How Can Generative AI Promote Corporate ESG Performance? Evidence from China

Abstract

Generative AI has surfaced as a key driving force for corporate sustainable development and strategic transformation, offering new perspectives for effectively enhancing corporate ESG performance practices. Utilizing panel data sourced from Chinese A-share listed firms spanning the years 2012 to 2024, this research establishes and substantiates a model elucidating the mechanism by which generative AI impacts corporate ESG performance. The findings reveal the subsequent points: First, generative AI can effectively drive improvements in corporate ESG performance. Second, the caliber of information disclosure acts, in part, as an intermediary factor influencing the correlation between generative AI and corporate ESG performance enhancement. Third, sustainable innovation partially mediates the relationship between generative AI and corporate ESG performance enhancement. Fourth, environmental regulations weaken the beneficial influence exerted by generative AI on a company’s ESG achievements. Fifth, compared to non-manufacturing firms, companies situated in the central and western parts of China, and non-technology-intensive firms, the application of generative AI exerts a more pronounced enhancing impact on ESG achievements in manufacturing firms, firms in eastern regions, and technology-intensive firms. The research findings provide new insights for improving corporate ESG performance and provide strategic guidance for businesses aiming to attain long-term sustainable growth through reliance on generative AI.

1. Introduction

Currently, the global economic and social landscape is at a critical juncture of transformation and upgrading. Faced with an increasingly severe environmental situation and resource constraints, accelerating the green transition of development models has become an inevitable requirement for promoting high-quality economic development. Corporate Environmental, Social, and Governance (ESG) performance has garnered significant attention from global regulators, investors, and consumers. It serves not only as a crucial metric for evaluating a company’s sustainable development capabilities but also as a key driver for long-term value creation. Effectively enhancing ESG performance has thus emerged as a vital issue confronting enterprises. Meanwhile, the rapid development of generative artificial intelligence and its deep integration into corporate operations have become focal points of attention in both academic and industrial circles, particularly concerning its impact on corporate ESG performance. Although existing literature has extensively explored the influence of AI technologies on corporate sustainable development [1], given that academic research on generative AI is still in its exploratory phase, there remain significant and testable research gaps regarding how generative AI specifically affects corporate ESG performance and the underlying mechanisms at play.

Existing research primarily focuses on traditional artificial intelligence, with studies often emphasizing its applications in enhancing production efficiency and optimizing decision-making processes. However, there is a lack of detailed mechanistic analysis regarding its unique impacts on ESG dimensions, particularly in the areas of Social and Governance [2]. For instance, some studies have found that the adoption of artificial intelligence significantly improves overall ESG performance, with more pronounced effects observed in large enterprises and non-heavily polluting industries [3]. Other research indicates that corporate digital transformation significantly promotes sustainable development goals through the adoption of generative artificial intelligence (GAI) technologies [4]. Nevertheless, these studies often conflate generative AI with broader digital transformation or narrower AI technologies, failing to adequately identify the unique causal effects of generative AI—especially those tools based on foundation-model-driven approaches with text/image/code generation capabilities—on ESG performance.

In fact, there are differences between generative artificial intelligence and traditional artificial intelligence in terms of underlying logic and form comprehension. Traditional AI focuses more on analyzing existing data and making decisions [5], whereas generative AI creates new and unique content “out of nothing” by learning complex patterns [6]. Artificial intelligence technology is continuously evolving, and generative artificial intelligence represents a new stage in the development of AI technology. It is capable of providing human-like responses and streamlining repetitive tasks, thereby helping to enhance corporate productivity. By implementing generative AI technology, enterprises can gain a competitive edge, innovate their business models, and capture commercial value. Digital transformation, on the other hand, is a broader and more strategic concept. It represents an all-encompassing process of change in which organizations leverage digital technologies (including, but not limited to, artificial intelligence, cloud computing, the Internet of Things, and big data) to reshape their business processes, customer experiences, operational models, and corporate cultures [7]. In summary, digital transformation is a strategic process through which enterprises achieve comprehensive upgrades using digital technologies, while traditional AI and GAI serve as powerful technological tools driving this transformation. Traditional AI excels at optimizing and automating existing processes, whereas generative AI leads digital transformation into a new phase by creating new content and providing innovative solutions, bringing about deeper-level changes and value creation.

At present, there is relatively little quantitative empirical research on the impact of specific generative AI technological applications on ESG performance. How generative AI empowers ESG performance remains an underexplored area. On the one hand, although studies have constructed AI measurement indicators using machine learning based on the annual reports of Chinese listed companies and found that AI usage significantly improves ESG performance [8], as well as studies that have utilized China’s AI pilot regions as quasi-natural experiments to evaluate the positive impact of AI on ESG [9], these studies primarily focus on AI in a broad sense rather than generative AI. Moreover, there is relatively little empirical research using large-sample, cross-industry, and longitudinal panel data on the impact of specific generative AI applications on ESG performance. On the other hand, although some studies have proposed that GAI can significantly enhance environmental performance [10], the detailed causal pathways through which generative AI influences ESG performance through specific generative processes remain unclear. This “black box” effect in terms of mechanisms limits our comprehensive understanding of how generative AI empowers ESG.

In light of the identified research gaps, the present study develops a model that elucidates the operational mechanism through which generative AI influences corporate ESG (Environmental, Social, and Governance) performance. This model is constructed using provincial panel data sourced from China, covering the period from 2012 to 2024. Within the research framework, the study integrates the quality of corporate information disclosure and sustainable innovation to scrutinize the interaction mechanism between generative AI and corporate ESG performance. Moreover, it delves into the conditional effects exerted by environmental regulations. This paper endeavors to uncover novel avenues for bolstering corporate ESG performance, providing a distinctive vantage point for dissecting the underlying driving forces and offering insightful guidance for fostering their progression.

The contributions of this study may lie in the following aspects: First, theoretical contributions. This study will integrate information asymmetry theory, the resource-based view, and institutional theory, and introduce variables such as information disclosure quality, sustainable innovation, and environmental regulation to elucidate how generative artificial intelligence drives the improvement of ESG performance. This goes beyond the single research perspective in existing literature and reveals, at a deeper level, its strategic role in shaping corporate sustainable development behaviors. Second, methodological contributions. This study will leverage the staggered adoption of generative artificial intelligence tools by firms and employ the Difference-in-Differences (DID) method to identify the causal effect of generative artificial intelligence on corporate ESG performance. This will help overcome the endogeneity issues prevalent in existing research and provide stronger causal evidence for the relationship between generative artificial intelligence and ESG performance. Third, empirical contributions. This study will offer the first causal evidence on the impact of generative artificial intelligence adoption on corporate ESG performance and empirically quantify the extent of its improvement in ESG performance. Additionally, this study will verify the mediating effects of sustainable innovation and information disclosure quality in this process and empirically examine how environmental regulation moderates the influence of generative artificial intelligence on ESG performance. This will provide comprehensive theoretical insights and empirical evidence for corporate sustainable development in the context of generative artificial intelligence, offering valuable references for corporate strategic decision-makers, policymakers, and investors.

2. Literature Review

2.1. Generative AI

Generative Artificial Intelligence, as one of the most transformative technological directions within the contemporary domain of artificial intelligence, is profoundly reshaping multiple dimensions of scientific research, industrial applications, and societal operations. Existing literature on generative AI primarily focuses on the following aspects: First, conceptual definition. The essence of generative AI resides in its capacity to learn the underlying distributions within large-scale datasets through models, thereby autonomously generating content that is similar to the original data yet novel. This encompasses textual content, visual imagery, video clips, musical compositions, 3D object models, programming code, among others [11]. Contrary to typical discriminative AI, which prioritizes classification functions, generative AI specializes in creative tasks, capable of producing realistic and original content from basic user prompts [12]. Second, significant value. The importance of generative AI is reflected in its ability to drive innovation, enhance efficiency, and solve complex problems across various domains. These values primarily stem from its core characteristic of autonomously generating novel and high-quality content [13]. In terms of content creation and innovation, generative AI demonstrates disruptive value, significantly enriching the means of creative industries and content production. Third, enabling domains. Generative AI has the potential to empower a wide range of industries, including higher education [14], business management [15], and the financial sector [16], among others. In conclusion, the academic community has yet to conduct in-depth exploration into the disruptive impact of generative artificial intelligence on corporate ESG performance, and the enabling mechanisms through which generative AI enhances corporate ESG performance remain unclear.

2.2. Corporate ESG Performance

Currently, ESG effectiveness has surfaced a fundamental indicator for evaluating a company’s management capabilities, risk management proficiency, and non-financial performance. Scholarly investigations into corporate ESG mainly focus on the subsequent areas: First, conceptual definition. ESG performance represents a multidimensional and comprehensive concept that extends beyond traditional financial performance, highlighting a company’s performance across three fundamental dimensions: Environmental, Social, and Governance [17]. Specifically, environmental performance (E) primarily focuses on a company’s impact on and management of the natural environment. Social performance (S) pertains to a firm’s interactions and obligations toward employees, customers, suppliers, communities, and additional stakeholders. Governance performance (G) centers on a company’s internal management structure, decision-making processes, ethical standards, and transparency, including elements such as corporate governance structure (e.g., board composition and independence), risk management, tax transparency, and protection of shareholder rights. Second, measurement methods. The measurement of ESG performance typically relies on scoring systems supplied by external rating entities like MSCI, Sustainalytics, and SynTao Green Finance [18]. These scoring systems quantify a company’s ESG performance by evaluating its disclosed information and actual actions across environmental, social, and governance dimensions [19]. Third, influencing factors. The factors influencing corporate ESG performance are multifaceted, encompassing both internal and external elements. These factors include innovation capabilities [20], government policies [21] and so on. In conclusion, research on ESG performance in the academic community is still at an exploratory stage, with most studies being fragmented analyses based on a single theory or perspective. There is a lack of in-depth exploration into its driving mechanisms from an integrated perspective, which restricts the systematic understanding of corporate ESG performance in both academic and practical circles. Against the backdrop of the rapid development of generative artificial intelligence such as DeepSeek, how enterprises can align with this trend and leverage this technology to enhance ESG performance remains to be explored.

2.3. Findings and Conflicts in Existing Research

Existing research has not specifically delved into the relationship between generative artificial intelligence and ESG performance. A substantial number of studies generally recognize the potential of AI in enhancing ESG performance. For instance, AI technologies have been found to significantly improve corporate ESG performance and further amplify this positive effect by enhancing corporate performance and attracting analyst attention [20]. Among A-share listed companies in China, digital transformation has been proven to significantly elevate corporate ESG performance, and this effect is more pronounced under the modulation of both formal and informal environmental regulations [22]. AI can also effectively curb corporate “greenwashing” behavior and promote corporate ESG performance by improving the accuracy and transparency of environmental reports [23]. Although the positive effects of AI have garnered widespread attention, some studies have begun to focus on its potential negative impacts and the risks of “ESG opportunism.” For example, while existing research overwhelmingly emphasizes the positive effects of generative AI on corporate ESG performance, it overlooks the risk of dimensional imbalance in ESG resource allocation and its potential spread within the supply chain [24]. This indicates that the application of AI may lead companies to over-focus on certain ESG dimensions (such as the environment) while neglecting others (such as social or governance), thereby resulting in “ESG opportunism” behavior.

It is noteworthy that the mechanism through which AI influences ESG performance does not follow a single path. Some studies emphasize that AI transforms ESG information and enhances environmental performance through its “ESG sense-making capability” [10], while others focus on its ability to promote sustainable performance through “business model innovation” (efficiency-oriented and novelty-oriented) [25]. These different mediating mechanisms suggest that the role of AI is multi-dimensional and complex, necessitating more nuanced analysis. Additionally, there are certain methodological differences among existing studies, as illustrated by the representative literature in

Table 1. Research Methods, Findings, and Shortcomings in the Previous literature.

Authors	Method	Findings	Shortcomings
wael AL-khatib (2023) [26]; Wang & Zhang (2024) [4]	Questionnaire survey	These studies explore the role of generative AI in promoting sustainable development goals within digital supply chains, as well as the driving factors behind the adoption of generative AI and its impact on exploratory and exploitative innovation.	The survey sample has limitations.
Huarng & Yu (2024) [19]	Qualitative Comparative Analysis (QCA)	The study examined the role of generative AI in promoting sustainable development goals within digital supply chains.	Lack of large-sample data analysis.
Li et al. (2024) [22]; Sun et al. (2024) [27]	Econometric model	These studies have uncovered the impact of digital transformation on ESG performance, as well as the positive effects of ESG performance.	No measurement and analysis have been conducted on generative AI.
Shen & Badulescu (2025) [25]	Structural Equation Modeling (SEM)	The relationship between generative AI and the sustainable performance of Chinese manufacturing enterprises was analyzed.	There are industry-specific limitations, and the analysis only centers on sustainable performance, neglecting a discussion on ESG performance.
Yu et al. (2026) [28]	Case study	An analysis was conducted on how digital transformation affects ESG outcomes.	Artificial intelligence was treated as part of digital transformation, without distinguishing between generative AI and traditional AI.

.

In summary, existing research on generative artificial intelligence (AI) and corporate ESG (Environmental, Social, and Governance) performance has yielded fruitful results. However, current studies have not explored how generative AI can empower the improvement of corporate ESG performance. The crucial question of how generative AI drives the enhancement of corporate ESG performance remains an unopened “black box.” Within the ESG context, there is a fundamental difference between the creative capabilities of generative AI and the analytical capabilities of previous technologies. Generative AI can generate novel content in various forms, including text, images, code, and music. In the ESG field, this implies that GAI can generate innovative concepts for sustainable products or services, automatically produce personalized ESG report summaries, press releases, or social media content tailored to the needs of specific stakeholders, and develop new environmental monitoring and emission reduction strategies. The powerful content generation, data analysis, and intelligent interaction capabilities of generative AI theoretically have the potential to bring about innovative transformations in corporate ESG practices. However, there is currently a lack of systematic research on its underlying mechanisms. Given the singularity of research perspectives and the limitations of research methods in existing literature, this study integrates information asymmetry theory, resource-based view, and institutional theory, introduces variables such as information disclosure quality, sustainable innovation, and environmental regulations, and elucidates how generative AI promotes ESG performance improvement. It employs text analysis to measure generative AI and then relies on econometric models to explore the mediating and moderating mechanisms through which generative AI empowers corporate ESG performance, as well as conducting heterogeneity analysis. This represents an update and enrichment of the empowerment theory of generative AI and a profound reshaping of theoretical research on corporate ESG performance improvement in the context of generative AI.

3. Theoretical Analysis and Research Hypotheses

3.1. Generative AI and Corporate ESG Performance

The Resource-Based View (RBV) theory emphasizes that by effectively managing and leveraging their unique tangible and intangible resources, enterprises can build competitive advantages and achieve sustainable development. In the realm of ESG (Environmental, Social, and Governance), an enterprise’s resource endowments and capability allocation directly influence its performance. Generative AI, as a disruptive technology, can be regarded as a key strategic resource for enterprises to acquire and sustain competitive advantages. Generative AI serves as a crucial driver for enterprises’ sustainable development and strategic transformation [29]. Existing research indicates that the application of AI technology can significantly enhance an enterprise’s ESG performance, primarily through alleviating financing constraints, promoting corporate innovation, and improving resource utilization efficiency, among other channels [8]. Generative AI can optimize an enterprise’s energy usage efficiency, reduce carbon emissions, and enable more sustainable practices within the supply chain through data analysis and prediction. For instance, the study by Lin et al. (2024) [30] found that the application of AI can enhance supply chain efficiency in enterprises, thereby improving ESG performance. Meanwhile, through the analysis of employee and customer data, generative AI can assist enterprises in better understanding and meeting the needs of diverse groups, thereby enhancing diversity and inclusivity and strengthening the enterprise’s sense of responsibility towards customers [31]. AI can also enhance social responsibility performance by improving working conditions, enhancing employee health and safety management, and ensuring that human rights are respected in the supply chain [17]. Furthermore, generative AI can aid enterprises in identifying weak links in governance. For example, research has shown that it can significantly enhance transparency and accuracy in ESG reporting through automated reporting, real-time data processing, and predictive analytics [32]. Based on the above analysis, this study proposes the following hypotheses:

H1: 

Generative AI can effectively promote the improvement of corporate ESG performance.

3.2. The Mediating Effect of Information Disclosure Quality

The theory of information asymmetry posits that in a market economy, information asymmetry is ubiquitous, meaning one party possesses more information than the other, which can lead to issues such as adverse selection and moral hazard. In corporate operations, particularly in the realm of ESG (Environmental, Social, and Governance), the problem of information asymmetry is more pronounced due to the often complex, non-standardized, and difficult-to-quantify nature of relevant information. Leveraging its powerful capabilities in data analysis, pattern recognition, and content generation, Generative AI can process and integrate ESG data from diverse sources and formats. For instance, generative AI can extract ESG-related insights from unstructured data (such as social media comments and news reports) and transform them into structured report content. Previous research has found that generative AI can delve deeply into unstructured data to uncover implicit information, providing investors and other stakeholders with a more comprehensive and in-depth view of corporate risks, thereby enhancing the quality and decision-making value of financial information disclosure [33]. generative AI can automate the report generation process, ensuring information consistency and accuracy, reducing human errors, and customizing reports according to the needs of different stakeholders to enhance the relevance and comprehensibility of the information. By refining complex ESG data into meaningful and actionable information, generative AI facilitates more efficient ESG information disclosure by enterprises, thereby improving the transparency, timeliness, and completeness of such disclosures.

High-quality information disclosure can effectively reduce information asymmetry, making it easier for external parties to access and understand a company’s ESG performance. When companies provide transparent, accurate, and comprehensive ESG information, investors can better assess the company’s sustainable development risks and opportunities, potentially leading them to favor companies with superior ESG performance [10]. This incentivizes companies to place greater emphasis on ESG practices to attract more responsible investments. Furthermore, high-quality information disclosure enhances a company’s reputation among consumers, employees, and the general public, fostering trust, attracting talented individuals, and garnering community support [34]. By clearly, comprehensively, and accurately disclosing information regarding their environmental policies, carbon emissions, resource consumption, pollution control, and environmental compliance, companies can effectively enhance their environmental transparency and prompt improvements in their practices [35]. Research indicates that when companies provide detailed disclosures of their efforts and progress in reducing greenhouse gas emissions, they can attract investors concerned with sustainable development, leading to higher environmental performance [36]. Simultaneously, high-quality information disclosure strengthens trust between companies and their stakeholders, enhances corporate reputation, and promotes improvements in social performance [37]. Moreover, high-quality information disclosure serves as the foundation for a sound corporate governance structure, enhances decision-making transparency, and safeguards shareholders’ rights and interests [38], effectively reducing a company’s tail risk (i.e., the likelihood of extreme negative events). Based on the above analysis, this paper proposes the following hypotheses:

H2: 

Generative AI can promote corporate ESG performance by improving the quality of corporate information disclosure.

3.3. The Mediating Effect of Sustainable Innovation

From the perspective of the Resource-Based View (RBV), generative AI represents a valuable and rare technological resource that can integrate with a company’s existing knowledge, organizational structure, and culture, thereby enhancing the value of its existing sustainability-related resources. Generative AI can accelerate the achievement of sustainable development goals by improving efficiency and fostering innovation [4]. Generative AI enables in-depth mining and analysis of vast amounts of literature and experimental data, reducing the uncertainty and R&D costs associated with sustainable innovation technologies through predictive assessments of technological environmental impacts and optimization of technological reaction pathways, thus enhancing corporate sustainable innovation efficiency. Generative AI can be applied to optimize energy consumption, reduce material waste, and improve energy efficiency across various industries [39]. For instance, in supply chain management, generative AI can assist in creating efficient and sustainable supply chains [40]. In the energy sector, AI can leverage advanced predictive analytics, blockchain integration, and autonomous energy systems to optimize energy production, consumption, and management, thereby driving the transition to sustainable energy [41].

According to the RBV, sustainable innovation can be regarded as a critical internal capability or resource combination that promotes the enhancement of a company’s ESG performance and ultimately forms a sustainable competitive advantage. Previous research suggests that sustainable innovation significantly improves corporate environmental performance by driving the adoption of clean technologies, optimizing resource utilization, and reducing pollution emissions [42]. Sustainable innovation encourages companies to re-examine their resource utilization patterns, such as transforming waste into valuable resources [18], laying the foundation for improved environmental performance. During the process of sustainable innovation, companies typically place greater emphasis on employees’ working environments, health and safety, and career development. For example, by introducing safer and more ergonomic new technologies, production accidents can be reduced, and employee satisfaction can be enhanced [17]. Sustainable innovation can also be reflected in a company’s interactions with local communities, such as addressing community environmental issues through technological innovation or providing employment opportunities and skills training. This proactive community engagement helps enhance a company’s social image and strengthens its “license to operate” within the community [43]. Furthermore, companies pursuing sustainable development often establish more robust governance structures that incorporate ESG factors into their decision-making processes. The establishment of such mechanisms helps ensure that sustainable innovation projects receive adequate resources and support while reducing agency costs [44]. Based on the above analysis, this study proposes the following hypotheses:

H3: 

Generative AI can improve corporate ESG performance by promoting sustainable innovation in enterprises.

3.4. The Moderating Effect of Environmental Regulation

Institutional theory posits that; external institutional factors inevitably restrict how generative AI impacts a company’s ESG performance. Environmental regulation, as one of the institutional arrangements through which the government intervenes in corporate behavior to promote green and environmentally sustainable progress, could influence the interplay between generative AI and corporate ESG performance regarding both implementation intensity and scope. Differing regional enforcement levels exist for environmental regulations [45]. Strict environmental regulations require companies to increase their investments in environmental protection equipment procurement, pollution control, environmental monitoring, and other areas, which undoubtedly raises their operational costs [46]. From the perspective of resource-based theory, a company’s resources are limited, and these cost investments will crowd out the resource allocation for generative artificial intelligence within the company. Meanwhile, stringent environmental regulations subject companies to significant compliance pressures. To avoid penalties and maintain a positive social image, companies will direct more strategic attention towards environmental compliance. Under such circumstances, companies may reduce their investments in and attention to projects related to generative artificial intelligence, shifting the time and effort originally dedicated to investigating the utilization of generative AI in enhancing ESG performance towards meeting specific environmental regulation requirements. Building on this, the current research posits the following hypothesis:

H4: 

As the intensity of environmental regulations increases, the enhancing effect of generative artificial intelligence on firm-level ESG outcomes may be weakened.

4. Research Design

4.1. Sample and Data Sources

The initial sample of this paper consists of data from Chinese A-share listed companies spanning from 2012 to 2024. The interpolation method is employed to eliminate corporate data with severe missing values and data from companies with observations only in a single year, ensuring the structural integrity of the panel data. Meanwhile, there are systematic differences in accounting standards, asset-liability structures, and regulatory rules between the financial industry and non-financial enterprises. The leverage levels, profit models, and risk structures of financial industry entities are not comparable to those of real sector enterprises [47]. To ensure business homogeneity among the research subjects, this paper, referring to the study by Xi (2025) [48], excludes samples from the financial industry, companies with financial abnormalities (ST, *ST), samples with missing variables, and single-observation samples. In accordance with the newly revised “Guidelines for the Industry Classification of Listed Companies” by the China Securities Regulatory Commission in 2022, the sample, after excluding the financial industry, covers 79 sub-sectors outside the financial sector. Furthermore, to eliminate potential interference from extreme values on the research findings, this paper also conducts a 1% winsorization at both the upper and lower tails for continuous variables, ultimately yielding 23,631 valid observations, of which the number of independent enterprises is 3733.

4.2. Variable Measurement

4.2.1. Core Explanatory Variable: Generative AI (GAI)

This study utilizes a text analysis method to measure the frequency of terms related to the application of generative artificial intelligence in annual reports, thereby assessing the level of GAI adoption by enterprises. Building upon the findings of Qiao et al. (2025) [49], we select keywords, segment the text content of annual reports, and extract relevant text to determine the number of characteristic terms associated with GAI. We then add one to this count and take the natural logarithm to derive the GAI variable. This paper draws reference from the research by Qiao et al. (2025) [49]. It selects keywords, conducts word segmentation and text extraction on the textual content of annual reports. It takes three dimensions—the foundational layer, the technological layer, and the model ecosystem layer—as the search scope for generative artificial intelligence (GAI) seed words. Using the “2022 White Paper on the Development of Large Models in China,” the “White Paper on Artificial Intelligence Generated Content (AIGC),” the “White Paper on the Development of AI Large Models,” and the “White Paper on the Technology of Large Artificial Intelligence Models in China (2023 Edition)” as corpora, it extracts the preceding and succeeding text from the total sample of listed companies and searches for frequently occurring text combinations. This process yields the count of generative artificial intelligence characteristic words, and by adding one to this count and then taking the natural logarithm, the GAI metric is obtained. The dictionary of keywords for generative artificial intelligence is shown in

Table 2. Dictionary of Keywords for Generative Artificial Intelligence.

Dimension	Categorized Terms	Word Segmentation Dictionary
Conceptual foundation layer	Artificial intelligence, generative artificial intelligence, large language model, pre-training	Generative AI, Large AI, Large model, AIGC, Pre-trained model, Pretrained Model, Large Language Model, AI Foundational Model, LLM
Core technology layer	Natural language, architecture, autoregressive, generative adversarial, autoencoder, diffusion, graph, model, multimodal	Natural Language Processing, NLP, Knowledge Graph, Transformer Architecture, TensorFlow, PyTorch, Keras, Caffe, MXNet, PaddlePaddle, Capsule Network, GAN, Diffusion Models, GAN, VAEs, Variational Autoencoder, Autoregressive Model, Autoregressive Models, Flow Model, Flow-based Models, Multimodal Generative Architecture, DALL-E
Model ecosystem type	Natural language processing models, large models, models based on the Transformer architecture, image generation models, multimodal large models	Bert, GPT, ChatGPT, XLM, ERNIE, Vit, iFLYTEK Spark Large Model, Qianwen, ERNIE Bot, Lenet, AlexNet, ResNet, Mobilenet, Catalyst, TFX, EfficientNet, Keras, transformers, Horovod, Luminous, DETR, GRU, Torch, Bloom, CTRL, GLM, Pangu Large Model, Hunyuan Large Model, LSTM, DGL, Caffe2, CPM, Pythia, LLaMA, Baichuan Large Model, T5, CPT, OPT, MPT, OpenFlamingo, mPLUG-Owl, KOSMOS-2, ImageBind

.

4.2.2. Dependent Variable: Corporate ESG Performance (ESG)

Currently, there are multiple ESG assessment institutions within China, including China Securities Index (CSI), ChinaBond, Bloomberg, and SynTao Green Finance. Considering the sample of Chinese A-share listed companies and the integrity of the data, this paper follows the approach of Xie & Lv (2022) [50] by utilizing the ChinaBond ESG rating system to assess corporate ESG performance. This rating system categorizes ESG performance into nine grades, which are assigned values from 1 to 9, respectively. This paper uses the annual average ESG score of enterprises to represent their ESG performance. The raw data are sourced from the Wind Database.

4.2.3. Mediating Variables: Information Disclosure Quality (IQ) and Sustainable Innovation (SI)

Information Disclosure Quality (IQ). Drawing on the research conducted by Shi & Mai (2025) [51], this paper selects the number of analysts tracking listed companies as an alternative indicator to measure the quality of information disclosure. In the real-world context of information asymmetry, investors and other relevant stakeholders often find it challenging to obtain accurate and timely information about enterprises. Serving as a crucial information communication bridge between listed companies and investors, as well as other stakeholders, securities analysts can leverage their professional skills to comprehensively collect, in-depth analyze, and accurately judge various types of information pertaining to listed companies. Based on this, when a greater number of analysts track a particular listed company, it to some extent reflects that the company has a higher quality of information disclosure. This is because it implies that securities analysts are able to grasp and convey more information related to the enterprise, thereby enabling external stakeholders to more easily access high-quality corporate information. Therefore, this paper takes the logarithm of the number of analysts tracking listed companies plus one as a key indicator for evaluating the quality of corporate information disclosure.

Sustainable Innovation (SI). Following Wang & Wang (2021) [52], the quantity of green patent applications (including both invention patents and green utility model patents) is selected as the measurement indicator. During the specific measurement process, this study obtained patent classification number information for invention patents and utility model patents from all A-share listed companies in China from the China Research Data Service Platform (CNRDS) and matched it with the “International Patent Classification Green Inventory” released by the World Intellectual Property Organization (WIPO) in 2010. Based on the matching results, the patents of listed companies were classified into green patents (green invention patents and green utility model patents) and non-green patents (non-green invention patents and non-green utility model patents). The total amount of green innovation, denoted as Total, was obtained by summing the number of applications for green invention patents and green utility model patents. The number of applications for green invention patents, Inva, was used to measure the quality of green innovation, while the number of applications for green utility model patents, Uma, was employed as a comparative indicator to measure the quantity of green innovation. To address the right-skewed distribution issue of green patent application data, this paper took the natural logarithm of the number of green patent applications after adding 1.

4.2.4. Moderating Variable: Environmental Regulation (ER)

This study incorporates methodologies from Chen et al. (2016) [53] and Chen et al. (2018) [54], employing the ratio of the frequency of terms associated with “environmental protection” in local government work reports to the overall word count of these reports as an indicator. Firstly, we manually collect the annual government work reports of various cities. As a programmatic document guiding governmental work, the content related to environmental governance within these reports can comprehensively reflect the policy orientation and implementation intensity of local governments’ environmental regulations, serving as an authoritative vehicle for measuring this variable. Secondly, we conduct word segmentation on the collected reports and filter out terms and corresponding sentences related to environmental governance. The proportion of the total number of words in these sentences relative to the total number of words in the government work report is used as the core indicator of environmental regulation intensity. Compared to the measurement approach based solely on the frequency proportion of single environmental protection terms, this method not only fully retains the semantic and syntactic meanings of the policy text but also effectively avoids measurement biases that may arise from simple term counting, thereby enhancing the accuracy of the indicator. Finally, since government work reports are typically released at the beginning of the year while economic activities span the entire year, this indicator possesses inherent exogeneity, effectively mitigating endogeneity issues caused by reverse causality and further ensuring the rigor of the empirical analysis.

4.2.5. Control Variables

Drawing on the research by Zhou (2025) [55], this study accounts for various factors that could potentially affect the link between generative AI and corporate ESG performance. Specifically, these variables include: corporate size (Size), corporate age (Age), debt-to-asset ratio (DAR), return on assets (ROA), proportion of accounts receivable (PAR), proportion of inventory (INV), Tobin’s Q ratio (TQR), and proportion of institutional investors (PII).

In

Table 3. Main Variable Measurement Methods.

Variable Names	Variable Symbols	Variable Measurement
Corporate ESG Performance	ESG	Annual average value of Huazheng ESG assessment index system
Generative AI	GAI	The logarithmic value (base e) of (the quantity of keywords related to the use of generative AI technology plus one)
Information Disclosure Quality	IQ	The logarithm of (the count of analysts monitoring listed firms, incremented by one)
Sustainable Innovation	SI	Natural log of (the combined total of a company’s green invention patent filings and green utility model patent filings, incremented by one)
Environmental Regulation	ER	The ratio of the occurrence frequency of environmental protection-related terms appearing within the textual corpus of local government reports to the total word count of the reports.
Corporate Size	Size	Logarithmic value (base e) of (a firm’s aggregate assets incremented by one)
Corporate Age	Age	Natural logarithm of (the difference between the current year and the company’s listing year, incremented by one)
Debt-to-Asset Ratio	DAR	The proportion of total liabilities to total assets
Return on Assets	ROA	Net profit relative to total assets
Proportion of Accounts Receivable	PAR	Net accounts receivable as a proportion of total assets
Proportion of Inventory	INV	Net inventory relative to total assets
Tobin’s Q Ratio	TQR	Market capitalization as a percentage of total assets
Proportion of Institutional Investors	PII	Institutional investors’ shareholding ratio in listed companies

, the measurement methods for the main variables included in this research are outlined.

4.3. Model Design

4.3.1. Benchmark Regression Model

Assessing the impact of generative AI technologies on a company’s ESG performance, this paper constructs a benchmark model with dual fixed effects for industries and years. To examine the rationality of the industry-year two-way fixed-effects model, this paper employs the Sigmamore robust Hausman test to compare the fixed-effects (FE) model with the random-effects (RE) model. The results show that χ²(95) = 548.02, p = 0.0000, strongly rejecting the null hypothesis that “the random-effects model is superior” at the 1% significance level. This indicates the presence of individual heterogeneity in the model that is correlated with the explanatory variables, suggesting that the fixed-effects model is the optimal choice. Therefore, all subsequent regressions in this paper utilize the industry-year two-way fixed-effects model. The specific formulation of the benchmark regression model is as follows:

ESG_i,t = α₀ + α₁GAI_i,t + β_jControl_j,i,t + ∑Year + ∑Industry + ε_i,t

(1)

Here, firm and year are indicated by the subscripts i and t; ESG_i,t signifies the ESG score of firm i in year t; GAI serves as the primary independent variable, measuring the firm’s level of generative artificial intelligence; Control denotes firm-level factors controlled for their potential impact on ESG outcomes. Based on relevant literature and the research needs of this paper, we primarily control for the effects of firm size, firm age, the relationship between leverage, profitability, and the accounts receivable ratio with corporate ESG performance. Furthermore, industry and year fixed effects are introduced to isolate the impact of other macroeconomic factors, and ε_i,t is the random disturbance term.

4.3.2. Mediation Effect Model

The mediation effect model employed in this research is structured as follows:

IQ = α₂ + δ₁GAI_i,t + δ2Control_j,i,t + ∑Year + ∑Industry + ε_i,t

(2)

ESG = α₃ + γ₁GAI + γ₂IQ + γ₃Control_j,i,t + ∑Year + ∑Industry + ε_i,t

(3)

SI = α4 + δ2GAIi,t + δ3Controlj,i,t + ∑Year + ∑Industry + εi,t

(4)

ESG = α5 + γ4GAI + γ5IQ + γ6Controlj,i,t + ∑Year + ∑Industry + εi,t

(5)

Here, IQ and SI serve as the mediating variables, representing corporate information disclosure quality and sustainable innovation capability, respectively. First, we examine the parameter β corresponding to Model (1). If β shows statistical significance, it indicates that the prerequisite for the existence of a mediation effect is satisfied. Next, we observe the coefficient δ₁ in Model (2); if δ₁ is significant, it suggests that generative artificial intelligence has an impact on corporate information disclosure quality. Subsequently, we examine the coefficients γ₁ and γ₂ in Model (3); if both are significant, we then compare coefficient β in Model (1) with coefficient γ₂ in Model (3) to determine whether a partial mediation effect exists.

4.3.3. Moderating Effect Model

The model examining moderating effects is specified as follows:

ESG = α6 + γ7GAIi,t + γ8GAIi,t∗ERi,t + γ9ERi,t + γ10Controlj,i,t + ∑Year + ∑Industry + εi,t

(6)

First, based on the primary econometric specification, the moderating effect is tested. If both α₁ and the interaction term γ₈ in Equation (6) are statistically significant, it provides evidence for the existence of a moderating effect. If α₁ and γ₈ share the same sign, it indicates that ER exerts a positive moderating effect; conversely, if their signs are opposite, it suggests that ER exerts a negative moderating effect.

5. Empirical Analysis

5.1. Descriptive Statistics

Table 4. Descriptive Statistical Results of Key Variables.

Variable	Observations	Mean	Standard Deviation	Minimum	Maximum
ESG	23,631	4.255	0.829	2.000	6.000
GAI	23,631	0.123	0.378	0.000	2.197
IQ	23,631	1.993	0.903	0.693	3.850
SI	23,631	1.134	1.363	0.000	7.466
ER	23,631	0.825	0.183	0.000	8.943
Size	23,631	22.601	1.327	20.287	26.631
Age	23,631	2.177	0.765	0.693	3.367
DAR	23,631	0.426	0.197	0.060	0.864
ROA	23,631	0.046	0.055	−0.170	0.205
PAR	23,631	0.121	0.101	0.000	0.465
INV	23,631	2.058	1.295	0.830	8.224
TQR	23,631	0.136	0.122	0.000	0.664
PII	23,631	0.461	0.250	0.009	0.922

displays the descriptive statistics for the key variables. Corporate ESG performance (ESG) has a mean of 4.255, ranging from a minimum of 2 to a maximum of 6, which reflects notable diversity in ESG outcomes among firms. Additionally, the average value of generative AI level (GAI) is 0.123, with a maximum of 2.197 and a minimum of 0.000, which is consistent with findings from existing literature and suggests notable differences in the adoption levels of generative artificial intelligence among firms. The relevant characteristics of all control variables align with descriptions in related studies.

5.2. Correlation Between Key Variables

Before conducting empirical analysis, it is essential to verify the correlations between the primary variables.

Table 5. Correlation Coefficient Matrix of Key Variables.

	1	2	3	4	5	6	7	8	9	10	11	12	13
1. ESG	1
2. GAI	0.085 ***	1
3. IQ	0.224 ***	0.026 ***	1
4. Patent	0.147 ***	0.089 ***	0.182 ***	1
5. ER	−0.010	0.038 ***	−0.003	−0.024 ***	1
6. Size	0.192 ***	0.024 ***	0.291 ***	0.474 ***	−0.003	1
7. Age	−0.023 ***	0.001	−0.024 ***	0.184 ***	0.023 ***	0.529 ***	1
8. DAR	−0.063 ***	−0.046 ***	−0.019 ***	0.256 ***	0.014 ***	0.552 ***	0.367 ***	1
9. ROA	0.139 ***	−0.050 ***	0.350 ***	−0.042 ***	−0.014 ***	−0.077 ***	−0.153 ***	−0.391 ***	1
10. PAR	−0.035 ***	0.071 ***	−0.040 ***	0.129 ***	0.037 ***	−0.221 ***	−0.215 ***	0.009	−0.024 ***	1
11. INV	−0.049 ***	0.047 ***	0.173 ***	−0.151 ***	−0.035 ***	−0.377 ***	−0.156 ***	−0.349 ***	0.287 ***	0.056 ***	1
12. TQR	0.063 ***	−0.061 ***	−0.012 *	−0.076 ***	0.004	0.118 ***	0.100 ***	0.292 ***	−0.070 ***	−0.070 ***	−0.072 ***	1
13. PII	0.082 ***	−0.040 ***	0.197 ***	0.145 ***	−0.006	0.459 ***	0.271 ***	0.222 ***	0.083 ***	−0.221 ***	−0.058 ***	0.037 ***	1

Note: *, and *** denote significance levels of 10% and 1%, respectively.

summarizes the detailed results. At the 1% significance level, a positive correlation (0.085) is observed between corporate GAI implementation and ESG outcomes. This result preliminarily indicates that a firm’s adoption of generative AI demonstrates a positive influence on its ESG performance.

5.3. Hypothesis Testing for Main Effects

Results from the baseline regressions are presented in

Table 6. Baseline Regression Results.

	(1) ESG	(2) ESG	(3) ESG
GAI	0.0687 *** (4.96)	0.164 *** (12.07)	0.0460 *** (3.56)
Size		0.239 *** (40.60)	0.264 *** (42.86)
Age		−0.147 *** (−18.34)	−0.143 *** (−18.29)
DAR		−0.969 *** (−26.59)	−0.898 *** (−24.08)
ROA		1.186 *** (11.03)	1.172 *** (10.51)
PAR		0.217 *** (4.05)	0.0157 (0.26)
INV		−0.0172 (−3.83)	−0.0066 (−1.43)
TQR		0.742 *** (16.95)	0.209 *** (3.72)
PII		−0.0292 (−1.24)	−0.0193 (−0.84)
_cons	4.246 *** (787.08)	−0.557 *** (−4.58)	−1.092 *** (−8.57)
Year	Yes	No	Yes
Industry	Yes	No	Yes
N	23,631	23,631	23,631
R2	0.093	0.115	0.194

Note: *** denote significance levels of 1%.

. Starting with a parsimonious model in column (1) that regresses ESG performance on GAI alone, the coefficient is positive and significant at the 1% level. Column (2) enriches this model by accounting for firm-level characteristics, yet the positive and significant (1% level) relationship for GAI holds. The full model in column (3), which also controls for industry-year fixed effects, continues to show a significantly positive coefficient for GAI at the 1% level. Specifically, the estimates imply a 0.046% gain in ESG performance per 1% increase in generative AI levels. This consistent pattern across progressively more rigorous models demonstrates a significant positive effect of generative AI on corporate ESG outcomes. The positive impact of generative AI on ESG performance indicates that firms can leverage this cutting-edge technology to enhance their environmental, social, and governance practices. This not only helps companies meet increasingly stringent regulatory requirements but also improves their reputation among stakeholders, including investors, customers, and the general public. A better ESG standing can lead to increased investor confidence, potentially lowering the cost of capital and facilitating access to new markets and business opportunities.

5.4. Mechanism Analysis

5.4.1. Mediation Effect Test

The baseline regression results confirm the study’s central hypothesis that generative artificial intelligence (GAI) supports the enhancement of corporate ESG performance. To assess the mediating influence of the quality of disclosure and sustainable innovation in the impact of GAI on corporate ESG performance, this paper incorporates information disclosure quality and sustainable innovation as mediating variables into the model (1) to evaluate the consequences of the pathways “GAI—quality of information disclosure—corporate ESG performance” and “GAI—sustainable innovation—corporate ESG performance.” The specifics are as follows:

Information Disclosure Quality. In column (1) of

Table 7. Results of the Mediation Effect Test.

Variables	(1) IQ	(2) ESG	(3) SI	(4) ESG
GAI	0.0308 ** (2.38)	0.0426 *** (3.33)	0.0442 ** (2.12)	0.0436 *** (3.38)
IQ		0.108 *** (15.59)
SI				0.0531 *** (−10.66)
_cons	−8.385 *** (−68.96)	−0.190 (−1.35)	−12.08 *** (−61.93)	−0.450 *** (−3.18)
Control	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
N	23,631	23,631	23,631	23,631
R2	0.388	0.202	0.480	0.198

Note: ** and *** denotes significance levels of 5% and 1%, respectively.

, the estimated coefficient for GAI on IQ is positive and statistically significant, indicating that after applying GAI, firms can more efficiently integrate and disclose information, improving the comprehensiveness and precision of disclosed information. As shown in column (2), the coefficient for IQ’s effect on ESG is also significantly positive, suggesting that informational clarity significantly improves corporate ESG performance. This result further supports hypothesis H2 regarding the mechanism of technology-enabled information disclosure, indicating that informational clarity contributes significantly to the enhancement of corporate ESG performance by GAI. Therefore, by enhancing the quality of information disclosure, generative AI can reduce information asymmetry in the market, attract more investors who prioritize transparency and sustainability, bring in more investments for enterprises, and support value-creating activities such as business expansion and R&D. Meanwhile, with access to comprehensive and accurate information, enterprises can make wise decisions on ESG issues, mitigate potential risks, and safeguard long-term financial stability and profitability.

Sustainable Innovation. As shown in column (3) of Table 7, the coefficient for GAI’s effect on SI is positive and statistically significant, which implies that after applying GAI, firms’ sustainable innovation capabilities are enhanced. Similarly, the regression result in column (4) shows a significantly positive effect of SI on ESG, suggesting that sustainable innovation significantly improves corporate ESG performance. This result further supports hypothesis H3 regarding the mechanism of technology-enabled sustainable innovation, indicating that sustainable innovation contributes substantially to the improvement of corporate ESG performance facilitated by GAI. This means that by leveraging generative AI to enhance sustainable innovation capabilities, enterprises are enabled to develop more eco-friendly products and services. This helps enterprises meet the ever-increasing demand from consumers and regulatory bodies for sustainable solutions. By adopting sustainable practices and technologies, enterprises can reduce resource consumption, waste generation, and energy use, thereby cutting operational costs. This contributes to improving corporate ESG performance and, in turn, helps build a more sustainable and responsible business ecosystem.

To further verify the transmission mechanisms through which generative artificial intelligence (GAI) affects corporate ESG performance, this paper employs the Bootstrap method to conduct 1000 repeated sampling tests to examine the mediating roles of information disclosure quality (IQ) and sustainable innovation (SI). The Bootstrap test results in

Table 8. Bootstrap test results.

	Conditional Indirect Effect	z	P > |z|	[95% Confidence Interval]
GAI → IQ → ESG	0.0027	1.98	0.048	[0.0002,0.0054]
GAI → SI → ESG	0.0069	5.84	0.000	[0.0046,0.0093]

indicate that the mediating effect of information disclosure quality (IQ) in the relationship between generative artificial intelligence and corporate ESG performance is significant (β = 0.0027, p < 0.05); similarly, the mediating effect of sustainable innovation (Patent) in this relationship is also significant (β = 0.0069, p < 0.001). These findings complement the conclusions drawn from the three-step mediation regression analysis presented earlier and further validate Hypotheses H2 and H3.

5.4.2. Moderating Effect Test

In this paper, both the core variables—generative artificial intelligence (GAI) and environmental regulation (ER)—which are included in the interaction term, have undergone Z-score standardization. Subsequently, interaction terms are constructed based on these standardized variables, and regression estimation is performed on the moderating effect model. The results presented in

Table 9. Results of the Moderating Effect Test.

	ESG
GAI_ER	−0.025 *** (−4.43)
_cons	−1.090 *** (−8.55)
Control	Yes
Year	Yes
Industry	Yes
N	23,631
R2	0.195

Note: *** denote significance levels of 1%.

indicate that the coefficient of the interaction term (GAI_ER_std) is −0.0250, which is significantly negative at the 1% level. This suggests that environmental regulation significantly weakens the positive impact of generative artificial intelligence on ESG performance, implying that stringent environmental regulations can inhibit the direct empowering effect of generative artificial intelligence. This may be because stringent environmental regulations have increased corporate compliance costs and operational pressures. To meet regulatory requirements, companies need to invest substantial resources in areas such as the construction of environmental protection facilities, pollution control, and environmental monitoring. These investments may crowd out the resources that would have otherwise been allocated to ESG innovation leveraging generative AI. For instance, a company might have to divert funds originally earmarked for researching and developing green production technologies based on generative AI to purchase costly environmental protection equipment to meet emission standards.

5.5. Robustness Test

5.5.1. Sub-Sample Test

Generative artificial intelligence entered the stage of large-scale application after 2022. This paper further restricts the research period to 2022–2024 and excludes samples from before the technological application to conduct a sub-sample test. The test results are shown in column (1) of

Table 10. Robustness Tests.

	(1) ESG	(2) ESG	(3) F_ESG	(4) ESG	(5) ESG
GAI	0.0291 * (1.82)		0.0383 ** (2.39)	0.0469 *** (3.31)	0.033 ** (2.14)
L_GAI		0.0407 ** (2.52)
Size	0.316 *** (27.71)	0.263 *** (37.15)	0.268 *** (37.93)	0.266 *** (38.90)	0.257 *** (35.57)
Age	−0.169 *** (−11.67)	−0.124 *** (−12.12)	−0.101 *** (−11.31)	−0.139 *** (−16.07)	−0.151 *** (−16.19)
DAR	−0.995 *** (−14.91)	−0.918 *** (−20.90)	−0.848 *** (−19.33)	−0.891 *** (−21.56)	−0.844 *** (−18.47)
ROA	0.620 *** (3.38)	1.064 *** (8.26)	2.034 *** (14.94)	1.165 *** (9.45)	1.518 *** (10.14)
PAR	0.450 *** (4.11)	−0.0445 (−0.62)	0.0081 (0.11)	−0.0057 (−0.08)	0.0734 (1.03)
INV	0.0333 *** (3.41)	0.0007 (0.13)	−0.0004 (−0.08)	−0.0044 (−0.87)	0.003 (0.55)
TQR	0.106 (0.99)	0.191 *** (2.89)	0.161 ** (2.47)	0.159 ** (2.53)	0.259 *** (3.93)
PII	0.0551 (1.32)	0.0321 (1.19)	0.0199 (0.67)	−0.0289 (−1.14)	0.0015 (0.06)
Growth					−0.048 *** (−9.63)
Dual					−0.0094 (−0.73)
_cons	−2.197 *** (−9.31)	−1.117 *** (−7.64)	−1.353 *** (−9.28)	−1.133 *** (−8.01)	−1.076 *** (−7.21)
Year	Yes	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes	Yes
N	4756	18,028	18,031	18,974	17,028
R2	0.331	0.196	0.205	0.194	0.194

Note: *, ** and *** denote significance levels of 10%, 5% and 1%, respectively.

, where the coefficient of GAI is significantly positive at the 10% level. Despite the limitations of a shorter time span and a reduced sample size in the sub-sample, which have led to a decrease in significance, the direction of the coefficient remains entirely consistent with that of the full sample, and the economic impact remains substantial.

The relatively low GAI values observed in the period before 2021 can be attributed to the nascent stage of corporate research and development (R&D) in the field of generative algorithms. During this early phase, companies were still in the process of exploring and laying the groundwork for GAI technologies, which led to lower levels of GAI implementation and thus lower GAI values. Including these years in the general panel is crucial as it allows us to capture the full evolution and development trajectory of GAI adoption within the corporate sector, providing a more comprehensive understanding of the long-term trends and impacts.

5.5.2. Lagged and Leading Terms

This paper conducts robustness tests by extending the time horizon over which generative AI affects corporate ESG performance. Column (2) in Table 10 applies a one-period lag to the explanatory variable (GAI) (L_GAI), while column (3) applies a one-period lead to the dependent variable (ESG) (F_ESG). The analysis reveals that the estimated coefficients are 0.0407 and 0.0383, each positive and statistically significant at the 5 percent level, indicating that generative AI significantly promotes corporate ESG performance.

5.5.3. Randomly Changing Sample Size

Changes in sample size can affect regression results. To test the reliability of the regression findings under different sample sizes, this paper randomly selects 80% of the sample observations for regression. As presented in Table 10’s column (4), the results indicate that the findings remain valid after randomly changing the sample size.

5.5.4. Adding Control Variables

Corporate growth potential may influence management decision-making and information disclosure motives, while a dual leadership structure (in cases where a single person serves as both chairman and general manager) can alter the degree of power balance, both of which may interfere with core indicators such as corporate performance. Based on this, this paper adds the following explanatory factors as controls: Growth: calculated as the natural logarithm of sales revenue growth, reflecting the potential effect of business expansion and development potential on the research relationship; Dual: an indicator variable set to 1 when the roles of chairman and general manager are combined, and 0 if they are separate, capturing interference arising from differences in governance structure. Column (5) of Table 10 shows that the inclusion of these control variables does not alter the support for the research hypothesis, thereby further solidifying the conclusions.

5.5.5. Instrumental Variable Method

In examining the influence of generative AI adoption on corporate ESG outcomes, there may be unobservable factors at the firm level that could lead to omitted variable bias or reverse causality issues. To address this possible endogeneity concern, this paper employs the IV (instrumental variable) method, referring to the approaches and ideas of Lewbel (1997) [56] and Zhou et al. (2017) [57]. Firstly, we calculate the annual average level of generative AI, excluding the firm itself, and use its lagged value as the instrumental variable IV1. The lagged design helps avoid the issue where the current industry average of AI directly affects corporate ESG performance, thereby satisfying the exogeneity constraint of the instrumental variable. Subsequently, we use the cube of the difference between the firm’s generative AI level and the mean generative AI level categorized by industry and province as the instrumental variable IV2. After these treatments, the generated instrumental variables IV1 and IV2 meet the requirements of relevance and exogeneity.

As shown in columns (1) and (3) of

Table 11. Endogeneity Treatment.

	(1) GAI	(2) ESG	(3) GAI	(4) ESG
IV1	0.826 *** (0.056)
IV2			0.469 *** (0.014)
GAI		0.383 *** (0.066)		0.118 ***(0.030)
Control	Yes	Yes	Yes	Yes
_cons	−0.532 *** (0.074)		−0.206 *** (0.058)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
N	17,978	17,389	17,978	17,389
R2	0.203		0.494
Kleibergen-Paap rk LM statistic		160.105 ***		267.272 ***
Kleibergen-Paaprk Wald F statistic		225.479		763.167

Note: *** denote significance levels of 1%.

, the instrumental variables exhibit strong correlation with the core explanatory variable. Columns (2) and (4) indicate that the Kleibergen-Paap rk LM statistics are 160.105 and 267.272, respectively, both significant at the 1% level. The Kleibergen-Paap rk Wald F statistics are 225.479 and 763.167, respectively, passing the tests for unidentifiability and weak instrumental variables, thus confirming the identifiability of the instrumental variables. Based on the aforementioned data, the instrumental variables selected in this paper are reliable and reasonable. The estimated coefficients of both IV1 and IV2 are significantly positive, further demonstrating the robustness of the conclusions drawn in this paper.

5.5.6. Difference-in-Differences (DID) Method

Given that the explosive application of generative AI in 2022 exhibits characteristics of an exogenous shock, this paper uses 2022 as the policy shock point to construct a Difference-in-Differences (DID) model to precisely identify the causal effect of GAI on corporate ESG performance. Here, the treatment group is defined as firms that mentioned GAI during the sample period (GAI > 0), while the control group consists of firms that did not mention it. The time dummy variable Post is defined as the years from 2022 onwards, and the core interaction term TreatPost captures the net effect of the policy shock.

The regression results, as presented in column (1) of

Table 12. Analysis results of the Difference-in-Differences (DID) method.

	(1) ESG	(2) ESG
GAI		−0.0060 (−0.29)
Treat	−0.0289 (−1.46)
TreatPost	0.155 *** (5.36)
Size	0.265 *** (42.99)	0.253 *** (35.31)
Age	−0.143 *** (−18.30)	−0.137 *** (−15.00)
DAR	−0.898 *** (−24.10)	−0.847 *** (−19.47)
ROA	1.171 *** (10.51)	1.438 *** (10.79)
PAR	0.0127 (0.21)	−0.0756 (−1.06)
INV	−0.0061 (−1.34)	−0.0135 *** (−2.64)
TQR	0.203 *** (3.61)	0.190 *** (2.97)
PII	−0.0187 (−0.82)	−0.0495 * (1.87)
_cons	−2.197 *** (−9.31)	−0.869 *** (−5.86)
Year	Yes	Yes
Industry	Yes	Yes
N	23,631	18,874
R2	0.331	0.178

Note: * and *** denote significance levels of 10% and 1%, respectively.

, indicate that the coefficient of the core interaction term TreatPost is 0.155 and is significantly positive at the 1% level. This demonstrates that, taking the technological breakthrough in 2022 as a demarcation point, firms adopting GAI experienced a significantly greater improvement in ESG performance compared to those that did not adopt it, confirming the presence of a significant positive causal enabling effect of GAI on ESG performance.

Moreover, the event study approach reveals that before the explosive application of generative AI technology in 2022, the dynamic effect coefficients for most years are not significant, with only short-term fluctuations observed in 2013 (−0.171) and 2018 (−0.195 **). Overall, this supports the parallel trend hypothesis. After the shock in 2022 (0.0804), the dynamic effect coefficients turned from negative to positive and became significantly positive in 2023 (0.0919 *), indicating that the enabling effect of GAI on ESG performance gradually emerged after the technology’s large-scale application, consistent with the conclusions of the baseline regression.

5.5.7. Placebo Test

This paper restricts the research period to 2012–2021, prior to the shock occurrence, and forcibly assumes the presence of a GAI shock effect during this timeframe. Column (2) of Table 12 presents the results of the placebo test, with the regression showing that the coefficient of the core variable is −0.060 and not statistically significant. This indicates that, before the actual technological transformation took place, there were no significant pre-existing differences in ESG performance between the treatment and control groups. This result effectively rules out the possibility that the core conclusions are driven by unobservable factors or pre-existing trends.

6. Heterogeneity Analysis

From the perspectives of firm characteristics in generative AI application and industry characteristics of ESG performance, this study investigates the heterogeneous effects of generative AI on corporate ESG performance from three perspectives: whether the company is in the manufacturing field, geographical location, and whether it is a technology-intensive enterprise.

6.1. Whether the Firm Belongs to the Manufacturing Industry

Due to significant differences in production modes, resource consumption intensity, and environmental responsibility boundaries across industries, these factors influence the mechanism and effectiveness of generative AI on corporate ESG performance. Accordingly, the sample is segmented into manufacturing and non-manufacturing firms for separate regression analyses. Table 11 reveals that for manufacturing firms, the coefficient of generative AI on ESG performance is 0.074, significant at the 1% level, whereas the coefficient for non-manufacturing firms does not reach statistical significance. This demonstrates that generative AI demonstrates a more pronounced promoting effect on ESG performance in manufacturing firms than in non-manufacturing firms. On one hand, manufacturing firms’ production processes are highly reliant on physical manufacturing, supply chain management, and energy consumption control. Generative AI can enhance ESG performance by optimizing production scheduling, simulating green process improvement paths, accurately predicting supply chain carbon footprints, and being applied in scenarios such as production safety compliance and employee skill training program generation. On the other hand, non-manufacturing firms’ use of generative AI may primarily manifest in leveraging intelligent risk control models to reduce compliance risks and generating standardized disclosure texts, which leads to a relatively weak impetus for corporate ESG performance.

6.2. Geographical Location

Firms located in different provinces have varying access to resources, local cultures, and government policies. In economically less developed regions, firms may pay less attention to ESG, referring to existing research by Shen et al. (2021) [58]. Based on geographic location, the sample provinces are classified into three regions: eastern, western, and central. For example, a dummy variable takes the value 1 for firms in the eastern region and 0 for firms in the western and central regions. The results reported in Table 11 indicate a significantly positive coefficient for generative AI in the eastern region (1% level), while the estimates for the other regions are not significant. This suggests that the positive influence of generative AI on ESG performance is confined to firms in the eastern region, with no observable effect in the western and central regions. The causes of this disparity can be explained from two perspectives: resource endowment and development priorities. Economically, the eastern region is more developed, with firms having more sophisticated digital infrastructures and facing relatively higher regulatory requirements and market attention regarding ESG disclosure. Generative AI can leverage its efficient information processing and compliance risk warning capabilities to meet the refined needs of eastern firms in ESG practices, thereby effectively enhancing ESG performance. Conversely, companies located in the central and western areas, with comparatively lower levels of economic development, may focus more on the basic expansion of production and operations, prioritizing ESG investment and digital transformation at a relatively lower level. Consequently, the application scenarios and supporting resources for generative AI are insufficient, and thus, no significant enabling effect on ESG performance has been observed yet.

6.3. Whether the Firm Belongs to a Technology-Intensive Industry

In accordance with the 2012 industry classification standards of the China Securities Regulatory Commission, all sample industries are categorized based on the intensity of production factors. The sub-industry codes for technology-intensive industries are as follows: N77, C36, M74, I65, C33, C35, C27, C29, C39, C38, C37, C41, and C40. Firms with higher R&D investment may adopt a more open and proactive attitude towards AI applications and allocate a higher proportion of their investment to AI compared to firms with lower R&D investment. Therefore, drawing on Lu’s [59] research approach, this study categorizes sample firms into technology-intensive and non-technology-intensive groups based on the ratio of fixed assets and the proportion of R&D expenditure to compensation, assigning values of 1 and 0, respectively. From the classification results, it is observed that there are more non-technology-intensive firms among samples of listed companies. As presented in

Table 13. Heterogeneity Results.

Variables	ESG		ESG			ESG
	Manufacturing Industry	Non-Manufacturing Industry	Western Region	Eastern Region	Central Region	Technology-Intensive Industry	Non-Technology-Intensive Industry
GAI	0.074 *** (4.08)	0.014 (0.76)	−0.0078 (−0.14)	0.063 *** (4.53)	−0.038 (−0.96)	0.051 *** (3.37)	0.019 (0.83)
_cons	−0.703 *** (−4.31)	−1.684 *** (−7.95)	−1.110 *** (−2.80)	−0.904 *** (−6.17)	−0.648 *** (−1.65)	−0.830 *** (−4.30)	−1.258 *** (−7.30)
Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	15629	7945	2747	17011	3701	10956	12612
R2	0.154	0.256	0.244	0.194	0.224	0.180	0.212

Note: *** denote significance levels of 1%.

, the regression findings reveal that generative AI exerts a notably positive influence on ESG performance at the 1% significance level for technology-intensive enterprises, whereas this effect is insignificant for non-technology-intensive ones. This disparity might stem from the reality that technology-intensive firms inherently possess strong R&D capabilities and a digital adaptation foundation, enabling them to have a lower threshold for applying generative AI and achieve higher integration efficiency. They can leverage its intelligent analysis capabilities to more precisely embed ESG concepts into core areas such as technological innovation and talent management, thereby directly translating into improvements in ESG performance. In contrast, non-technology-intensive firms have business models that are less reliant on technological tools and have relatively limited resource investments in digital transformation and ESG management. Consequently, the application scenarios for generative AI have not been fully expanded, and thus, no significant driving effect on ESG performance has been observed yet.

7. Conclusions and Implications

7.1. Discussion and Conclusions

This research constructs an impact mechanism model of generative artificial intelligence (GAI) propelling corporate ESG performance and conducts an empirical analysis using panel data from Chinese A-share listed companies from 2012 to 2024. The following are the key findings and in-depth discussions on the underlying reasons and potential leveraging strategies:

7.1.1. GAI and Corporate ESG Performance Enhancement

GAI significantly contributes to enhancing corporate ESG performance, demonstrating positive effects across environmental, social, and governance dimensions. The reasons behind this are multifaceted. In the environmental realm, GAI’s ability to conduct real-time data analysis and intelligent decision-making is pivotal. It can predict energy consumption patterns with high accuracy, enabling enterprises to implement energy-saving measures proactively. For instance, in a manufacturing plant, GAI can analyze historical energy usage data and external factors like weather conditions to optimize production schedules, thereby reducing energy waste and lowering the environmental footprint. From a social perspective, GAI-powered chatbots and customer service systems revolutionize the way enterprises interact with customers and employees. These intelligent systems can handle a large volume of inquiries simultaneously, providing quick and accurate responses. This not only improves customer satisfaction but also enhances the working experience of customer service staff, as they can focus on more complex and value-added tasks. In terms of governance, GAI plays a crucial role in risk assessment and decision-making processes. It can analyze vast amounts of internal and external data to identify potential risks, such as financial fraud or regulatory non-compliance. By providing data-driven insights, GAI helps enterprises make more informed decisions, improving the overall governance structure and transparency. To leverage this positive impact, enterprises should actively invest in GAI technologies, allocate sufficient resources for employee training to ensure effective utilization, and collaborate with GAI technology providers to customize solutions according to their specific ESG goals.

7.1.2. The Mediating Role of Information Disclosure Quality

The quality of information disclosure serves as a partial mediating factor in the relationship between GAI adoption and improvements in corporate ESG performance. Generative AI technologies improve the accuracy of financial data, reduce human errors, and promote transparency, thus improving the caliber of financial reporting and, consequently, corporate ESG performance. High-quality information disclosure is of utmost importance for stakeholders to accurately assess a company’s ESG performance. In the absence of reliable information, stakeholders may make incorrect judgments, leading to sub-optimal investment decisions or reputational damage for the company. GAI can process large volumes of data at an unprecedented speed and with high accuracy. It can automatically detect anomalies in financial data, such as unusual transactions or incorrect accounting entries, and correct them promptly. This reduces the risk of misinformation and ensures that the disclosed ESG-related information is trustworthy. Moreover, GAI can generate more detailed and user-friendly ESG reports. It can analyze complex data and present it in a clear and concise manner, making it easier for stakeholders to understand the company’s ESG performance. To leverage this mediating effect, enterprises should establish strict data management and information disclosure systems integrated with GAI. Regular audits should be conducted to ensure the effectiveness of these systems. Additionally, they can use GAI to engage with stakeholders more effectively, for example, by providing real-time updates on ESG performance through interactive platforms.

7.1.3. The Mediating Role of Sustainable Innovation

Sustainable innovation partially mediates the correlation between generative AI and the enhancement of corporate ESG outcomes. Generative AI enables in-depth mining and examination of extensive volumes of literature and experimental data, reducing uncertainties and R&D costs associated with sustainable innovation technologies through predictive assessments of technological environmental impacts and optimization of technological response pathways. This elevates the efficiency of corporate sustainable innovation, supplying assistance to improvements in ESG performance. Sustainable innovation is the cornerstone of long-term ESG performance improvement. In today’s competitive business environment, enterprises need to continuously develop new products, processes, and business models that are environmentally friendly, socially responsible, and economically viable. GAI can significantly accelerate the sustainable innovation process. By analyzing large datasets, including market trends, consumer preferences, and scientific research, GAI can provide valuable insights for enterprises. For example, it can predict the environmental impacts of new technologies in the early stages of R&D, allowing companies to make adjustments to minimize negative effects. Moreover, GAI can optimize the technological response pathways, reducing the time and cost required for sustainable innovation. To leverage this mediating role, enterprises should allocate more resources to R&D departments for the integration of GAI and sustainable innovation. They can establish cross-functional teams that include GAI experts, sustainability specialists, and R&D personnel to foster collaboration and innovation. Additionally, enterprises can establish partnerships with research institutions and universities to access the latest GAI technologies and sustainable innovation ideas.

7.1.4. The Moderating Effect of Environmental Regulations

Environmental regulations weaken the favorable effect of generative artificial intelligence on corporate ESG outcomes. Stringent environmental regulations require enterprises to increase investments in areas such as acquiring equipment for environmental safeguarding, implementing pollution mitigation strategies, and environmental monitoring. This will undoubtedly raise enterprises’ operational costs and, consequently, crowd out resource allocation toward generative artificial intelligence within these enterprises. Environmental regulations are designed to protect the environment and promote sustainable development. However, they can also have unintended consequences on enterprises’ investment decisions. When faced with strict environmental regulations, enterprises are often required to make significant upfront investments in environmental protection measures. For example, a manufacturing company may need to install expensive pollution control equipment to comply with emission standards. These investments can strain the company’s financial resources, leaving less budget for GAI-related projects. Moreover, the ongoing costs of operating and maintaining environmental protection equipment can also be substantial, further reducing the funds available for GAI investment. To mitigate this moderating effect, governments can provide incentives to encourage enterprises to invest in GAI while complying with environmental regulations. Tax breaks, subsidies, or low-interest loans can be offered to enterprises that adopt GAI technologies for environmental management. Enterprises, on the other hand, can explore cost-effective GAI solutions that can also help them meet environmental requirements. For example, they can use GAI for energy management to reduce both costs and environmental impact, achieving a win-win situation.

7.1.5. Heterogeneous Effects Across Different Types of Firms

Compared to non-manufacturing firms, firms established in the central and western territories, and non-technology-intensive firms, generative AI exerts a more pronounced enhancing effect on ESG performance in manufacturing firms, firms in the eastern region, and technology-intensive firms. The reasons for these heterogeneous effects are rooted in the characteristics of different types of firms. Manufacturing firms typically have more complex production processes and a larger environmental footprint compared to non-manufacturing firms. The application of GAI in manufacturing can bring about significant improvements in environmental performance. For example, GAI can optimize supply chain management in manufacturing, reducing waste and energy consumption through real-time monitoring and intelligent decision-making. Firms in the eastern region usually have better access to technology, talent, and capital. These advantages facilitate the adoption and effective use of GAI. The eastern region is often more developed economically, with a higher concentration of high-tech companies and research institutions. This creates a favorable environment for enterprises to access the latest GAI technologies and attract skilled professionals. Additionally, firms in the eastern region face more intense market competition, driving them to use GAI to improve their ESG performance for a competitive edge. Technology-intensive firms are more likely to have the technical capabilities and infrastructure to integrate GAI into their operations. They are also more familiar with the potential of emerging technologies and are more willing to invest in them. To address these heterogeneous effects, non-manufacturing firms, firms in the central and western regions, and non-technology-intensive firms should learn from the successful experiences of their counterparts. They can seek partnerships with technology companies or universities to enhance their GAI capabilities. Governments can also provide targeted support, such as technology transfer programs and training initiatives, to help these firms catch up.

In conclusion, this study not only presents empirical evidence on the relationship between GAI and corporate ESG performance but also provides in-depth insights into the underlying mechanisms and heterogeneous effects. These findings can offer valuable guidance for enterprises in formulating GAI-based ESG strategies and for policymakers in designing relevant policies. However, it should be noted that these conclusions are based on the current empirical analysis, and further research is needed to explore the long-term and dynamic effects of GAI on corporate ESG performance under different economic and regulatory environments.

7.2. Policy Implications

This study has constructed the driving mechanism of generative AI on corporate ESG performance, yielding the following practical implications:

First, enterprises should formulate an “AI + ESG” integrated strategy, deeply embedding generative AI into their internal information governance systems to leverage technology in resolving information asymmetry challenges during green transformation. This study confirms that the quality of information disclosure serves as a crucial transmission pathway through which generative AI enhances corporate ESG performance. Enterprises should not merely regard AI as a tool for cost reduction and efficiency enhancement but should incorporate it into the core framework of ESG governance. In practice, enterprises need to proactively establish an ESG data hub based on generative AI, utilizing its natural language processing capabilities to automatically clean, integrate, and validate the compliance of unstructured data scattered across various departments (such as supply chain communication records, environmental monitoring logs during production processes, and employee safety reports). This will result in the formation of a real-time, comparable, and traceable ESG information repository. This not only significantly enhances the transparency and completeness of external disclosures, reducing regulatory and reputational risks arising from “greenwashing” behaviors but, more importantly, high-quality internal information supports management in accurately identifying weaknesses in environmental and social dimensions, providing a decision-making basis for subsequent targeted investments and rectifications, thereby achieving a transition from passive compliance to proactive governance.

Second, enterprises should leverage generative AI to reconstruct their R&D paradigms, driving sustainable innovation through “AI-assisted invention” to transform technological dividends into long-term ESG competitive advantages. This study indicates that sustainable innovation is another core mediator through which generative AI exerts its effects. Enterprises should move beyond superficial applications of AI that replace repetitive labor and instead utilize its knowledge generation and simulation capabilities to overcome green technology bottlenecks. Specifically, R&D departments can employ generative AI for rapid screening of new material formulations, simulation optimization of low-carbon production processes, and scenario construction for circular economy models. For instance, during the product design phase, AI can generate alternative solutions that combine high performance with low environmental impact based on extensive chemical and material databases; in production and operations, AI can analyze energy consumption data in real-time and generate dynamic optimization instructions. This empowerment enables enterprises to accelerate green technology iteration with lower trial-and-error costs, swiftly respond to market demand for green products, and ultimately establish differentiated ESG performance that is difficult to imitate, with technological innovation serving as a barrier.

Third, industry organizations should establish an open-source green AI sharing platform to alleviate the dual resource constraints of “compliance-innovation” faced by small and medium-sized enterprises (SMEs) through knowledge spillover. The negative moderating effect of environmental regulations may be more pronounced among SMEs, as they lack sufficient resources to meet increasingly stringent compliance requirements and cannot independently invest in building advanced AI-empowered systems. To alleviate this structural contradiction, industry leaders and industrial alliances should play a synergistic role in taking the lead in constructing a vertical-sector-specific open-source green AI platform. This platform could include open-source sharing of industry-specific carbon emission factor databases, the accumulation of knowledge bases for validated low-carbon technology solutions, and the low-cost provision of compliance self-check tools based on generative AI. Through such industry-level knowledge spillover, SMEs can acquire AI empowerment capabilities at extremely low costs, rapidly enhancing their information disclosure quality and innovation efficiency, thereby achieving ESG performance catch-up rather than falling behind in a heavily regulated environment.

Fourth, policymakers should be vigilant against the negative impacts of “regulatory overload” and establish a flexible regulatory system compatible with the characteristics of AI technology. The empirical results of the negative moderating effect of environmental regulations provide an important warning for policy design: overly rigid or high-frequency compliance requirements may leave enterprises too busy coping with regulatory inspections to utilize cutting-edge technologies for fundamental green transformations. Therefore, regulatory authorities need to optimize the design logic of regulatory tools. Specifically, “technology-adaptive clauses” can be introduced into environmental policies, offering incentives such as reduced compliance frequencies, data reporting exemptions, or tax deductions to enterprises that actively adopt generative AI to achieve carbon reduction, resource recycling, or supply chain transparency. Simultaneously, exploring the establishment of a precise regulatory model based on AI dynamic monitoring to reduce compliance costs for enterprises using technological means is essential. This flexible approach, which combines regulatory objectives with technological empowerment, can transform external pressures into internal motivations, avoiding the crowding-out effect of regulations on innovation.

Fifth, in practical terms, given the crucial role of the “Governance” (G) component within the ESG framework, it is imperative to establish a robust regulatory framework for the ethical use of generative AI in corporate reporting. Regulators should take the lead in formulating and implementing comprehensive guidelines that specifically address the ethical challenges associated with generative AI-powered reporting. Firstly, strict data privacy protection regulations must be enforced as generative AI relies heavily on vast amounts of data for training and generating reports, making the confidentiality and security of sensitive corporate and stakeholder data of utmost importance; regulators should set clear standards for data collection, storage, and usage to prevent any unauthorized access or misuse. Secondly, algorithm transparency should be a key focus of the regulatory framework since the inner workings of generative AI algorithms are often complex and opaque, leading to concerns about bias, discrimination, and lack of accountability; regulators should require companies to disclose relevant information about the algorithms used in report generation, such as the data sources, decision-making processes, and potential biases, enabling stakeholders to better understand and evaluate the reliability and fairness of the reports. Moreover, to ensure fairness in decision-making processes driven by generative AI, regulators should establish ethical review mechanisms involving independent experts who assess whether the generative AI-generated reports and related decisions comply with ethical principles, such as non-discrimination, equality, and social responsibility; companies should be required to submit their generative AI-based reporting practices for regular ethical reviews, and any non-compliant behavior should be subject to penalties. By implementing these specific regulatory measures for the ethical use of generative AI in reporting, we can not only enhance the credibility and integrity of corporate ESG reporting but also promote the sustainable and responsible development of the corporate sector in the era of generative AI.

7.3. Limitations and Future Research

This investigation also has particular limitations: Firstly, it only includes enterprises in China as the sample. However, the path through which technology drives ESG performance may diverge due to varying institutional backgrounds across different countries or regions. Future research could establish a global sample database, integrating enterprise data from areas including China, the European Union, the United States and Southeast Asia, to analyze the differentiated impacts of GAI (Generative Artificial Intelligence) technology under different institutional environments (e.g., the EU’s carbon border tax and the US’s ESG disclosure regulations) and to refine a universal theoretical framework. Secondly, this research fails to differentiate between the effectiveness of generative AI and traditional AI in ESG management, which may obscure the following key issues: Traditional AI is more efficient in processing structured data, while generative AI excels at integrating unstructured information, and these differences have not been revealed. Meanwhile, the deployment cost of generative AI is markedly higher compared to traditional AI, which may restrict its application within small and medium-sized enterprises and lead to an uneven distribution of technological dividends. Future research could select Chinese enterprises within the same industry, deploy generative AI and traditional AI in separate groups, and compare the differences in the advancement of enterprises’ ESG performance using the Difference-in-Differences (DID) method. Thirdly, the measurement of some variables relies on proxy indicators, which may introduce bias. Drawing on the research by Shi and Mai (2025) [59], we take the natural logarithm of the number of analysts tracking listed companies plus one as a key indicator for evaluating the quality of corporate information disclosure. However, this approach may not accurately reflect the actual quality of information disclosure within companies. Future research should explore alternative measurement methods to mitigate proxy measurement bias. One promising direction is to conduct content analysis of corporate reports for semantic transparency. By analyzing the language, structure, and level of detail in the reports, researchers can obtain a more direct and in-depth understanding of the quality of information disclosure, thus validating the findings of the current study and providing a more comprehensive and accurate assessment of corporate information disclosure.