a novel approach to measuring the application of AI technology at the firm level by analyzing AI-related key terms in annual reports of listed companies, thereby avoiding biases associated with automation and expanding empirical research possibilities in terms of data selection. Finally, we show that the use of AI in enterprises in the eastern regions, the period after the pandemic, and in low-pollution industries has a more significant impact on green transformation, assisting the government in formulating targeted, effective, and sustainable economic policies.

The subsequent sections are as follows: Section 2 is the literature review and hypothesis. Section 3 outlines the methodology. Section 4 presents the empirical analysis. Section 5 examines the possible economic mechanisms of the impact of AI on corporate green transformation. Section 6 analyzes the moderating effect of marketization. Section 7 offers the conclusions and implications.

2. Literature Review and Hypothesis

2.1. Resource-Based View Theory

The resource-based view (RBV) theory is a pivotal strategic management framework that identifies strategic resources that underpin sustained competitive advantage, suggesting that corporate intangible assets are no less vital than tangible resources [32]. It highlights firm performance determinants through heterogeneous resource allocation. Barney [33] believes that value, rarity, imitability, and substitutability are four major empirical indicators of the potential of firm resources to generate sustained competitive advantage.

On the one hand, the emphasis on firm R&D serves as the core driving force behind green transformation, transforming “strategic intentions” into “resource allocation”. When management regards AI technology as a growth mechanism for future profits, R&D investment, the talent team, and the patent layout will systematically tilt towards the green direction. Conversely, if insufficient attention is paid, funds and human resources will flow towards expansion within a traditional capacity. Therefore, the degree of emphasis directly determines the scale, sustainability, and pertinence of R&D investment, and further determines whether an enterprise can accumulate key green knowledge, break through technical bottlenecks, and ultimately achieve dual optimization of emissions and output. On the other hand, firm green innovation capabilities provide the organizational internal soft foundation necessary for technological R&D. These capabilities lower the marginal cost of assimilating external scientific advances, mitigate the risk of technological lock-in, and accelerate the iterative cycle from lab prototype to scalable market application. Together, the “hard” commitment of R&D investment and the “soft” architecture of green transformation capability form a mutually reinforcing loop: emphasis expands the technological frontier, while capabilities ensures that new knowledge is efficiently internalized, recombined, and diffused within the firm, thereby sustaining a dynamic path of green total factor productivity growth, that is, the dynamic path of corporate green transformation.

Based on this framework, this research investigates the emphasis on firm R&D and firm green innovation capabilities, the mechanisms through which AI influences corporate green transformation.

2.2. AI Application and Corporate Green Transformation

In the contemporary landscape of technological advancement, AI has emerged as a pivotal force, poised to revolutionize various facets of business operations. The integration of AI into corporate strategies is anticipated to catalyze a significant shift towards sustainable practices, thereby fostering corporate green transformation.

Firstly, AI is capable of optimizing resource allocation and enhancing operational efficiency [26,34]. Also, AI contributes to green production [35] by meticulously analyzing vast datasets to identify inefficiencies and redundancies within production processes, thereby promoting green transformation.

Secondly, AI can foster innovation, including innovation capacity [22,36] and innovation efficiency [37]. It provides firms with tools to develop novel products and services that are more sustainable. AI can facilitate the design of energy-efficient products, which allows for the optimization of product features and materials to reduce energy consumption during usage. These are crucial in a rapidly evolving market where firms must stay ahead of the curve to remain competitive and environmentally responsible.

Lastly, AI enhances decision-making processes within firms by providing real-time data analysis and insights. AI can analyze consumer behavior and market trends to identify opportunities for green marketing strategies, thereby encouraging consumer adoption of sustainable products and services. This capability empowers managers to make informed decisions that align with sustainability goals. Therefore, we propose Hypothesis 1.

Hypothesis 1.

AI can promote corporate green transformation.

2.3. The Mechanisms That Link AI and Corporate Green Transformation

While promoting corporate green development, AI not only directly promotes corporate green transformation, but also indirectly contributes to it by enhancing the emphasis on firm R&D and firm green innovation capabilities. With the advancement of AI technology, enterprises will face accelerated technological iteration. To maintain a leading position and to gain market competitiveness, it is essential to strengthen the role of R&D activities in corporate operations and continuously increase investment in order to carry out innovative activities. R&D investment is important for promoting corporate green transformation [38]. It not only drives enterprises to make technological improvements in energy saving, emission reduction, clean energy, and waste recycling, thereby reducing production energy consumption and improving utilization efficiency, but it also optimizes production processes [39] and promotes clean production [40], thereby advancing the green transformation of enterprises.

AI can effectively enhance a company’s green innovation capabilities, which include firm green independent innovation capabilities and firm green collaborative innovation capabilities. In independent corporate R&D activities, machine simulation and testing has shown that AI technology can significantly shorten the product development cycle and reduce trial-and-error costs in independent R&D activities, providing resources and a better internal environment for corporate green transformation [41]. In collaborative R&D activities, AI technology can achieve information integration and sharing more effectively, reducing coordination costs that arise from information asymmetry in the cooperation process, and establishing an efficient and stable collaborative platform, stimulating corporate green transformation [42].

Both of the above situations can promote corporate green transformation. Based on this, we propose Hypotheses 2 and 3.

Hypothesis 2.

AI promotes green transformation by enhancing the emphasis on firm R&D.

Hypothesis 3.

AI promotes green transformation by enhancing firm green innovation capabilities.

2.4. The Moderating Effect of Marketization

The marketization level captures provincial differences in resource-allocation efficiency, competitive intensity, and policy predictability [43]. It plays a pivotal moderating role in the relationship between AI and corporate green transformation.

The marketization level significantly shapes resource-allocation efficiency [44]. In highly marketized regions, price signals are more sensitive and cross-firm mobility costs are lower. Price mechanisms guide capital, technology, and other high-end elements toward high-productivity new-growth sectors more effectively, thereby amplifying the growth-boosting role of AI. Firms seeking excess profits are therefore inclined to utilize AI for the purpose of green transformation.

A higher degree of marketization helps lower the cost of AI adoption and raises the marginal return on AI application. Because of strengthened intellectual-property protection and reduced policy uncertainty, provinces with advanced marketization significantly reduce transaction costs from R&D to application. At the same time, well-developed market intermediaries reduce uncertainty when adopting AI, thereby increasing the marginal-return curve of green transformation investment.

In regions where marketization is more advanced, market competition mechanisms are more complete, and competition there is fiercer, leading firms to place greater emphasis on increasing technological innovation inputs and improving production efficiency [45], enhancing the quality and competitiveness of products and services, and thereby providing endogenous momentum for corporate green transformation. Based on this analysis, we propose the following hypothesis:

Hypothesis 4.

Marketization exerts a positive moderating effect between AI and corporate green transformation.

3. Methodology

3.1. Model Setting

3.1.1. Baseline Regression Model

To control the potential endogeneity problem, we adopted the dynamic panel data model for estimation [46] and introduced the second-order lag term of the explained variable to better capture the dynamic structure. Although the statistical diagnostic test results of the model are good, the estimation results show unreasonable economic implications that green total factor productivity itself lacks sustainability, which is seriously inconsistent with theoretical expectations and the existing literature. The GMM model may have overcorrected or filtered out the key variations in the data in the context of this study, resulting in the estimation results losing their economic significance. Therefore, the lagged dependent variable should not be included in the estimation. The following multidimensional fixed effects linear model was applied to explore the impact of AI on corporate green transformation:

$${GTFP}_{it}={\alpha }_0+{\alpha }_1{AI\_1}_{it}+{\sum }_{q=1}^Q{\gamma }_q{Controls}_{it}^q+{\delta }_t+u_k+{\epsilon }_{it}$$

(1)

where ${AI\_1}_{it}$ is the independent variable representing the degree of AI usage with a one-period lag of enterprise i in year t. ${GTFP}_{it}$ is the dependent variable representing a firm’s green transformation performance for year t. ${Controls}_{it}^q$ is a set of control variables. ${\delta }_t$ and ${\mathrm{u}}_k$ represent the time fixed effect and the firm fixed effect. ${\epsilon }_{it}$ is the random disturbance term. The coefficient ${\alpha }_1$ is of particular interest. If ${\alpha }_1$ is significantly positive, it indicates that AI can promote corporate green transformation, and Hypothesis 1 holds true.

3.1.2. Mediation Effect Model

To validate Hypothesis 2 and examine the mediating role of the emphasis on corporate R&D, this chapter constructs the following mediation effect model based on the baseline regression model.

$${RDemphasis}_{it}={\beta }_0+{\beta }_1{AI\_1}_{it}+{\sum }_{q=1}^Q{\gamma }_q{Controls}_{it}^q+{\delta }_t+u_k+{\epsilon }_{it}$$

(2)

In this equation, ${RDemphasis}_{it}$ is our dependent variable, measuring the emphasis on firm R&D.

To validate Hypothesis 3 and examine the mediating role of firm green innovation capabilities, this chapter constructs the following mediation effect model based on the baseline regression model.

$${FGIC}_{it}={\beta }_0+{\beta }_1{AI\_1}_{it}+{\sum }_{q=1}^Q{\gamma }_q{Controls}_{it}^q+{\delta }_t+u_k+{\epsilon }_{it}$$

(3)

Here, ${FGIC}_{it}$ represents our mediating variable, measuring a firm’s green innovation capabilities, which are constituted by two components: firm green independent innovation capabilities (FGIC1) and firm green collaborative innovation capabilities (FGIC2).

3.1.3. Moderating Effect Model

To validate Hypothesis 4 and verify whether the marketization level can empower the role of AI in promoting corporate green transformation, we established the following moderating effect model.

$${GTFP}_{it}={\alpha }_0+{\alpha }_1{AI\_1}_{it}+{\alpha }_2{market}_{it}+{\alpha }_3{{AI\_1}_{it}\times market}_{it}+{\sum }_{q=1}^Q{\gamma }_q{Controls}_{it}^q+{\delta }_t+u_k+{\epsilon }_{it}$$

(4)

This model allowed us to examine the role of the NERI index of marketization in the use of AI, which affects the green transformation. We focused on the coefficient ${\alpha }_3$ of the interaction term ${{AI\_1}_{it}\times market}_{it}$, which indicates the presence and direction of the moderating effect.

3.2. Variables

3.2.1. Explained Variable: Green Transformation

Green transformation (GTFP): We used the firm-level GTFP to measure green transformation. Referring to Färe and Grosskopf [47], we measured firm-level GTFP with the non-radial slacks-based measure (SBM) model and the Malmquist–Luenberger (ML) index. Green technical efficiency change (GTEC) and green technological change (GTC) are two sub-indicators of GTFP. GTC focuses on technological advancements, while GTEC emphasizes the optimization of efficiency in the production process.

3.2.2. Explanatory Variable: Artificial Intelligence

Artificial intelligence (AI_1): We calculated the word frequency of AI-related keywords in the annual reports of listed companies. Considering the typical “right-bias” characteristic of the original data, we used the natural logarithm of the original data plus 1. Because of the lagged effect of AI application on corporate GTFP, and to control the endogeneity caused by reverse causality, we adopted a lagged one-period term of AI in the regression.

3.2.3. Control Variables

We incorporated corporate control variables of three dimensions: (1) general characteristics variables including the enterprise age ($\mathrm{l}\mathrm{n}\mathrm{a}\mathrm{g}\mathrm{e}$), the firm size ($\mathrm{l}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{z}\mathrm{e}$), the enterprise growth rate ($\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{w}$), the nature of enterprise ownership ($\mathrm{s}\mathrm{t}\mathrm{o}\_\mathrm{e}\mathrm{n}\mathrm{t}$), and firm R&D intensity ($\mathrm{R}\_\mathrm{D}$); (2) governance characteristics variables including a dual chairman–CEO role ($\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}$) and top shareholder ownership ($\mathrm{o}\mathrm{w}\mathrm{n}\mathrm{e}\mathrm{r}$); and (3) financial characteristics variables including enterprise financial health ($\mathrm{f}\mathrm{i}\mathrm{n}\_\mathrm{h}\mathrm{e}\mathrm{a}\mathrm{l}$), the return on net assets ($\mathrm{R}\mathrm{O}\mathrm{E}$), Tobin’s Q ratio ($\mathrm{T}\mathrm{B}\mathrm{Q}$), the book-to-market ratio of enterprises ($\mathrm{B}\mathrm{E}$), the current ratio ($\mathrm{C}\mathrm{R}$), and audit opinion ($\mathrm{a}\mathrm{u}\mathrm{d}\mathrm{i}\mathrm{t}$).

3.2.4. Mediating Variables

(1)

Emphasis on firm R&D (RDemphasis). In financial fields, the amount of funds allocated to R&D-related activities reflects the level of corporate emphasis on research and development. Given this, we used the natural logarithm of corporate R&D expenditures to measure the level of emphasis on firm R&D.

(2)

Firm green innovation capabilities (FGIC). We used the number of corporate green patents to measure the green innovation capabilities of enterprises. Specifically, we employed the number of independent green invention patents applied for by the enterprise to measure its green independent innovation capability. The number of joint green invention patents applied for by the enterprise was used to measure its green collaborative innovation capability. To exclude the “right-skewness” issue of green patent data, the original data mentioned above were all processed by adding one and then taking the logarithm.

3.2.5. Moderating Variable: Marketization Level

Marketization level (market): We measured the marketization level using the National Economic Research Institute (NERI) marketization index for China’s provinces, which includes five primary variables of the marketization index, the government–market relationship, the development of the non-state-owned economy, the development degree of product markets, the development degree of factor markets, and the development of market intermediary organizations and the legal institutional environment [43]. All components were measured on a 0–10 scale. The definitions and quantification of all variables can be found in

Table 1. Definition of variables.

Variable	Definition	Description or Calculation Method
GTFP	Green total factor productivity	Eco-efficient output per composite input, estimated via the non-radial slacks-based measure (SBM) model and the Malmquist–Luenberger (ML) index.
AI_1	Artificial intelligence	The lagged one-period word frequency of AI-related keywords in the annual reports of listed companies.
lnage	Firm age	The logarithm of years since incorporation.
lnsize	Firm size	The logarithm of year-end total assets.
grow	Enterprise growth rate	The year-over-year percentage change in operating revenue.
sto_ent	Nature of enterprise ownership	A value of 1 is assigned if the ultimate controller is the state; otherwise, a value of 0 is assigned.
R_D	Firm R&D intensity	R&D expenditure divided by the operating revenue.
dual	Dual chairman–CEO role	A value of 1 is assigned if one person holds both titles; otherwise, a value of 0 is assigned.
owner	Top shareholder ownership	The percentage of shares held by the largest shareholder.
fin_heal	Enterprise financial health	The net annual amount of cash and cash equivalents generated or consumed by a firm’s core operating activities.
ROE	Return on net assets	The net profit divided by average net assets.
TBQ	Tobin’s Q ratio	(The market value of equity + book value of debt)/total assets.
BE	Firm book-to-market ratio	The book equity value divided by market capitalization.
CR	Current ratio	Current assets divided by current liabilities.
audit	Audit opinion	A value of 1 is assigned if the external auditor issues a standard unqualified opinion; otherwise, a value of 0 is assigned.
RDemphasis	Emphasis on firm R&D	The logarithm of the amount of funds allocated to R&D-related activities.
FGIC1	Green independent innovation capability	The number of independent green invention patents applied for by the enterprise.
FGIC2	Green collaborative innovation capability	The number of joint green invention patents applied for by the enterprise.
market	Marketization level	The National Economic Research Institute (NERI) marketization index for China’s provinces, which includes five primary variables of the marketization index.

.

3.3. Sources of Data and Preliminary Analysis

Our initial sample included all A-share listed companies on the Shanghai Securities Exchange and Shenzhen Securities Exchange from 2015 to 2022. The sample data was sourced from the China Stock Market and Accounting Research (CSMAR) database. To enhance the reliability of the data, the following screening and processing procedures were applied to the raw data: (1) Listed companies from the financial and real estate industries were excluded. (2) Listed companies marked as ST and *ST were excluded. A company was labeled ST (Special Treatment) when it had posted two consecutive years of losses or when auditors issued a negative opinion. In addition, a company was upgraded to *ST when it was already ST and faced imminent delisting risk—typically after three straight loss-making years, severe asset impairment, or major compliance violations. (3) Listed companies with missing key variables were excluded.

Table 2. Descriptive statistics.

Variable	Obs	Mean	SD	Min	Max
GTFP	25,681	1.0631	0.0635	0.9333	1.1760
AI_1	25,681	0.4090	0.8072	0.0000	5.5568
lnage	25,681	2.2502	0.7746	0.6931	3.4965
lnsize	14,974	7.4465	1.1972	0.6931	13.2535
grow	25,543	3.5670	376.5769	−29.4764	59,400
sto_ent	25,018	0.3359	0.4723	0.0000	1.0000
R_D	18,678	5.1925	8.0056	0.0000	424.9300
dual	24,845	0.3012	0.4588	0.0000	1.0000
owner	25,577	32.8832	14.6848	0.2863	89.9910
fin_heal	25,574	0.0480	0.0788	−1.6863	2.2216
ROE	25,426	−0.0073	3.0119	−207.3971	281.9892
TBQ	25,164	2.2739	5.3632	0.6245	729.6293
BE	25,164	0.6117	0.2641	0.0014	1.6012
CR	25,577	0.5639	0.1995	0.0087	1.0000
audit	24,744	1.1307	0.6800	1.0000	6.0000
RDemphasis	18,454	101.6637	30.9501	0.0000	251.6907
FGIC1	25,612	0.5807	0.9713	0.0000	6.8046
FGIC2	25,612	0.1807	0.5800	0.0000	6.7226
market	25,560	10.0658	1.6455	1.1260	12.8640

presents descriptive statistics for a dataset comprising 25,681 observations of 4103 listed companies across all industries, excluding the financial and real estate industries. The green total factor productivity has a maximum value of 1.1760, a minimum of 0.9333, a mean of 1.0631, and a standard deviation of 0.0635. This indicates that the sample exhibits uniformly good green productive performance with modest cross-firm differences, providing a robust baseline for further empirical analysis.

4. Empirical Analysis Results

4.1. Benchmark Regression Results

The impacts of listed companies’ use of AI on their green transformation are shown in

Table 3. Baseline regression results.

VARIABLES	(1)	(2)	(3)	(4)	(5)
	GTFP	GTFP	GTFP	GTEC	GTC
AI_1	0.0195 ***	0.0029 ***	0.0144 ***	0.0021 ***	0.0023 ***
	(0.0005)	(0.0003)	(0.0006)	(0.0004)	(0.0003)
lnage			0.0046 ***	0.0299 ***	0.0315 ***
			(0.0007)	(0.0009)	(0.0008)
lnsize			−0.0048 ***	−0.0006	−0.0006
			(0.0005)	(0.0008)	(0.0007)
grow			−0.0001	−0.0000 **	−0.0000 *
			(0.0000)	(0.0000)	(0.0000)
sto_ent			0.0079 ***	0.0014	−0.0001
			(0.0021)	(0.0013)	(0.0013)
R_D			0.0003 ***	0.0001	0.0001 *
			(0.0001)	(0.0001)	(0.0001)
dual			0.0049 ***	0.0005	−0.0005
			(0.0010)	(0.0007)	(0.0007)
owner			−0.0001 ***	−0.0002 ***	−0.0002 ***
			(0.0000)	(0.0000)	(0.0000)
fin_heal			0.0860 ***	0.0102 ***	0.0114 ***
			(0.0069)	(0.0031)	(0.0035)
ROE			0.0000	0.0001	0.0000
			(0.0002)	(0.0001)	(0.0000)
TBQ			−0.0002	−0.0001	−0.0001
			(0.0003)	(0.0001)	(0.0001)
BE			0.0491 ***	0.0234 ***	0.0193 ***
			(0.0028)	(0.0016)	(0.0017)
CR			0.0183 ***	0.0132 ***	0.0143 ***
			(0.0029)	(0.0027)	(0.0023)
audit			0.0010	−0.0002	−0.0002
			(0.0008)	(0.0005)	(0.0006)
Constant	1.0551 ***	1.1324 ***	1.0294 ***	1.0372 ***	1.0360 ***
	(0.0004)	(0.0003)	(0.0046)	(0.0076)	(0.0076)
Time FE	NO	YES	YES	YES	YES
Firm FE	NO	YES	YES	YES	YES
Observations	25,681	25,471	12,631	12,272	12,272
R-squared	0.0617	0.9065	0.1033	0.8986	0.9010

Note: Robust standard errors are shown in parentheses; *, **, and *** denote significance at the 10%, 5%, and 1% levels, respectively.

. In Column (1), we regress without control variables and fixed effects, finding that the coefficient is positive and significant. Column (2) shows the results with added fixed effects. In Column (3), we regress on a series of company-level control variables and year-fixed and firm-fixed effects. The results suggest that when observed and unobserved variables are controlled, firms can obtain better green transformation performance after using more AI technology. The empirical findings prove that Hypothesis 1 holds true.

Columns (4)–(5) in Table 3 show the separate impacts of AI on green technical efficiency change (GTEC) and green technological change (GTC). The effect on green technical efficiency change and green technological change is significant. On one hand, AI could make an impact by optimizing production processes and improving the allocation of resources from an economic perspective. On the other hand, AI could also make an impact by updating technology and equipment.

4.2. Robustness Checks

We used alternative sample selections to confirm the robustness of our findings. The estimated coefficients, as shown in Columns (1)–(3) of

Table 4. Alternative sample selections.

VARIABLES	(1)	(2)	(3)
	GTFP	GTFP	GTFP_w
AI_1	0.0019 ***	0.0018 ***	0.0020 ***
	(0.0004)	(0.0004)	(0.0004)
Constant	0.9619 ***	1.0382 ***	1.0383 ***
	(0.0105)	(0.0079)	(0.0075)
Controls	YES	YES	YES
Time FE	YES	YES	YES
Industry FE	YES	YES	YES
Observations	8452	11,784	12,272
R-squared	0.9104	0.8987	0.8988

Note: Robust standard errors are shown in parentheses; *** denotes significance at the 1% level, respectively.

, are steadily positive and significant, indicating that our findings are still valid, excluding non-continuously operating enterprises, excluding stocks from the artificial intelligence sector, and performing a two-sided 1% winsorizing on all continuous variables. This suggests that a positive effect exists despite the selection of samples.

4.3. Endogeneity Concerns

The baseline regression used the lagged one-period term of AI to partially mitigate endogeneity, but the endogeneity of AI still requires further discussion. In order to eliminate the potential influence of unobserved effects correlated with both AI and firm GTFP, we employed the instrumental variable approach and the Heckman two-step model to deal with the potential endogeneity problem.

This research followed the approach of Li et al. [47], employing the average frequency of AI-related terms among enterprises within the same province and industry as an instrumental variable (IV). Columns (1) and (2) present the results. This IV is correlated with the endogenous explanatory variable and is strictly exogenous. An individual firm’s GTFP is one of many thousands of observations used to compute this average, so a firm’s performance is mechanically too small to influence the aggregate. Also, the province and industry average is driven by policy shocks, technology diffusion waves, and supply-side spillovers that are orthogonal to a single firm’s idiosyncratic shocks. Nevertheless, it is not expected to have a significant impact on a particular firm’s GTFP. As such, it should be an effective instrumental variable that satisfies both exclusion and relevance criteria. In the first stage of the analysis, the coefficient of the IV was significantly positive, indicating a strong correlation between the IV and the endogenous explanatory variable. The positive and significant LM statistics confirm the validity of the IV selection, while the positive Wald F statistic demonstrates that the IV makes a positive contribution to the regression equation.

To further address endogeneity issues arising from sample selection bias, we used the Heckman two-step model, the first stage of which requires a strong exogenous variable [48]. Columns (3) and (4) illustrate the application. In this context, local government attention to AI, measured by the frequency of references to AI in government reports, was selected as an exogenous variable. The adoption of AI technology is a costly and long-term strategic decision, and local government attention to AI significantly influences a firm’s propensity to adopt AI. In the first stage, the binary variable “whether a firm has adopted AI technology” (AI_01) served as the dependent variable, with the exogenous variable “local government’s attention to AI” (ai_sum_1) being introduced into the model. In the second stage, the Inverse Mills Ratio (IMR) was incorporated to correct for sample selection bias, and the impact of AI application on firms’ green total factor productivity (GTFP) was re-examined. The coefficient for “local government’s attention to AI” is significantly positive at the 1% level, indicating that listed companies in regions with greater government focus on AI are more likely to adopt AI technology. This finding aligns with theoretical expectations and validates the reasonableness of the exogenous variable selection. The significantly negative IMR coefficient in the second stage confirms the presence of sample selection bias and the effectiveness of the Heckman model in correcting for this bias. After accounting for sample selection bias, the coefficient for AI remains significantly positive. As shown in

Table 5. Endogeneity results.

VARIABLES	(1)	(2)	(3)	(4)
	AI_1	GTFP	AI_01	GTFP
AI_1		0.0182 ***		0.0022 ***
		(0.0036)		(0.0004)
IV	0.7143 ***
	(0.0643)
ai_sum_1			0.0276 ***
			(0.0104)
IMR				−0.0313 ***
				(0.0044)
Constant			−2.3889 ***	1.1131 ***
			(0.2337)	(0.0125)
Controls	YES	YES	YES	YES
Time FE	YES	YES	YES	YES
Industry FE	YES	YES	YES	YES
Kleibergen–Paap rk	9.671 [0.0019]
LM statistic
Kleibergen–Paap rk Wald F statistic	123.372
Observations	12,628	12,628	12,920	12,272
R-squared		−0.0653		0.8993

Note: Robust standard errors are shown in parentheses; *** denotes significance at the 1% level, respectively.

, all results are consistent with the baseline conclusion.

4.4. Heterogeneity Tests

4.4.1. The Effect of Different Regions

China is a vast country and there exists a certain degree of development imbalance among its eastern, central, and western regions. Factors such as the local industrial structure, development goals of each region, the economic development level, and the quality of the labor force may lead to differences in the regression results of the model [14]. Based on this, this section examines the sensitivity of the empirical results across different regional samples. The eastern part includes Beijing, Tianjin, Hebei, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, Guangxi, and Hainan, while the rest of the provinces are non-eastern. The results for the eastern and non-eastern regions are shown, respectively, in Columns (1) and (2) in

Table 6. Heterogeneity test results.

VARIABLES	Different Regions		Time Periods		High-Pollution Industries
	(1)	(2)	(3)	(4)	(5)	(6)
	GTFP	GTFP	GTFP	GTFP	GTFP	GTFP
AI_1	0.0023 ***	0.0010	−0.0001	0.0026 ***	0.0003	0.0025 ***
	(0.0004)	(0.0009)	(0.0009)	(0.0005)	(0.0015)	(0.0004)
Constant	1.0301 ***	1.0483 ***	1.1511 ***	0.9679 ***	1.0334 ***	1.0355 ***
	(0.0084)	(0.0148)	(0.0249)	(0.0108)	(0.0158)	(0.0091)
Controls	YES	YES	YES	YES	YES	YES
Time FE	YES	YES	YES	YES	YES	YES
Industry FE	YES	YES	YES	YES	YES	YES
Observations	9148	3111	4248	7850	2966	9277
R-squared	0.8989	0.8983	0.7192	0.8571	0.9021	0.8980

Note: Robust standard errors are shown in parentheses; *** denotes significance at the 1% level, respectively.

, revealing that the impact of AI application on corporate green transformation varies significantly between these regions. Specifically, the influence of AI is more pronounced in the eastern region. This finding can be attributed to the relatively more developed economy and more advanced transportation infrastructure in the eastern region, which likely facilitates the adoption and effectiveness of AI technologies in promoting green transformation among firms.

4.4.2. The Effect of Time Periods

The outbreak of COVID−19 drew urgent attention from the international community, causing health, life, and economic losses in over 200 countries and regions, including China [49]. Columns (3) and (4) in Table 6 present the empirical results for two distinct temporal segments: the period preceding 2020 and the period after it. The year 2020 marks a significant inflection point, as it coincides with the onset of the COVID−19 pandemic, an exogenous shock that exerted profound and systemic influences on both the Chinese and the global economy. The analysis reveals that the impact of AI on corporate green transformation exhibits notable variation across these different time periods. Specifically, the differential effects observed suggest that the role of AI in facilitating green transformation within firms may be contingent upon the broader economic and environmental conditions prevailing during distinct periods.

4.4.3. The Effect of High-Pollution and Low-Pollution Industries

As per the Environmental Protection Guide for Information Disclosure of Listed Companies promulgated by the Ministry of Environmental Protection of China, the subsequent 16 industries are categorized as high-pollution industries: thermal power generation, steel, cement, electrolytic aluminum, coal, metallurgy, chemical, petrochemical, building materials, paper-making, brewing, pharmaceuticals, fermentation, textiles, leather, mining, and quarrying. Liu and Zhou [1] believe that enterprises in high-pollution industries encounter higher credit barriers when adopting artificial intelligence technology and are at a disadvantage in the competition for green resources. Given this, to further investigate the impact of industry heterogeneity on corporate green transformation, this research divided the sample into high-pollution industries and low-pollution industries based on the above classification.

Columns (5) and (6) in Table 6 present the results for two distinct industrial classifications: high-pollution industries and low-pollution industries. The analysis indicates that the application of artificial intelligence (AI) exerts a significant influence on the green transformation of firms within low-pollution industries, whereas its impact is relatively muted in high-pollution industries. This differential effect can be attributed to the distinct characteristics of these industrial sectors. Compared to high-pollution industries, low-pollution industries generally possess more favorable conditions for green transformation, including greater business flexibility, lower transformation costs, and reduced transition costs and risks. These attributes facilitate a more seamless integration of AI technologies, thereby enhancing the efficacy of green transformation initiatives within low-pollution industries.

5. Possible Economic Mechanisms

Building upon the preceding analysis, this study ascertained that AI technology exerts a significant positive influence on the green transformation of Chinese enterprises. This section delves further into the underlying mechanisms through which this impact is realized, presenting empirical tests of Hypothesis 2 and Hypothesis 3. The construct of firm green innovation capabilities is bifurcated into two distinct components: firm green independent innovation capabilities and firm green collaborative innovation capabilities. Based on Equations (2) and (3), the results from

Table 7. Mechanism test results.

VARIABLES	(1)	(2)	(3)
	RDemphasis	FGIC1	FGIC2
AI_1	0.8294 ***	0.0698 ***	0.0173 ***
	(0.1102)	(0.0183)	(0.0043)
Constant	−40.7563 ***	−1.3993 ***	−0.4939 ***
	(6.3862)	(0.3256)	(0.1455)
Controls	YES	YES	YES
Time FE	YES	YES	YES
Industry FE	YES	YES	YES
Observations	12,230	12,269	12,269
R-squared	0.9524	0.7257	0.6761

Note: Robust standard errors are shown in parentheses; *** denotes significance at the 1% level, respectively.

, Column (1), report the results of the mechanism test for the emphasis on firm R&D, while Columns (2) and (3) report the results for firm green innovation capabilities. More precisely, Column (2) details the mechanism test results for firm green independent innovation capabilities, and Column (3) details the results for firm green collaborative innovation capabilities.

AI has a positive and statistically significant effect on the emphasis on firm R&D, with a coefficient of 0.8294 at the 1% significance level. This finding suggests that AI induces firms to accord greater priority to R&D activities, thereby augmenting financial support for these initiatives. Consequently, firms are able to undertake more R&D endeavors in areas such as energy conservation, emission reduction, clean energy, and waste recycling. These efforts culminate in technological advancements, reduced energy consumption, and enhanced efficiency, thereby catalyzing the green transformation of Chinese enterprises. The aforementioned results substantiate Hypothesis 2, which posits that AI fosters corporate green transformation by augmenting the emphasis on corporate R&D.

In Table 7, Column (2), the estimated coefficient that quantifies the impact of AI application on firm green independent innovation capabilities is 0.0698, passing the 1% significance test. The adoption of AI technology can improve the efficiency of R&D activities and shorten development cycles, aiding firms in overcoming technological bottlenecks and achieving green transformation. Correspondingly, in Column (3), the estimated coefficient that estimates the impact of AI on firm green collaborative innovation capabilities is 0.0173, also achieving significance at the 1% level. During collaborative R&D activities, the use of AI serves to diminish communication costs stemming from information asymmetry and to establish an efficient and stable collaborative framework. The integration of environmental and sustainability principles throughout the entire product life-cycle, encompassing design, production, and after-sales services, thereby propels the green transformation of enterprises. The results indicate that AI not only bolsters firm green independent innovation capabilities but also fortifies firm green collaborative innovation capabilities. This finding corroborates Hypothesis 3, which asserts that AI promotes corporate green transformation by reinforcing firm green innovation capabilities.

6. Moderating Effect of Marketization

A stable and mature market environment requires enterprises to improve their efficiency and competitiveness, providing an incentive for enterprises to adopt AI technology. This section evaluates how the National Economic Research Institute (NERI) index of marketization of China’s provinces influences the relationship between AI and corporate green transformation, as shown in

Table 8. Moderating effect test results.

VARIABLES	(1)	(2)	(3)	(4)	(5)	(6)
	GTFP	GTFP	GTFP	GTFP	GTFP	GTFP
AI_1	0.0020 ***	0.0019 ***	0.0018 ***	0.0021 ***	0.0019 ***	0.0020 ***
	(0.0004)	(0.0005)	(0.0006)	(0.0005)	(0.0004)	(0.0006)
market	0.0049
	(0.0029)
AI_1×market	0.0000
	(0.0003)
GMR		−0.0041 **
		(0.0015)
AI_1×GMR		0.0004
		(0.0005)
DNSOE			0.0029
			(0.0019)
AI_1×DNSOE			0.0007 **
			(0.0003)
DDPM				−0.0013 **
				(0.0006)
AI_1×DDPM				0.0003 *
				(0.0001)
DDFM					0.0035 ***
					(0.0009)
AI_1×DDFM					0.0000
					(0.0001)
DMIOLIE						0.0034 ***
						(0.0009)
AI_1×DMIOLIE						0.0001
						(0.0001)
Constant	0.9881 ***	1.0712 ***	1.0014 ***	1.0447 ***	0.9923 ***	0.9937 ***
	(0.0282)	(0.0131)	(0.0251)	(0.0086)	(0.0113)	(0.0151)
Controls	YES	YES	YES	YES	YES	YES
Time FE	YES	YES	YES	YES	YES	YES
Industry FE	YES	YES	YES	YES	YES	YES
Observations	12,219	12,219	12,219	12,219	12,219	12,219
R-squared	0.8997	0.9001	0.8993	0.8991	0.9011	0.9018

Note: acronyms used in the table are as follows: GMR (government–market relationship); development of non-state-owned economy (DNSOE); development degree of product markets (DDPM); development degree of factor markets (DDFM); development of market intermediary organizations and the legal institutional environment (DMIOLIE). Robust standard errors are shown in parentheses; *, **, and *** denote significance at the 10%, 5%, and 1% levels, respectively.

.

Columns (2)–(6) show separately the results of the five primary variables of the marketization index, which are the government–market relationship, the development of the non-state-owned economy, the development degree of product markets, the development degree of factor markets, and the development of market intermediary organizations and the legal institutional environment. The interaction coefficients in Columns (3) and (4) are significantly positive, while the other coefficients are insignificant. According to the results, the development of the non-state-owned economy and the development degree of product markets have positive effects on the influence of AI application on corporate green transformation.

This indicates that the development of the non-state-owned economy and the development degree of product markets are vital in creating a favorable external environment for the operation and development of enterprises and enabling AI technology to play a greater role in promoting the green transformation of enterprises. However, the government–market relationship, the development degree of factor markets, and the development of market intermediary organizations and the legal institutional environment have no significant impact. One possible reason is that the role of the actual economic environment does not exist in isolation but interacts with other factors, and the legal system has a lag. The other possible reason is that the Chinese government has relatively strong control over state-owned enterprises, so these indicators have strong exogeneity and play a relatively minor role in moderating the influence of AI on corporate green transformation.

7. Conclusions and Implications

This study empirically explores the impact of AI on green transformation in Chinese firms using panel data of listed companies from 2015 to 2022, which enriches the research on AI. According to the empirical results, a firm’s use of AI increases its green transformation performance. This finding is robust to different specifications and the Heckman Model. Augmenting the emphasis on corporate R&D and reinforcing firm green innovation capabilities are the two mechanisms through which AI drives corporate green transformation. This study further explores the heterogeneity effect underlying this impact, discovering that the impact of AI on corporate green transformation is more pronounced in the eastern region, after the year 2020, and in low-pollution industries. The development of the non-state-owned economy and the development degree of product markets have a moderating effect.

The significant theoretical contribution of this study is a complement to AI research on the firm-level impact of AI, offering new empirical evidence regarding the effect of AI on green transformation. Additionally, our findings hold significant importance for corporate strategic choices and the formulation of government policies. The results of the heterogeneity effects indicate that specific and targeted policies should be applied to firms in different regions and industries.

The research findings yield several policy implications: (1) To realize classified and precise policies for the use of AI, differentiated policies should be formulated based on heterogeneity. Considering the regional differences where enterprises are located, in the short term, the government should increase policy support and financial subsidies for the central and western regions. In the medium term, the construction of digital infrastructure should be advanced to narrow the digital divide among the eastern, central, and western regions. In the long term, regional connectivity should be strengthened to achieve coordinated development and complementary advantages among regions, improve the efficiency of resource allocation based on local conditions, and promote the wide application of AI. Considering the differences between high-pollution and low-pollution industries, financial subsidies for AI energy-saving renovations for enterprises in high-pollution industries should be provided. Their emission reduction volumes will be included in pollutant discharge rights trading to form market-based incentives. For enterprises in low-pollution industries, cloud-based AI diagnosis and shared models should be promoted to reduce the cost of AI implementation. (2) To enhance the green total factor productivity of enterprises and promote green and sustainable development, differentiated green development policies should be formulated based on heterogeneity. Considering the regional differences in areas where enterprises are located, in the short term, cooperation among enterprises in different regions should be strengthened to improve the efficiency of resource utilization. In the long term, the distinct resource advantages of the eastern, central, and western regions should be leveraged. The first-mover advantage of the eastern region could spur subsequent development in the central and western regions, promoting the growth of green transformation. Considering the differences between high-pollution and low-pollution industries, the policy priority is to promote the green development of enterprises in high-pollution industries, especially in carbon-intensive sectors, helping them adopt green development mindsets, coordinating governance costs with enterprise benefits, preventing structural misallocation of resources, and thus removing the worries that hold back green technological innovation. (3) To meet enterprises’ needs for AI adoption and green transformation, practical, professional, and technical talents should be cultivated. In the short term, the government should increase educational investment and optimize the allocation of educational resources. Also, the government should encourage colleges and vocational schools to set up cutting-edge disciplines and majors and introduce courses related to digital and green technologies, such as AI, big data, new energy, and ecological environmental protection, to carry out targeted talent cultivation to meet the demand of enterprises for highly skilled professionals. In the medium term, the government should guide universities and other educational institutions to strengthen relevant vocational and professional education. The government should also guide enterprises to deeply participate in the talent cultivation process. Through school–enterprise cooperation, order-based training, and jointly building training bases, students’ professional skills and practical operation abilities can be enhanced. In the long term, the government should expand channels for talent introduction, offer relevant policy support such as tax incentives and financial subsidies to enterprises, and encourage them to widely absorb digital talents worldwide. Furthermore, the government should also provide high-level service work after the introduction of new talent providing a good living and working environment so as to retain high-end talent.

However, this study has four principal limitations that require further investigation. First, this study lacks a granular decomposition of AI’s differential impacts on GTEC and GTC, necessitating future research to delineate their distinct transmission mechanisms through decomposition modeling. Because of the data limitation and the focus on the Chinese perspective, the generalizability of the findings needs further examination, potentially by employing decomposition models to delineate their distinct transmission mechanisms. Second, constrained data coverage (listed firms only) and the exclusive Chinese context limit external validity, necessitating cross-economy replication with broader samples. Third, given endogenous innovation decisions and unobserved firm heterogeneity, the inherent difficulties in establishing causality remain empirically unresolved, calling for quasi-experimental designs or instrumental variable approaches. Finally, since the GMM model is not applicable to this study, the influence of the lagged dependent variables was not considered in this paper. If the dependent variables exhibit autoregressive characteristics and the lagging dependent variable is omitted in the model, this may lead to omitted variable bias. This kind of bias will cause the estimated coefficients of other explanatory variables to be biased and inconsistent, thereby affecting the reliability of the research conclusion.

Future research should advance in two directions. In terms of data, data sources should be expanded to collect generalizable enterprise-level data, including listed and non-listed companies worldwide. Future studies may also delve into more micro-level investigations, sectoral analyses, and cross-country comparisons. In terms of models and methods, future research can focus on establishing causality, thereby more accurately and comprehensively analyzing and reflecting the relationship between AI and corporate green transformation.