The Impact of Artificial Intelligence on ESG Performance of Manufacturing Firms: The Mediating Role of Ambidextrous Green Innovation

Abstract

In the context of the worldwide quest for green and sustainable development, there is a growing importance in enhancing the environmental, social, and governance (ESG) performance of manufacturing companies. With the rise of digital transformation and pressing environmental challenges, artificial intelligence (AI) has emerged as a crucial tool for manufacturing organizations to gain a competitive edge in sustainability. While the role of digital technologies in driving ESG improvements has been widely discussed, there is limited scholarly exploration of the specific impact of AI on the ESG performance of manufacturing firms, as well as the underlying mechanisms at play from an AI perspective. Addressing this research gap, this study constructs a theoretical model of AI affecting manufacturing firms’ ESG performance using a sample of Chinese-listed manufacturing firms from 2012–2022. Additionally, this study examines the role of mediating mechanisms of ambidextrous green innovation as well as differences in the intrinsic mechanisms triggered by the equilibrium of ambidextrous green innovation and firm size. The findings indicate that the application of AI technology effectively promotes improvements in the ESG performance of manufacturing firms, with ambidextrous green innovation playing a positive mediating role. Furthermore, manufacturing companies with a strong balance of ambidextrous green innovation and larger scale exhibit enhanced effects of AI on ESG performance. This study serves to supplement existing literature on ESG performance enhancement, elucidate the theoretical rationale behind the non-economic performance of AI-enabled firms, and extend the application of organizational dualism theory to new contexts.

1. Introduction

With a growing global emphasis on green and sustainable development, an increasing number of countries and enterprises are focusing on enhancing their environmental, social, and governance (ESG) performance [1,2]. In particular, manufacturing firms play a crucial role in the global economy, serving not only as primary industrial production entities but also as comprehensive organizations integrating technological innovation, market responsiveness, and social responsibility [3]. In today’s economic and environmental context, these firms must enhance operational efficiency while also prioritizing environmental protection, social responsibility, and corporate governance improvements [4]. However, during this process, enterprises encounter a series of technical barriers that hinder the improvement of their ESG performance. For instance, achieving higher energy efficiency, reducing waste emissions, and implementing circular economy practices often require advanced technological support [5]. Additionally, the complexity of data collection and analysis poses challenges for enterprises when measuring their ESG performance [6]. These issues not only increase operational costs but also limit progress toward sustainability.

The development of artificial intelligence (AI) technologies provides new avenues to overcome these technical barriers [7]. AI, as a key driver of growth, is becoming a critical competitive advantage and resource for companies of all sizes [8]. It profoundly influences business operations and organizational processes. Companies can leverage AI to play a significant role in innovation, facilitating open collaboration and precise innovation [8], as well as to shape dynamic capabilities and enhance resource efficiency [9]. AI can also assist in better monitoring and reporting of ESG metrics, ensuring accuracy and transparency. Thus, AI plays a crucial role in enhancing ESG performance. Currently, AI has entered industrial production, reshaping traditional industry models and leading future value creation, making significant contributions to the global economy and social activities. The United Nations, the European Union (EU), and the World Economic Forum have explicitly stated that the use of AI should contribute to achieving sustainable development goals [10]. Consequently, AI, with its high level of intelligence and digitalization, can help companies overcome issues related to low energy efficiency, weak innovation foundations, and information transmission bottlenecks [11], providing a supportive environment and necessary elements for improving ESG performance in manufacturing enterprises.

Currently, most research focuses on the impact of AI on corporate financial performance [12] and production efficiency [13], with fewer studies examining how AI affects ESG performance, particularly in the manufacturing sector. Manufacturing firms face numerous environmental and social challenges, and AI holds significant potential in addressing these issues, but existing research has not fully explored the specific mechanisms and effects in this domain. In response to the call by scholars like Lv et al. (2023) [14], this paper aims to accelerate the deep integration of AI with industrial manufacturing and fully leverage the low-pollution and clean nature of intelligent technologies to continuously stimulate the environmental optimization effects of AI, assisting manufacturing enterprises in achieving green emission reduction goals. Existing studies often directly examine the relationship between AI and ESG performance, neglecting potential mediating and moderating variables. For instance, what specific pathways does AI use to influence ESG performance? Do different firm characteristics affect this relationship differently? These questions remain unanswered. On this basis, this paper further explores whether dual green innovation can mediate the relationship between AI and ESG performance. Dual green innovation, as a sustainable dual-wheel driver of innovation, can enhance intrinsic motivation in enterprises, resolve development contradictions, and solve interest conflicts to achieve balanced development. This element is significant for enterprise ESG performance and sustainable development. However, existing research overlooks the role of dual green innovation in the relationship between AI and manufacturing enterprise ESG performance and does not adequately distinguish between exploratory and exploitative green innovations, leading to an insufficient understanding of specific green innovation behaviors. Furthermore, current research predominantly uses cross-sectional data or single-case studies, lacking large-sample, long-time-series empirical analyses. This limitation restricts the external validity and generalizability of the findings. Additionally, existing studies have limitations in data sources and measurement methods, making it difficult to comprehensively and accurately reflect the complex relationship between AI and ESG performance.

Building upon the preceding analysis, this study employs a dataset encompassing 1068 A-share manufacturing firms listed on the Shanghai and Shenzhen stock exchanges in China between 2012 and 2022 to conduct an in-depth examination of the mechanisms through which AI influences the ESG performance of manufacturing enterprises, alongside an assessment of the mediating role played by ambidextrous green innovation. In comparison with the existing literature, this study makes marginal contributions with the following: (1) The impact of AI on the ESG performance of manufacturers in the context of green and sustainable development is clarified, effectively expanding and complementing the antecedent research on enhancing ESG performance of manufacturing enterprises, which helps further clarify the theoretical logic of how AI empowers non-economic performance of enterprises. (2) The complete pathway of “AI, ambidextrous green innovation, ESG performance of manufacturing enterprises” is successfully validated, thereby elucidating the mechanism through which ESG performance is enhanced. This contributes to unraveling the black box surrounding the mechanisms through which AI impacts corporate ESG performance, thereby furnishing enterprises with actionable strategies to bolster their competitive edge in sustainable development. (3) The mediating roles of exploratory and exploitative green innovation are examined, which to some extent strengthens the knowledge and understanding of the differentiation of corporate ambidextrous green innovation, enriches relevant research on the subject, and expands the practical domains in which organizational dual theory finds application. (4) Heterogeneity analysis, conducted separately in terms of ambidextrous green innovation balance capability and firm size, provides a nuanced understanding of the internal challenges faced by enterprises, clarifies certain prerequisites upon which the ESG enhancement mechanisms rely, and broadens the analytical lens through which heterogeneity is studied in this context.

2. Theoretical Analyses and Hypothesis Formulation

2.1. AI and ESG Performance of Manufacturing Firms

Against the backdrop of a global impetus towards achieving high-quality economic growth and green sustainable development, enterprises must not only prioritize financial returns but also take the initiative to take into account environmental protection, social responsibility, corporate governance, and the active fulfillment of ESG responsibility, as these constitute pivotal determinants of long-term resilience and competitive edge. AI, a digital technology that mimics human cognitive thinking and behavior, processes and analyzes data, learns and improves performance, understands and generates natural language, and performs reasoning and decision-making tasks [15], provides a powerful tool for enhancing ESG performance in enterprises. It also guides organizations in developing new knowledge, improving decision-making processes, and better meeting consumer needs [8].

Firstly, the application of AI technology by manufacturing enterprises has significantly enhanced their capability to excavate, process, and analyze unstructured and non-standardized data, thereby facilitating smoother information flow both within and outside the enterprise [16]. Subsequently, with the acceleration in the speed of information exchange and product development, business processes have been streamlined, and internal organizational management has been upgraded, enabling the effective implementation of efficient and refined management within the organization. AI plays a pivotal role in enabling companies to make decisions more rapidly and accurately, thereby steadily elevating corporate governance standards.

Secondly, if enterprises want to achieve long-term development, they must avoid short-sighted behaviors that damage the environment and long-term social benefits. In terms of environmental protection, AI is capable of conducting extensive analyses of data about production processes and environmental conditions, thereby empowering enterprises to formulate and execute more adaptive climate interventions and environmental protection strategies more objectively and rationally. This analytical capacity facilitates the identification of climate risks, the enhancement of resource efficiency, and the diminution of carbon emissions [17]. Simultaneously, AI strengthens cross-organizational collaboration between enterprises, reduces the application threshold of environmental protection technologies such as environmental governance and energy conservation, and accelerates the green transformation of enterprises, enhancing their environmental performance.

Finally, social responsibility, as an intangible asset, plays a decisive role in whether manufacturing enterprises can achieve long-term development. AI can deeply interpret and mine information, helping enterprises quickly identify and capture the value demands of stakeholders including shareholders, employees, consumers, creditors, upstream and downstream enterprises, and the general public, speeding up the response to market demand, breaking information barriers, and enhancing the efficiency of external resource allocation. Concurrently, under the watchful gaze and scrutiny of diverse stakeholders, enterprises become increasingly attentive to their public image and reputation and are conducive to assuming a broader spectrum of social responsibilities.

In summary, manufacturing firms can harness the power of AI to steer their decision-making towards environmentally responsible practices, social accountability, and sound corporate governance, ultimately enhancing their ESG performance, giving rise to Hypothesis 1.

H1: 

AI contributes to the ESG performance of manufacturing enterprises.

2.2. The Mediating Role of Ambidextrous Green Innovation

With the rapid development of the ESG connotation, ambidextrous green innovation has become a new norm for industry development. Ambidextrous green innovation, while pursuing corporate development and economic growth, focuses more on improving natural resource utilization and reducing environmental pollution emissions through technological, product, and process innovations, etc. [18]. Ambidextrous green innovation is a crucial manifestation of ESG and is essential for enhancing corporate competitiveness and creating a sustainable circular economy pattern [19]. However, the unique knowledge spillover and dual external attributes of environmental protection of ambidextrous green innovation [20], combined with the inherent long-term and high uncertainty of traditional innovation [21], determine the difficulty for manufacturing enterprises to implement ambidextrous green innovation. At this point, AI can leverage its characteristics such as digitization, intelligence, openness, and sharing to provide new development opportunities for enterprises, promote ambidextrous green innovation in manufacturing enterprises, and enhance their ESG performance. Ambidextrous green innovation in enterprises can be divided into exploratory green innovation and exploitative green innovation [22], clarifying the differences and connections between different types of green innovation driving paths, and helping enterprises carry out targeted ambidextrous green innovation activities based on their conditions [23].

Exploratory green innovation refers to inventive activity that disrupts and reconstructs the prevailing green technological paradigm, manifesting as the subversion and reconfiguration of extant green knowledge and technologies, often accompanied by the advent of novel green products and processes. The deployment of AI within manufacturing enterprises exhibits characteristics of self-learning, adaptability, and autonomous action in response to external environmental shifts, thereby catalyzing the pursuit of exploratory green innovation and the subsequent reconfiguration of green manufacturing workflows [24]. This enables the optimization of existing product portfolios, precise allocation of production factors, reduction in material waste, energy consumption, and pollution discharge, as well as the curtailment of production costs [25], thereby operationalizing ESG principles. Moreover, with the public’s growing environmental consciousness, consumers increasingly favor eco-friendly green products [26]. AI technologies aid businesses in accurately pinpointing consumer demand for green offerings, spurring them towards exploratory green innovation, enhancing their product market returns, and providing retroactive incentives for green innovation endeavors. The actualization of exploratory green innovation facilitates the development of novel, difficult-to-emulate products, and distinctive operating models, progressively engendering a portfolio of differentiated green products that confer enduring competitive advantages [27], ultimately aligning to enhance corporate ESG performance.

Exploitative green innovation denotes inventive activity that adheres to the established green technological trajectory, manifesting as enhancements and refinements to extant green knowledge and technologies, frequently accompanied by upgrades to green products and services and reductions in green development costs. Manufacturing enterprises implementing exploitative green innovation can leverage the advantages of artificial intelligence technology to enhance connectivity between manufacturing enterprises, research institutes, and government agencies, effectively promote information sharing, knowledge recombination, and resource complementarity among different entities, accelerate the dissemination and aggregation of innovative elements, and increase the reserve of green innovation knowledge in enterprises [20]. Additionally, exploitative green innovation continuously stimulates the vitality of data and information elements, integrates and coordinates internal resources, improves existing product quality, efficiency, and production efficiency, reduces operational costs [28], enhances the contribution of enterprises to the environment and sustainable development, and strengthens the ESG practice capabilities of enterprises.

In conclusion, whether it is exploratory green innovation or exploitative green innovation, the empowerment of AI can help enterprises establish a good green image. While maintaining competitive advantages, it can achieve green sustainable development; ultimately, this fosters an elevation in the ESG performance of manufacturing enterprises. Therefore, ambidextrous green innovation serves as an intermediary mechanism between AI and the ESG performance of firms, giving rise to Hypothesis 2.

H2a: 

Exploratory green innovation exerts a positive mediating influence on the relationship between AI and the ESG performance of manufacturing enterprises.

H2b: 

Exploitative green innovation exerts a positive mediating influence on the relationship between AI and the ESG performance of manufacturing enterprises.

The theoretical model is illustrated in Figure 1.

3. Data and Methods

3.1. Data Sources

To ensure the credibility and validity of this study, a multi-channel, multi-source data collection method was employed. The specific data collection process is as follows: The study sample includes 1068 A-share manufacturing companies listed on the Shanghai and Shenzhen stock exchanges from 2012 to 2022. Companies with abnormal financial conditions and severe data deficiencies were excluded to ensure sample representativeness and data completeness.

Among these, ESG performance data were obtained from the Huazheng ESG ratings and scores in the Wind database. Wind, a leading financial data provider in China, offers extensive, authoritative, high-quality, and timely data, providing a reliable foundation for this study. Green invention and utility patent data were sourced from the green patent database and listed company patent database disclosed by the China Research Data Service Platform (CNRDS). CNRDS is an authoritative academic data platform that ensures the accuracy and reliability of patent data. Additional financial and market data were obtained from the Choice Financial Terminal, a widely used platform for financial research and investment analysis, known for its comprehensive and timely financial data, which further enhances the reliability and diversity of the research data.

After collecting all data, the following steps were taken: First, strict cleaning and preprocessing were conducted to remove duplicate data and outliers, ensuring data accuracy and consistency. Second, data from different sources were integrated into a unified research database to facilitate subsequent analysis and modeling. Third, cross-validation and comparative analysis were performed to ensure data consistency and reliability, minimizing the impact of data bias on the research results. Through these detailed data collection and processing steps, this study ensures high-quality and reliable data, providing a solid foundation for exploring the relationship between artificial intelligence and the ESG performance of manufacturing enterprises.

3.2. Regression Model

To examine the relationship between artificial intelligence and the ESG performance of manufacturing enterprises, we refer to the work of scholars such as Li et al. (2024) [17] and formulate Model (1) based on Hypothesis 1:

$$ES{G}_{it}={\alpha }_0+{\alpha }_{1}A{I}_{it}+{\alpha }_{i}Control{s}_{it}+\Sigma Compan{y}_{it}+\Sigma Yea{r}_{it}+{\epsilon }_{it}$$

(1)

Among them, i denotes firm, t denotes year, $ES{G}_{it}$ is the ESG performance score of firm i in period t, $A{I}_{\mathrm{it}}$ is the degree of AI of firm i in period t, $Contro{l}_{it}$ denotes the control variable, $Compan{y}_{it}$, $Yea{r}_{it}$ represent firm and year, respectively, and ${\epsilon }_{it}$ is the random error term.

3.3. Variable Description

3.3.1. Explained Variable: ESG Performance (ESG)

The Huazheng ESG Rating System is a specialized framework focusing on ESG metrics. It employs scientific evaluation methods, offers broad coverage, and maintains high data quality and timeliness. Compared to other domestic ESG rating systems, the Huazheng system provides detailed and accurate ESG ratings and data, with greater breadth and depth. Furthermore, its ratings and data are highly recognized and authoritative in the industry. This study uses the 2012–2022 Huazheng ESG rating data, dividing the ratings from low to high into 9 levels, scoring companies from 1 to 9 based on their annual ESG rating level [29]. The higher the number, the better the ESG performance of manufacturing enterprises.

3.3.2. Explanatory Variable: Artificial Intelligence (AI)

At the micro-level, existing literature mainly uses patents [30] or surveys [31] to measure AI, without building comprehensive indicators to reflect the micro-level AI level of enterprises. For listed companies, their annual reports disclose the company’s main business information, operating conditions, and management’s judgment on future development direction [32], providing important reference value for understanding the company’s business strategy and decision-making. Therefore, this study refers to the AI dictionary summarized and compiled by Yao et al. (2024) [33], including 73 keywords related to AI (Appendix A 

Table A1. Artificial intelligence dictionary.

Artificial Intelligence	AI Products	AI Chips	Machine Translation	Machine Learning
Computer Vision	Human-Computer Interaction	Deep Learning	Neural Networks	Biometrics
Image Recognition	Data Mining	Feature Recognition	Speech Synthesis	Voice Recognition
Knowledge Graph	Intelligent Banking	Intelligent Insurance	Human–Machine Collaboration	Intelligent Supervision
Intelligent Education	Intelligent Customer Service	Smart Retail	Intelligent Agriculture	Intelligent Investment
Augmented Reality	Virtual Reality	Intelligent Healthcare	Smart Speaker	Intelligent Voice
Intelligent Government	Automated Driving	Smart Transport	Convolutional Neural Network	Voice Recognition
Feature Extraction	Driverless	Smart Home	Question and Answer System	Face Recognition
Business Intelligence	Intelligent Finance	Recurrent Neural Network	Reinforcement Learning	Intelligent Body
Intelligent Elderly Care	Big Data Marketing	Big Data Risk Control	Big Data Analytics	Big Data Processing
Support Vector Machine	Long Short-Term Memory	Robotic Process Automation	Natural Language Processing	Distributed Computing
Knowledge Representation	Smart Chip	Wearables	Big Data Management	Smart Sensors
Pattern Recognition	Edge Computing	Big Data Platforms	Intelligent Computing	Smart Search
Internet of Things	Cloud Computing	Augmented Intelligence	Voice Interaction	Intelligent Environmental Protection
Human–computer Dialogue	Deep Neural Networks	Big Data Operations

). Through text analysis, the annual report data of listed companies was analyzed to sum up the occurrences of the vocabulary included in the AI dictionary in the annual report of the company, adding 1 and taking the natural logarithm to measure the extent of AI application in that manufacturing company.

3.3.3. Mediating Variable

Ambidextrous green innovation encompasses exploratory green innovation (GreInva) and exploitative green innovation (GreUma). Given the relative comprehensiveness of green patent data, which allows for accurate reflection and quantification of green innovation efforts, and considering the substantial time and financial investments entailed in the progression from patent application submission to approval, it is posited that the number of green patent applications offers a more stable and authentic representation of a firm’s green innovation prowess compared to alternative metrics [34]. Thus, this study employs the count of green invention patent applications as a proxy for exploratory green innovation, while the count of green utility model patent applications serves as a measure of exploitative green innovation. The raw patent counts are subjected to data preprocessing by adding 1 to each value and subsequently computing the natural logarithm.

3.3.4. Control Variables

In recognition of the influence exerted by various other factors intrinsic to manufacturing enterprises on their ESG performance, and to ensure the robustness of the study outcomes, the following control variables were selected: enterprise ownership nature, equity concentration degree, and board size that significantly affect the company’s strategic decision-making and response speed; enterprise size and total asset return ratio reflecting the company’s R&D investment intensity and asset operation efficiency; the uncertainty of cash flow that inhibits enterprises from introducing artificial intelligence technology, developing environmental protection technology, and undertaking social responsibility.

Specific variable measurement methods are shown in

Table 1. Definition and description of main variables.

Variable	Definition	Symbol	Calculation Method
Explained variable	ESG Performance	ESG	Huazheng ESG Rating Index Score
Explanatory variable	Artificial Intelligence	AI	ln(total number of AI word frequencies + 1)
Mediating variable	Exploratory Green Innovation	GreInva	ln(Green Invention Patent Applications + 1)
	Leveraging Green Innovation	GreUma	ln(Green Utility Model Patent Applications + 1)
Control variables	Firm Size	Size	ln(Total Assets + 1)
	Corporate Ownership	Owner	National Ownership: 1 if state-owned, 0 otherwise
	Shareholding Concentration	Int10	Cumulative Shareholding Percentage of Top 10 Shareholders
	Board Size	Boarsize	Number of Board of Directors
	Total Return on Assets	Roa	Net Profit/Total Assets at Period End
	Cash Flow Level	Cash	Operating Cash Flow/Operating Revenue

.

4. Empirical Analysis

4.1. Descriptive Statistics

Descriptive statistics of the main variables are shown in

Table 2. Descriptive statistics of the sample.

Variable	Observations	Mean	Std. Dev.	Min	Max
ESG	11,748	4.010	1.050	1.000	7.250
AI	11,748	0.740	1.100	0.000	5.570
Size	11,748	3.564	1.184	0.000	8.505
Owner	11,748	0.275	0.447	0.000	1.000
Int10	11,748	56.860	15.340	0.000	100.000
Boardsize	11,748	8.397	1.522	0.000	18.000
Roa	11,748	6.647	19.530	−888.810	1206.390
Cash	11,748	0.905	5.140	−117.310	105.180

, with a mean ESG score of 4.010, a minimum of 1.000, and a maximum of 7.250, implying a need for heightened ESG commitment among the listed manufacturing enterprises. The mean of AI is 0.740, with a minimum of 0.000 and a maximum of 5.570, indicating differences in the application of artificial intelligence technology among different manufacturing companies.

4.2. Correlation Analysis

The results of the correlation analysis for the main variables are displayed in

Table 3. Correlation analysis results.

	ESG	AI	Size	Owner	Int10	Boardsize	Roa	Cash
ESG	1.000
AI	0.068 ***	1.000
Size	0.166 ***	0.115 ***	1.000
Owner	0.033 **	−0.041 ***	0.264 ***	1.000
Int10	0.130 ***	−0.090 ***	−0.022 **	−0.071 ***	1.000
Boardsize	0.028 **	−0.078 ***	0.165 ***	0.179 ***	0.008	1.000
Roa	0.076 ***	−0.026 **	−0.040 ***	−0.038 ***	0.090 ***	0.019 **	1.000
Cash	0.064 ***	0.030 **	0.105 ***	−0.010	−0.043 ***	0.007	0.047 ***	1.000

*** p < 0.01, ** p < 0.05, * p < 0.1.

. The coefficient between ESG and AI is 0.068, significantly positively correlated at the 1% level, suggesting that the higher the level of AI in manufacturing companies, the more it can promote their ESG performance, preliminarily confirming the theoretical analysis and Hypothesis 1 of this study. All selected control variables have a significant relationship with ESG, indicating their appropriateness. Further examination utilizing the Variance Inflation Factor (VIF) method reveals that the VIF values for all control variables are below 5, indicating the absence of severe multicollinearity concerns in the baseline regression model.

4.3. Baseline Regression Results

Table 4. Baseline regression results.

	(1)	(2)	(3)	(4)
	ESG	ESG	ESG	ESG
AI	0.065 ***	0.0603 ***	0.075 ***	0.021 *
	(0.009)	(0.009)	(0.009)	(0.012)
Size		0.136 ***	0.152 ***	0.216 ***
		(0.009)	(0.010)	(0.018)
Owner		0.0165	0.003	0.0326
		(0.023)	(0.023)	(0.724)
Int10		0.009 ***	0.009 ***	0.006 ***
		(0.001)	(0.001)	(0.001)
Boardsize		0.003	−0.001	−0.015 *
		(0.007)	(0.007)	(0.008)
Roa		0.003 ***	0.003 ***	0.0003
		(0.001)	(0.001)	(0.0004)
Cash		0.010 ***	0.006 ***	0.001
		(0.002)	(0.002)	(0.002)
Constant	3.964 ***	2.902 ***	3.109 ***	2.496 ***
	(0.012)	(0.072)	(0.081)	(0.699)
Company	NO	NO	NO	YES
Year	NO	NO	YES	YES
Observations	11,748	11,748	11,748	11,748
R-squared	0.005	0.052	0.062	0.609

Standard errors in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1.

presents the empirical results examining the relationship between AI and the ESG performance of manufacturing enterprises. For an accurate assessment of AI’s impact, the model incorporates multiple influencing factors: Column (1) includes only the dependent and independent variables, with an R-squared of 0.005, indicating the limited explanatory power of AI on ESG performance. Column (2) introduces control variables, increasing the R-squared to 0.052, suggesting a modest improvement in the model’s explanatory capacity. Column (3) adds year effects, raising the R-squared to 0.062, reflecting a slight but significant enhancement in explanatory power. Column (4) further incorporates firm-specific and year-fixed effects, significantly boosting the R-squared to 0.609. This underscores the crucial role of these fixed effects in enhancing the model’s explanatory power and interpreting ESG performance. Regression results consistently demonstrate a significant positive correlation between AI and ESG performance, regardless of the inclusion of control variables or fixed effects, thereby validating Hypothesis H1.

In summary, regardless of whether control variables are included or whether firm-specific and year-specific effects are controlled for, AI and ESG remain significantly positively correlated. This indicates that AI positively influences the ESG performance of manufacturing firms, thereby validating Hypothesis H1. This study uses text analysis to measure AI, where term frequency reflects the importance and application level of AI technology in manufacturing firms, including aspects such as R&D investment, technology acquisition, and product innovation. The Huazheng ESG evaluation system is employed to assess the environmental, social, and governance performance of firms. This system considers multiple dimensions to provide a comprehensive and objective reflection of a firm’s sustainability capabilities. The findings suggest that as firms increase their investment and application of AI technology, their performance in environmental, social, and governance areas also improves, thereby promoting sustainable development. This insight is crucial for guiding manufacturing firms in strategically deploying AI technology to achieve a win–win scenario of economic and social benefits.

4.4. Robustness Test

4.4.1. Instrumental Variable Method

The baseline regression results suggest that manufacturing enterprises with higher levels of AI tend to promote enhanced ESG performance. However, this outcome may also stem from the possibility that firms with deeper ESG convictions are more inclined to proactively upgrade their digital capabilities, thereby introducing a bidirectional causal endogeneity issue. To mitigate this potential endogeneity problem, we employ lagged AI as an instrumental variable in a two-stage least squares (2SLS) regression. In the first stage of the 2SLS approach, the manufacturing firm’s AI is the dependent variable, with lagged AI serving as the independent variable, along with the control variables from Model (1) and firm and year fixed effects, subject to ordinary least squares regression. In the second stage, ESG performance assumes the role of the dependent variable, with AI as the explanatory variable. The results are summarized in

Table 5. Regression results of the instrumental variable.

	(1) First Stage	(2) Second Stage
	AI	ESG
AI		0.063 ***
		(0.011)
LagAI	0.905 ***
	(0.005)
Size	0.013 ***	0.145 ***
	(0.005)	(0.009)
Owner	−0.028 **	0.014
	(0.012)	(0.023)
Int10	0.0004	0.008 ***
	(0.000)	(0.001)
Boardsize	−0.006	0.003
	(0.004)	(0.007)
Roa	0.0003	0.003 ***
	(0.000)	(0.000)
Cash	−0.002 *	0.010 ***
	(0.001)	(0.002)
Constant	0.138 ***	2.894 ***
	(0.038)	(0.073)
Observations	11,748	11,748
R-squared	0.759	0.053

Standard errors in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1.

. In the first-stage regression outputs, the coefficient attributed to LagAI exceeded zero at the 1% significance threshold, indicating a strong correlation between the instrumental variable and the endogenous explanatory variable. Furthermore, the outcomes from both the weak instrument test and the overidentification test exhibited a p-value of 0.000, decisively refuting the null hypothesis, thereby negating the presence of a weak instrument. This evidence supports the efficacy of the instrumental variable selection in this study, concurrently substantiating the robustness of the findings derived from the baseline regression analysis.

4.4.2. Propensity Score Matching Method

Due to the possibility that companies with high levels of AI may have better basic conditions, to further mitigate potential sample selection issues, propensity score matching (PSM) is employed to further address endogeneity concerns. Based on the average number of AI terms in manufacturing companies being 5, the sample is divided into a high AI group and a low AI score group. Subsequently, firm size, firm ownership, equity concentration, board size, return on total assets, and cash flow levels are selected as covariates. A 1:1 nearest neighbor matching is conducted between the treatment and control groups, and the model is re-estimated to validate the regression. The post-matching regression results, as per Model (1) in

Table 6. PSM regression results.

Variable	Sample	Treatment Group	Control Group	Difference	Standard Deviation	T-Statistic
ESG	Unmatched	4.114	3.966	0.148	0.022	6.800
	ATT	4.114	3.989	0.125	0.031	4.080
	ATU	3.967	4.049	0.082
	ATE			0.096

, reveal that the ATT’s t-statistic exceeds 2.58, signifying statistical significance at the 1% level. This indicates that, even after partially mitigating sample selection bias, the level of AI in manufacturing companies still significantly affects their ESG performance, further ensuring the robustness of the conclusions.

4.4.3. Effectiveness Analysis of DID Estimation

To further validate the credibility of the empirical results, this study supplements the difference-in-differences (DID) method to examine the impact of AI on ESG performance in manufacturing companies and conducts tests to assess the effectiveness of the DID estimation through parallel trend tests and placebo tests. Here, the deployment in 2017 by the Chinese State Council of the “New Generation Artificial Intelligence Development Plan” (hereinafter referred to as the “Plan”) is used as a quasi-natural experiment. This document is China’s first systematic deployment in the field of AI, guiding the creation of China’s advantage in intelligent development in the future. Specifically, the DID model constructed in this study is as follows:

$$ES{G}_{it}={\delta }_0+{\delta }_{1}Trea{t}_i\times Pos{t}_t+rControl{s}_{it}+{\mu }_t+{\epsilon }_{it}$$

(2)

where $ES{G}_{it}$ represents the ESG performance of manufacturing firm i in year t. $Trea{t}_i$ is a dummy variable. The “Plan” mentions that existing funds, bases, and other stock resources should be fully utilized, the coordinated allocation of domestic and international innovative assets, the incentivization of increased investment from enterprises and society, and the encouragement of leading enterprises to support AI development. Therefore, manufacturing companies with a basic foundation in AI technology are classified as the treatment group, with $Trea{t}_i$ defined as 1; firms are designated as the control group, with $Trea{t}_i$ defined as 0. $Pos{t}_t$ is a time dummy variable. Given that the “Plan” was introduced in 2017, when t is greater than or equal to 2017, it equals 1; when t is less than 2017, it equals 0. $Trea{t}_i\times Pos{t}_t$ are the variables of main concern in this paper, which represent the interaction term between the processing group and time virtual variables of the virtual group after the planning. $Control{s}_{it}$ is a combination of control variables incorporating a range of firm-level factors that vary over time. ${\mu }_t$ represents year-fixed effects, while ${\epsilon }_{it}$ stands for the error term. The regression coefficient associated with the core variable $Trea{t}_i\times Pos{t}_t$ is statistically significant at the 10% level, affirmatively indicating that the implementation of the initiatives outlined in the Plan has led to an improvement in the ESG performance of Chinese manufacturing enterprises.

Parallel Trend Test: Event Study Method

The fundamental prerequisite for a valid DID estimation is the requirement of parallel trends between the treatment and control groups before the policy intervention. Following the approach adopted by scholars such as Lv et al. (2019) [35], this study employs an event study methodology to meticulously investigate the trend evolution of both groups. To mitigate multicollinearity concerns, the period immediately preceding the policy implementation is excluded from the analysis. The regression results, depicted in Figure 2, reveal that before 2017, non-significant outcomes prevailed, indicating the absence of any substantial disparity between the treatment and control groups. Notably, a significant divergence in ESG performance emerges in the treatment and control groups coinciding with and after the policy onset. Although there was a slight decline during the COVID-19 pandemic from 2019 to 2021, which inflicted severe shocks on the economy, China’s economy rebounded resiliently in 2022, overcoming internal difficulties, with the economy recovering and industries advancing solidly towards high-quality development. This suggests that there is a clear upward trend in the dynamic effect of the policy’s implementation in promoting AI-driven improvements in the ESG performance of manufacturing firms, satisfying the parallel trend hypothesis.

Placebo Test

To address potential endogeneity issues, this study set the sample period from 2012 to 2017, randomly set the year of the initiative to 2015, and then generated new treatment and control groups for a placebo test. To improve the identification capability of the placebo test, 500 regression runs were repeated, with the results shown in Figure 3. From Figure 3, it can be observed that most estimates in the “pseudo-treatment group” are concentrated in the range of −0.05 to 0.05, while the true estimate is 0.077. The majority of p-values exceed 0.1, indicating non-significance at the 10% level. This signifies that our true estimate is an outlier and cannot produce significant estimation results by random simulation. This further substantiates the notion that the policy implementation has exerted a pronouncedly facilitative effect on the ability of AI to enhance the ESG performance of manufacturing firms, enhancing the reliability of this study’s conclusions.

5. Mechanism Testing

Based on the analysis above, AI primarily enhances the ESG performance of manufacturing companies by promoting breakthrough innovations in green technologies and upgrading green products, among other methods. To this end, this section will analyze and examine the above impact mechanisms from an empirical perspective to reveal the logical chain behind intelligence and greening in a more comprehensive manner.

5.1. Mediation Effect Model

To investigate the mediating role of ambidextrous green innovation in the relationship between AI and the ESG performance of manufacturing firms, Models (3) and (4) are formulated. In Model (3), $GreInv{a}_{it}$ represents exploratory green innovation for firm i in period t, operationalized as the count of green invention patent applications. Similarly, in Model (4), $GreUm{a}_{it}$ signifies exploitative green innovation for firm i in period t, defined as the number of green utility model patent applications.

$$ES{G}_{it}={\beta }_0+{\beta }_{1}A{I}_{it}+{\beta }_{2}GreInv{a}_{\mathrm{it}}+\Sigma {\beta }_{m}Control{s}_{it}+\Sigma Year+{\epsilon }_{\mathrm{it}}$$

(3)

$$ES{G}_{it}={\chi }_0+{\chi }_{1}A{I}_{it}+{\chi }_{2}GreUm{a}_{\mathrm{it}}+\Sigma {\chi }_{m}Control{s}_{it}+\Sigma Year+{\epsilon }_{it}$$

(4)

5.2. Analysis of Mediation Effect Regression Results

Table 7. Results of the mediation effect test.

	(1)	(2)	(3)	(4)
	GreInva	ESG	GreUma	ESG
AI	0.137 ***	0.036 ***	0.087 ***	0.047 ***
	(0.006)	(0.009)	(0.005)	(0.009)
GreInva		0.171 ***
		(0.015)
GreUma				0.153 ***
				(0.016)
Size	0.198 ***	0.102 ***	0.158 ***	0.112 ***
	(0.006)	(0.010)	(0.006)	(0.010)
Owner	0.018	0.013	−0.012	0.019
	(0.015)	(0.023)	(0.014)	(0.023)
Int10	−0.0004	0.009 ***	9.38 × 10−5	0.009 ***
	(0.001)	(0.001)	(0.0004)	(0.001)
Boardsize	0.016 ***	0.0003	0.007 *	0.002
	(0.004)	(0.007)	(0.004)	(0.007)
Roa	0.0001	0.003 ***	4.24 × 10−6	0.003 ***
	(0.0003)	(0.001)	(0.0003)	(0.001)
Cash	−0.006 ***	0.011 ***	−0.005 ***	0.011 ***
	(0.001)	(0.002)	(0.001)	(0.002)
Constant	−0.633 ***	3.010 ***	−0.428 ***	2.968 ***
	(0.047)	(0.072)	(0.043)	(0.072)
Observations	11,748	11,748	11,748	11,748
R-squared	0.151	0.065	0.106	0.060

Standard errors in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1.

Columns (1) and (2) reflect the influencing mechanism of exploratory green innovation as a mediating variable between the main effects. Column (1) shows that AI has a positive and significant effect on exploratory green innovation, with an R-squared of 0.151. In Column (2), exploratory green innovation is reintroduced into Model (1) while keeping other variables constant, resulting in an R-squared of 0.065. Although this R-squared is relatively low, it indicates that exploratory green innovation has a certain influence on ESG performance. The regression results indicate that exploratory green innovation significantly promotes the improvement of the ESG performance of manufacturing companies at the 1% level, forming a positive pathway of “AI, exploratory green innovation, manufacturing company ESG performance”. Combining the above analysis, hypothesis H2a is confirmed.

Table 7 Columns (3) and (4) reflect the influencing mechanism of exploitative green innovation as a mediating variable between the main effects. Column (3) shows that AI enhances exploitative green innovation, with an R-squared of 0.106. In Column (4), exploitative green innovation is added to Model (1) while keeping other variables constant, resulting in an R-squared of 0.060. This indicates that exploitative green innovation serves as a mediator with some explanatory power in the model. The regression results demonstrate that exploitative green innovation can drive manufacturing companies to enhance their ESG performance at the 1% level, forming a positive pathway of “AI, exploitative green innovation, manufacturing company ESG performance”, confirming hypothesis H2b.

This study uses the number of green invention patents applied for by manufacturing firms as a measure of exploratory green innovation. Invention patents typically involve novel technical solutions, reflecting the firm’s frontier exploration and innovation capabilities in green technology. For exploitative green innovation, this study employs the number of green utility model patents applied for, which generally involve improvements to existing technologies, reflecting the firm’s practical application and enhancement capabilities in green technology. The regression results indicate that both exploratory and exploitative green innovations mediate the positive relationship between AI and ESG performance. This suggests that the application of AI technology can promote ambidextrous green innovation in manufacturing firms, leading to increased publication of green patents and improved green technology levels. Through ambidextrous green innovation, firms can achieve better performance in environmental, social, and governance areas, thereby advancing their long-term goals of sustainable development.

6. Further Research

6.1. Heterogeneity Analysis

6.1.1. Ambidextrous Green Innovation Balancing

Organizational dual theory suggests that enterprises can achieve a better balance between exploratory and exploitative green innovation for better balancing current development and future development. Therefore, the level of balance in ambidextrous green innovation in manufacturing companies may have different impacts on the role of AI in ESG performance. Following the approach of scholars like Gastaldi et al. (2022) [36], this study calculates the balance of ambidextrous green innovation in enterprises based on ${\displaystyle \frac{Exploitation+Exploration}{\left|Exploitation-Exploration\right|}}$ and conducts regression tests by grouping according to high and low mean values. The results in Columns (1) and (2) of

Table 8. Further analysis of regression results.

	(1)	(2)	(3)	(4)	(5)
	High Ambidexterity	Low Ambidexterity	LE	SME	TFP
AI	0.004 ***	0.001	0.004 ***	−0.000
	(0.001)	(0.001)	(0.001)	(0.001)
ESG					0.042 ***
					(0.005)
Size	0.242 ***	0.133 ***	0.106 ***	0.058 *	0.420 ***
	(0.024)	(0.016)	(0.010)	(0.031)	(0.008)
Owner	−0.050	0.044	0.052 **	−0.083	−0.228
	(0.067)	(0.039)	(0.025)	(0.057)	(0.320)
Int10	0.001	0.006 ***	0.005 ***	0.017 ***	−0.001 *
	(0.002)	(0.001)	(0.001)	(0.002)	(0.0004)
Boardsize	−0.013	0.019 *	−0.00500	−0.025	0.016 ***
	(0.020)	(0.011)	(0.007)	(0.017)	(0.004)
Roa	0.025 ***	0.001	0.021 ***	0.001	0.002 ***
	(0.004)	(0.001)	(0.001)	(0.001)	(0.0002)
Cash	0.007	0.009 ***	0.007 ***	0.007 **	0.002 ***
	(0.006)	(0.003)	(0.002)	(0.003)	(0.001)
Constant	3.106 ***	3.046 ***	3.270 ***	2.778 ***	6.124 ***
	(0.211)	(0.120)	(0.0801)	(0.179)	(0.309)
Observations	979	3819	8592	2149	11,748
R-squared	0.143	0.036	0.057	0.055	0.906

Standard errors in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1.

demonstrate that manufacturing firms with a high balance of ambidextrous green innovation significantly enhance the impact of AI on their ESG performance. Conversely, in firms with a low balance of ambidextrous green innovation, the application of AI does not exert a statistically significant influence on ESG performance. This could be due to high-balance companies being able to better synchronize economic and environmental benefits, adapt to rapid market changes, create greater value for the company, stimulate disruptive technology development, possess greater capacity to resolve internal conflicts, and allocate resources efficiently to enhance overall efficiency, leading to better application of AI technology. AI applications can maximize the potential of data, optimize manufacturing processes, and achieve green sustainable development goals, thereby enhancing the overall ESG performance of companies.

6.1.2. Enterprise Size

Acknowledging the disparities among firms of varying sizes in terms of risk resilience, governance standards and financing strategies, it is plausible that the influence of AI on corporate ESG performance may exhibit heterogeneity across these scales. Consequently, to investigate such scale-specific variations, this study stratifies the sample based on the median annual industry size, categorizing firms into large- and small-to-medium-sized enterprise groups (SMEs) for separate empirical scrutiny. As per the regression outcomes presented in Columns (3) and (4) of Table 8, within the cohort of large firms, the coefficient associated with AI is positive and statistically significant at the 1% level, implying that the adoption of AI technologies positively contributes to the enhancement of ESG performance among large-scale manufacturing entities. Conversely, in the SME group, the coefficient for AI fails to attain statistical significance, indicating no discernible relationship between AI utilization and ESG outcomes in this smaller-scale context. This may be because the development and application of AI in manufacturing companies require collaboration between digital talent, infrastructure hardware, and digital software. Large enterprises excel in risk management, human capital, and technological innovation potential compared to SMEs. Additionally, large-scale companies have ample R&D funding, sophisticated AI algorithms, and data accumulation, enhancing their research success rate and thus boosting their ESG performance.

6.2. Economic Consequences of Enhanced ESG Performance in Manufacturing Companies

According to the analysis above, manufacturing companies can improve their ESG performance through AI technology. Existing literature has found that high ESG performance not only gains the trust and support of stakeholders and reduces corporate financing costs and difficulties [37] but also signals a focus on green sustainable development to the external world, helping to enhance the social reputation and image of the company [38], promoting the construction of beneficial value networks for the company, and strengthening long-term cooperative relationships with various stakeholders. Moreover, companies focusing on ESG also tend to prioritize environmental protection and sustainable development, attracting more resources and financial support [39], leading them to proactively fulfill environmental and social responsibilities, and ultimately driving high-quality development in manufacturing companies.

Consequently, this study delves further into the impact of ESG performance on the high-quality development of manufacturing enterprises. To this end, this study employs total factor productivity (TFP) calculated via the LP method as an indicator of high-quality development within the manufacturing sector, examining the influence of enhanced ESG performance on this dimension. The findings in Column (5) of Table 8 reveal that the regression coefficients for ESG are consistently and significantly positive at the 1% level, substantiating a strong positive association between ESG performance and high-quality development in manufacturing firms. Hence, manufacturing enterprises should strive for a balanced development that harmonizes economic, social, and environmental benefits, thereby boosting TFP and ultimately facilitating their progression towards high-quality growth.

7. Conclusions

Against the backdrop of the global pursuit of green sustainability, enhancing the ESG performance of manufacturing enterprises assumes paramount significance. As a disruptive technology, the application of AI has been shown to facilitate corporate transitions towards sustainable development within production and innovation contexts [40]. Despite this, the question of whether and how AI contributes to improving the ESG performance of manufacturing firms remains underexplored. The key findings are as follows: (1) The adoption of AI significantly enhances the ESG performance of manufacturing firms. This conclusion remains robust after a series of sensitivity checks and the mitigation of endogeneity concerns. This study overcomes the limitation of the existing research, which primarily focuses on the economic effects of AI, by broadening the scope to include non-economic consequences, particularly ESG performance, in manufacturing firms. (2) Mechanism analysis reveals that AI drives ESG performance improvements through two channels: explorative green innovation and exploitative green innovation. Through a systematic analysis of the specific channels through which AI influences ESG performance, this study provides deeper insights into how AI can be utilized to enhance ESG performance, thereby contributing to the theoretical framework of sustainable development. (3) Heterogeneity analysis shows that the effect of AI on the ESG performance of manufacturing companies varies significantly based on different balances of ambidextrous green innovation and company size characteristics. Specifically, in manufacturing companies with a high balance of ambidextrous green innovation and large scale, the promotion effect of AI on the ESG performance of manufacturing companies is stronger. This nuanced understanding helps tailor AI strategies to different contexts, enhancing the practical applicability of the findings.

7.1. Recommendations

Based on the aforementioned research conclusions, this study aims to better connect the research outcomes with economic reality, considering how different stakeholders can benefit and effectively utilize these findings.

Government Level: First, the government should accelerate the deep integration of AI with industrial manufacturing, promoting the implementation of smart and green manufacturing initiatives to foster a market and policy environment conducive to AI-driven innovation. Second, the government should avoid one-size-fits-all policies for intelligent transformation and instead implement differentiated policies tailored to different industries and enterprise sizes, with increased support for SMEs in adopting AI technologies.

Manufacturing Enterprise Level: First, manufacturing enterprises should actively leverage AI to drive digital transformation and achieve green sustainability. By developing AI plans focused on environmental sustainability, companies can enhance operational efficiency and business agility while reducing environmental impacts. Second, enterprises should establish a comprehensive digital dual-green innovation management system to standardize operations, ensure the rational allocation of dual-innovation resources, and facilitate efficient coordination of innovation processes. This will drive the balanced development of dual-green innovation, further enhancing the company’s overall competitiveness and sustainability.

7.2. Limitations and Future Research

From the perspective of limitations, while exploratory green innovation and exploitative green innovation play a significant mediating role in enhancing ESG performance, the path mechanisms examined in this study do not represent the sole avenues through which AI can elevate ESG metrics for manufacturing firms. More efficacious mechanisms and pathways warrant further exploration and empirical validation. Future research is encouraged to employ broader datasets, encompassing diverse industries for comparative analysis, to augment the comprehensiveness and generalizability of the derived insights.