ESG Performance, Artificial Intelligence Adoption, and Knowledge Diversification: Pathways to Sustainable Innovation Capability

Abstract

Against the backdrop of the global shift from scale-driven competition to innovation-driven development, together with the dual trends of climate governance and digital transformation, whether firms can transform ESG performance into sustainable innovation capability has become a central question in assessing the real economic value of ESG practices. Using panel data of Chinese A-share listed companies from 2009 to 2025, this study constructs a two-way fixed-effects model to systematically examine the effects of overall ESG performance and its three dimensions, namely, Environmental (E), Social (S), and Governance (G), on corporate sustainable innovation capability. Furthermore, this study incorporates artificial intelligence (AI) adoption and knowledge diversification (TD) into a unified analytical framework and investigates their roles as parallel mediating mechanisms. The results show that overall ESG performance is positively related to corporate sustainable innovation capability. In the joint regression, all three ESG dimensions have positive effects, with the social dimension exhibiting the largest marginal effect. Artificial intelligence adoption and knowledge diversification constitute two parallel and complementary transmission channels through which ESG enhances sustainable innovation capability. Further analyses find that higher ESG levels are positively associated with corporate sustainable innovation capability, whereas rapid short-term improvements in ESG are negatively related to corporate sustainable innovation capability. It shows that innovation performance experiences short-term pressure before a substantial ESG improvement but gradually surpasses its previous level afterward, indicating the existence of genuine short-term adjustment costs in ESG-driven innovation. This study simultaneously incorporate artificial intelligence adoption and knowledge diversification into the mechanism analysis of the ESG–innovation relationship, providing more fine-grained micro-level evidence on how ESG is transformed into real productive capacity. The findings offer important implications for firms’ differentiated ESG investment strategies and the coordinated design of innovation and sustainability policies.

1. Introduction

Over the past decade, the fundamental logic of global industrial competition has undergone a structural transformation, shifting from growth driven by the expansion of factor inputs toward innovation-driven development centered on technological breakthroughs and sustainability-oriented capabilities. At the same time, two major trends that had previously evolved largely independently have increasingly converged. The first is the rapid rise in environmental, social, and governance (ESG) as a core framework for corporate sustainability in global capital markets. By 2025, sustainable and responsible investment had gained considerable prominence in global capital markets, with the latest Global Sustainable Investment Review documenting approximately US$16.7 trillion in fund assets managed under responsible-investment approaches (https://www.gsi-alliance.org/members-resources/gsir2024/?utm_source, accessed on 27 August 2026). The second is the emergence of a new wave of technological revolution driven by artificial intelligence (AI), which is reshaping corporate innovation processes. The widespread adoption of generative AI, large language models, and intelligent research and development (R&D) tools can enhance the efficiency of knowledge search, integration, and transformation. Consequently, the level of AI application has become a potentially important enabler of firms’ innovative capability.

The convergence of these two trends raises a theoretically significant and practically urgent question: can superior ESG performance contribute to corporate sustainable innovation capability? If so, through what mechanisms does this transformation occur? The answer is far from self-evident. On the one hand, the resource-based view [1] and stakeholder theory [2] provide theoretical foundations for the proposition that ESG enhances innovation. ESG-related practices may help firms develop strategically valuable resources, including environmental compliance capabilities, skilled human capital, and transparent governance structures, which can be translated into organizational capabilities, knowledge assets, and external legitimacy that support innovation. On the other hand, several arguments suggest that this relationship may not be straightforward. ESG-related investments require financial resources and managerial attention, potentially crowding out short-term R&D expenditures [3]. Moreover, improvements in ESG performance often depend on simultaneous enhancements in firms’ internal knowledge and technological capabilities, among which AI adoption and knowledge diversity may represent two critical yet insufficiently integrated transmission mechanisms. Furthermore, firms differ substantially in their ability to transform ESG-related resources into innovation outcomes. Such absorptive capacity may vary systematically across firms with different sizes, R&D intensities, and ownership structures. Failure to account for these differences may obscure important heterogeneity in the relationship between ESG performance and sustainable innovation capability. This “resource empowerment” perspective has received growing empirical support. Some researchers demonstrated that superior corporate social responsibility performance significantly reduces firms’ cost of capital [4], while others further showed that firms systematically integrating sustainability into their organizational processes and governance structures consistently outperform their peers in both long-term corporate performance and innovation outcomes [5]. These findings provide direct empirical support for the applicability of the resource-based view in the ESG context.

These inconsistent findings may partly reflect competing theoretical explanations of the relationship between ESG performance and innovation capability. The resource-based perspective argues that improvements in ESG performance may expand firms’ resource base and strengthen their external legitimacy. The seminal “Porter Hypothesis” was the first to suggest that stringent environmental regulations can partially or even fully offset compliance costs by stimulating innovation that may offset compliance costs [6]. Subsequent theoretical reviews and empirical evidence from the European manufacturing sector offer varying degrees of support for this hypothesis [7,8]. However, evidence supporting the “resource crowding-out” perspective is equally compelling. One study found that firms engaging in ESG decoupling where externally disclosed ESG ratings diverge from firms’ actual sustainability performance subsequently invest significantly less in R&D than their peers, suggesting that symbolic ESG compliance may divert financial resources and managerial attention away from substantive innovation activities [9]. Similarly, research examining the interaction between intellectual property governance and corporate social responsibility documented evidence of resource competition between these two types of corporate activities [10]. Differences in ESG measurement may also contribute to the mixed findings. A landmark study systematically documented substantial discrepancies in ESG ratings assigned to the same firms by major rating agencies such as MSCI and Sustainalytics, attributing these differences to systematic variations in rating scope, measurement methodologies, and weighting schemes [11]. Other studies further demonstrated that such ESG rating disagreements themselves contain information that is priced by capital markets [12,13].

Similarly, a study using data from Chinese A-share listed firms found that ESG rating disagreement is significantly associated with stock returns [14]. Meanwhile, two related strands of the literature, namely, the effects of AI adoption and knowledge diversity on innovation performance, exhibit a pattern of debate strikingly similar to that observed in the ESG–innovation relationship. Regarding AI, one study using data from U.S. firms found that companies actively developing AI capabilities achieve faster growth and generate more product innovations [15]. In contrast, studies based on German enterprise survey data showed that the positive effects of AI adoption on total factor productivity and innovation probability depend critically on complementary digital investments and organizational capabilities, rather than arising automatically from AI adoption itself [16,17]. These findings suggest that when examining the mediating role of AI application, the moderating effects of firms’ complementary capabilities should also be taken into account. Evidence from China further supports this view. One study confirmed that the impact of AI adoption on innovation exhibits substantial heterogeneity across industries and ownership types [18]. The literature on knowledge diversity presents a similarly divided picture. Early research found that technological diversity suggests that the innovation benefits of technological diversity depend partly on technological coherence, whereas expanding knowledge breadth alone may reduce innovation output because of rising coordination costs [19]. Another study reported that technological diversity significantly improves firms’ innovation capability, with a stronger effect on exploratory innovation than on exploitative innovation [20]. Using Chinese firm-level data, some researchers likewise demonstrated that both the direction and magnitude of the relationship between knowledge diversity and innovation vary substantially depending on measurement approaches and moderating conditions [21,22].

In summary, although previous studies have separately examined the relationship between ESG and corporate innovation capability, commonly assessed using patent-based indicators or R&D intensity as well as the separate roles of AI adoption and knowledge diversity on innovation capability, few studies have simultaneously incorporated AI adoption and knowledge diversity into a unified mediation framework to systematically identify the specific mechanisms through which ESG enhances sustainable innovation. This study aims to fill this important research gap. Nevertheless, to the best of our knowledge, no existing study has integrated ESG, artificial intelligence, and knowledge diversity into a unified multiple-mediation framework to systematically examine their joint transmission mechanisms.

Using Chinese A-share listed firms as the research sample offers a relevant empirical setting for this study for three reasons. To begin with, the sample provides a long observation window that captures the evolution of China’s ESG rating system. Covering the period from 2009 to 2025, the dataset spans the entire process through which China’s ESG evaluation system evolved from its early stage to increasing maturity. Since its establishment in 2009, the Huazheng ESG rating system has continuously expanded its coverage and refined its evaluation methodology. This extended observation period enables the present study to identify not only the cross-sectional relationship between ESG performance and corporate sustainable innovation capability but also the dynamic evolution of this relationship as the ESG rating system becomes more mature.

In addition, China provides a relevant setting for examining AI adoption in an evolving policy environment. Compared with firms in developed economies, Chinese firms have experienced a much steeper trajectory of AI adoption, particularly following the release of the New Generation Artificial Intelligence Development Plan by the State Council in 2017 (https://www.xinhuanet.com/politics/2017-07/20/c_1121353544.htm, accessed on 27 August 2026). Since then, substantial heterogeneity has emerged in firms’ AI application levels, providing firm-level variation for examining the potential mediating role of AI adoption in the relationship between ESG performance and innovation. Consistent with previous literature, using a sample of Chinese A-share listed firms, Gao and Feng [23] found that the productivity-enhancing effects of AI adoption vary significantly across ownership types and industries.

Lastly, the research is relevant to China’s innovation and sustainability policy objectives and addresses an issue of considerable practical importance. China is undergoing a critical transition from a factor-driven growth model toward an innovation-driven development strategy. The 14th Five-Year Plan and the Long-Range Objectives Through 2035 explicitly emphasize strengthening firms’ role as the principal driver of innovation (https://en.ndrc.gov.cn/policies/202203/P020220315511326748336.pdf, accessed on 27 August 2026). At the same time, reforms mandating ESG information disclosure are being progressively implemented. Consequently, there is an urgent need for rigorous micro-level evidence to determine whether the institutionalization of ESG genuinely supports China’s innovation-driven development strategy.

The marginal contributions of this study are threefold. First, this study develops a “dual-engine” transmission framework through which links ESG performance to sustainable innovation capability by simultaneously incorporating AI application (the technological capability dimension) and knowledge diversity (the knowledge structure dimension) into a unified mediation framework. This approach differs from prior studies that typically focus on a single transmission channel, such as patent output or R&D investment. Furthermore, this study finds that the innovation-enhancing effect of ESG depends critically on firms’ existing R&D intensity and environmental capability. By doing so, the study refines the boundary conditions of the resource-based view in explaining the ESG–innovation relationship. The proposed moderating mechanism is consistent with the absorptive-capacity effects identified in the contexts of digital transformation and regional collaborative innovation [24,25]. It also complements the recent findings of those who demonstrate that the benefits of knowledge diversity are contingent upon organizational context [26,27]. This study contributes to the ESG–innovation literature by jointly examining technological and knowledge-related pathways and exploring variation across firm capabilities.

Second, this study employs a two-way fixed-effects panel model to estimate the effect of ESG performance on corporate sustainable innovation capability. To ensure the credibility of the empirical findings, we implement a comprehensive set of endogeneity analysis and robustness strategies, including instrumental variable two-stage least squares (IV-2SLS), the propensity score matching method, double machine learning, Heckman two-step sample selection correction [28], mediation analysis using the test [29], and system GMM. Compared with the conventional robustness checks adopted in related studies, the methodological framework employed in this paper provides a more comprehensive cross-validation of the baseline results.

Third, this study systematically integrates firm-level AI adoption data and knowledge diversity measures with Huazheng ESG ratings and a corporate sustainable innovation capability within a unified analytical framework. Although the resulting matched sample is constrained by data availability, this data integration represents a valuable contribution to the existing micro-level research infrastructure and provides a replicable data-processing framework for future studies using larger samples. Moreover, motivated by recent evidence on ESG rating disagreement and its economic consequences (e.g., [30,31]), we further employ alternative estimation methods as robustness checks to mitigate potential measurement bias arising from reliance on a single ESG rating provider.

The remainder of this paper is organized as follows. Section 2 reviews the literature on ESG performance and corporate sustainable innovation capability, AI adoption, and knowledge diversity, develops the theoretical framework and presents the research hypotheses. Section 3 describes the research design, variable measurement, and descriptive statistics. Section 4 reports the baseline empirical results and robustness tests, and addresses endogeneity concerns. Section 5 examines the parallel mediating roles of AI adoption and knowledge diversity. Section 6 provides additional analyses. Finally, Section 7 concludes the paper and discusses its policy implications.

2. Literature Review

2.1. ESG Performance and Corporate Sustainable Innovation Capability

Empirical research on the relationship between ESG performance and corporate sustainable innovation capability has examined both the overall relationship and its potential underlying mechanisms. An important theoretical antecedent is the Porter hypothesis, Porter and van der Linde [6] primarily examined whether environmental regulation could offset compliance costs through an “innovation compensation effect.” They argued that stringent environmental standards may stimulate technological innovation, thereby partially or even fully compensating firms for the additional costs associated with environmental compliance. Subsequent empirical research has tested the applicability of this hypothesis under different institutional settings, yet the findings remain mixed. Some studies report a positive relationship between the stringency of environmental regulation and firms’ patent output [7,8], whereas others argue that the effect depends critically on the flexibility of regulatory design and firms’ pre-existing innovation capabilities. In particular, rigid command-and-control regulations may crowd out innovation activities [32].

As ESG ratings have become more widely available, research has expanded beyond environmental regulation to examine firms’ overall ESG performance. Using cross-country panel data, related studies documented a significant positive association between ESG ratings and both patent output and R&D intensity [33]. The underlying mechanisms are primarily attributed to the alleviation of financing constraints [4] and the enhancement of stakeholder support for corporate innovation [5]. In the Chinese context, some researchers further found that ESG performance has a stronger positive effect on green patents and substantive innovation than on strategic innovation, suggesting that the ESG–innovation relationship may vary across types of innovation [34,35].

More recent studies have refined these findings by examining ESG rating disagreement and contextual differences. Liu et al. [36] showed that the relationship between ESG ratings and corporate performance is moderated by disagreement among ESG rating agencies, with greater rating divergence weakening the positive association between ESG ratings and corporate performance. Rossi et al. [37] further proposed a geospatial information framework to address the methodological challenges posed by ESG rating disagreement. Regarding innovation measurement, Zhang et al. [38] distinguished between the quantity and quality of green patents and found that ESG performance is more strongly associated with green patent quality than with patent quantity. This finding is closely related to the central question addressed in this study concerning the transmission mechanisms through which ESG performance promotes corporate sustainable innovation capability. In addition, Wan et al. [39] explicitly examined the moderating role of absorptive capacity in the relationship between ESG and innovation efficiency and found that the positive effect is significant only for firms possessing strong internal knowledge integration capabilities. Complementing this resource-access perspective, Byun and Lee [40] found that ESG reputational risks are associated with tighter financial constraints and higher equity financing costs among Korean wholesale and retail firms. Furthermore, Qu and Zhang [41] showed that the value-enhancing effect of ESG varies across different stages of the corporate life cycle, with stronger marginal effects observed among growth-stage firms. Li and Chen [42] further reported that extreme climate events significantly affect the stability of ESG ratings, providing an external explanation for the year-to-year fluctuations in ESG scores. Further evidence highlights the joint roles of organizational, technological, and environmental conditions in sustaining green innovation. Using data from 48 Chinese listed textile firms, Yan et al. [43] combine necessary condition analysis and fuzzy-set qualitative comparative analysis to identify configurations associated with sustained green patenting, with no individual condition identified as necessary. Their findings offer a complementary perspective on how resources and capabilities jointly support innovation in the ESG context.

Building upon the resource-based view [1], stakeholder theory [2], and absorptive capacity theory [44], this study develops an integrated analytical framework to explain how ESG performance enhances corporate sustainable innovation capability. To start with, there is the resource-based mechanism. According to the resource-based view, sustainable competitive advantage arises from strategic resources that are valuable, rare, inimitable, and non-substitutable. Firms with superior ESG performance accumulate organizational capabilities in areas such as environmental management, human capital development, and corporate governance transparency. These capabilities are difficult for competitors to replicate in the short run and therefore constitute essential complementary assets for innovation activities. For example, a well-established environmental management system provides an operational foundation for green technology development, while effective corporate governance reduces agency conflicts and managerial short-termism in innovation-related decision-making. Related evidence from distribution research highlights the complementary role of technological capabilities. Bui et al. [45] examined green standards adoption, distribution channel innovation, and international business performance, showing how combinations of green standards implementation and technological capabilities contribute to superior performance outcomes.

Moreover, stakeholder relationships provide a potential channel linking ESG performance to innovation. Firms with strong ESG performance possess greater capabilities in managing relationships with diverse stakeholders, including employees, suppliers, regulators, and local communities [2]. Such capabilities are particularly important for innovation because sustained innovation relies heavily on knowledge exchange and collaboration across organizational and institutional boundaries. Higher employee satisfaction and engagement stimulate idea generation and intrapreneurship, close partnerships with suppliers and research institutions facilitate the acquisition of external knowledge, and constructive relationships with regulators and local communities may help firms navigate regulatory uncertainty and manage reputational risk associated with the commercialization of new technologies and products.

Finally, absorptive capacity may condition the relationship between ESG performance and innovation. According to absorptive capacity theory [44], firms’ ability to transform external knowledge and resources into innovation outputs fundamentally depends on their existing knowledge base and R&D investment intensity. Even when firms obtain comparable external resources such as improved financing conditions or enhanced reputational capital resulting from superior ESG performance—those with insufficient absorptive capacity may be unable to effectively convert these resources into innovation outcomes. This mechanism suggests that the innovation-enhancing effect of ESG is unlikely to be homogeneous across firms and instead depends critically on firms’ technological capabilities and R&D foundations. This prediction will be systematically examined in the heterogeneity analysis presented later in the paper. Drawing on these theoretical arguments, this study proposes the following overarching hypothesis:

Hypothesis H1.

ESG performance has a positive effect on corporate sustainable innovation capability.

2.2. Artificial Intelligence Adoption and Corporate Sustainable Innovation Capability

The role of AI adoption in corporate sustainable innovation has emerged as an important research topic in recent years. Theoretically, AI influences corporate sustainable innovation capability through three major pathways. First, AI may facilitate knowledge search and recombination. Machine learning and natural language processing technologies substantially reduce the cost of searching, identifying, and recombining knowledge from massive patent databases and the scientific literature, thereby accelerating the generation of recombinant innovation [46]. Second, the R&D process optimization pathway: AI-assisted experimental design, material screening, and simulation may improve R&D efficiency and shorten development cycles [47]. Third, the organizational decision-making pathway: AI enhances firms’ information-processing capabilities and resource allocation efficiency, enabling innovation resources to be allocated more precisely to high-value projects [15].

Empirical studies provide evidence of a positive association between AI adoption and several dimensions of corporate innovation. Babina et al. [15] using patent text and human capital data from U.S. firms, find that AI talent intensity is positively associated with patent output and product innovation, with the effect being particularly pronounced in traditional industries that adopted AI relatively late, indicating a clear catch-up effect. In the Chinese context, Qin et al. [48] and Zhao and Chen [49] show that AI adoption significantly improves total factor productivity and innovation output in manufacturing firms, primarily through optimizing R&D resource allocation and strengthening knowledge spillovers. Sullivan and Wamba [50] further extend AI adoption to supply chain management and demonstrate that AI-driven demand forecasting and resource allocation release organizational resources for exploratory innovation activities. Heo et al. [51] found that digital transformation is positively associated with product and process innovation performance among Korean manufacturing firms. This provides complementary evidence on digital technologies and innovation, without isolating the specific role of AI adoption.

The relationship between AI, ESG, and corporate sustainable innovation capability has also attracted increasing scholarly attention. Xie et al. [52] find that AI adoption is positively associated with firms’ ESG performance. This finding points to a potential reverse relationship between AI adoption and ESG performance proposed in this study and suggests that AI and ESG may reinforce each other beyond the mediation framework examined here. Wu and Zhu [53], based on data from Chinese listed firms, report that the productivity-enhancing effect of AI application exhibits significant industry heterogeneity, with stronger effects observed in knowledge-intensive industries. This pattern is compatible with a knowledge-based explanation, although it does not directly establish the knowledge-search mechanism. Complementary evidence from environmental governance highlights the institutional and organizational conditions underlying AI adoption. Zhou [54] identified configurations of institutional, cognitive, technological, and social-structural factors associated with AI adoption in Chinese environmental agencies. Although based on public-sector evidence, the study offers a complementary perspective on the conditions that may enable AI adoption.

Firms with superior ESG performance are generally more likely to invest in and adopt AI technologies for at least three reasons. First, firms with higher ESG performance typically enjoy greater financial resources and lower financing costs [4], providing the financial foundation for AI procurement, deployment, and talent acquisition. Second, the governance dimension of ESG emphasizes digital governance and information management capabilities, which naturally complement AI adoption. Third, ESG disclosure requirements encourage firms to establish more comprehensive data governance systems, thereby creating the high-quality structured data infrastructure necessary for effective AI implementation.

Once AI technologies are adopted, they directly enhance corporate sustainable innovation capability by reducing knowledge search costs, optimizing R&D processes, and improving organizational decision-making efficiency [15,46]. Nevertheless, existing studies on the relationship between AI application and corporate innovation predominantly treat AI either as an independent explanatory variable or as an outcome variable. Few studies have incorporated AI into the broader “ESG–innovation” framework to examine whether AI serves as a specific transmission mechanism through which ESG enhances innovation. Specifically, firms with superior ESG performance generally possess stronger governance structures and more abundant strategic resources. Whether these advantages systematically promote AI adoption and subsequently transform ESG-related resources into sustainable innovation outputs. Accordingly, the following hypothesis is proposed:

Hypothesis H2.

Artificial intelligence adoption positively mediates the relationship between ESG performance and corporate sustainable innovation capability.

2.3. Knowledge Diversification and Corporate Sustainable Innovation Capability

The concept of knowledge diversification (or technological diversification) originates from organizational learning theory and the knowledge-based view of the firm [55]. It refers to the breadth and balance of a firm’s technological knowledge across different technological fields or knowledge domains. The impact of knowledge diversification on corporate sustainable innovation capability has generated two competing theoretical perspectives. One perspective argues that knowledge diversification enhances sustainable innovation capability by facilitating recombinant innovation, as the combination and integration of knowledge from different technological domains constitute an important source of breakthrough innovation [56]. The alternative perspective contends that excessive knowledge diversification may dilute firms’ core technological capabilities, increase the coordination costs of internal knowledge integration, and consequently exert a nonlinear effect on innovation capability [19].

Research has further extended the knowledge diversification perspective toward the diversification of external knowledge sources. Yun et al. [57] from the perspective of firms’ intermediate goods import diversification, find that the geographical and technological diversification of knowledge sources jointly promotes innovation capability through complementary effects. Similarly, Sun et al. [58], in the context of strategic alliances, demonstrate that greater knowledge heterogeneity among alliance partners significantly enhances corporate sustainable innovation capability, although this relationship also exhibits an inverted U-shaped pattern characterized by diminishing marginal returns. At the theoretical level, Zahra and George [59] systematically reconceptualized the absorptive capacity framework proposed by Cohen and Levinthal [44] by distinguishing between potential absorptive capacity (knowledge acquisition and assimilation) and realized absorptive capacity (knowledge transformation and exploitation). This refinement provides a more comprehensive theoretical foundation for the absorptive capacity threshold mechanism. Specifically, the threshold effect of R&D intensity identified in this paper is conceptually more closely related to the dimension of realized absorptive capacity proposed by Zahra and George [59]. Chen and Yu [60] further demonstrated that the performance-enhancing effect of knowledge management capability is itself contingent upon firms’ internal knowledge integration mechanisms, which is consistent with the transformation mechanism linking knowledge diversification and corporate sustainable innovation capability. Complementary evidence from the hospitality sector further emphasizes the importance of knowledge management processes. Dat et al. [61] found that knowledge acquisition, dissemination, and utilization positively influence innovation capacity, with knowledge utilization exhibiting the strongest effect. These findings underscore the relevance of knowledge utilization when considering how a diversified knowledge base may support innovation.

From a theoretical perspective, ESG performance is directly associated with knowledge diversification. Firms with superior ESG performance generally possess more effective governance structures and stronger capabilities for cross-functional knowledge integration and coordination, consistent with the stakeholder management framework proposed by Freeman [2]. Furthermore, ESG investments require firms to develop competencies across environmental, social, and governance domains that were previously outside their core knowledge base, thereby objectively expanding the diversity of their knowledge structures. However, to examine knowledge diversification as a mediating mechanism through which ESG performance influences corporate sustainable innovation capability. Accordingly, the following hypothesis is proposed:

Hypothesis H3.

Knowledge diversification mediates the relationship between ESG performance and corporate sustainable innovation capability.

3. Research Design

3.1. Data Sources and Sample Selection

The data used in this study cover the period 2009–2025 and were obtained from five primary sources. First, firm-level financial and corporate governance data are collected from the China Stock Market and Accounting Research (CSMAR) database, including information on balance sheets, income statements, ownership structure, and board governance. Second, ESG performance data are obtained from the Huazheng ESG Rating Database, which provides both an overall ESG score and separate scores for the environmental (E), social (S), and governance (G) dimensions. Since its introduction in 2009, the Huazheng ESG rating system has continuously expanded its coverage and has become one of the most widely used ESG databases in China [62,63,64]. Its disaggregated E, S, and G scores have been extensively employed to examine the heterogeneous effects of different ESG dimensions on corporate outcomes such as green total factor productivity and green innovation [65]. Third, the corporate sustainable innovation capability (hereafter, IIP) is constructed following recent studies on comprehensive measures of sustainable and green innovation performance [66]. Fourth, the level of artificial intelligence (AI) application is measured using AI-related patent identification based on patent texts and International Patent Classification (IPC) codes. This approach has recently been widely adopted to quantify AI technology adoption among Chinese listed firms [18,67]. Fifth, the knowledge diversification (TD) is constructed based on the diversification of firms’ patent technology classifications. This measure originates from the patent entropy index methodology and has recently been extended to research on knowledge diversification among Chinese high-tech listed firms [68,69]. These five datasets are merged by firm stock code and fiscal year to construct the final analytical database. The sample selection procedure consists of three steps. First, firms designated as ST or *ST are excluded based on their stock abbreviations. Second, firms in the financial and insurance industries are excluded. Third, all continuous variables are winsorized at the 1st and 99th percentiles to mitigate the influence of extreme values (detailed sample selection procedures are reported in Appendix A Table A1).

3.2. Variable Definitions and Measurement

3.2.1. Variable Definition

(1) Dependent variable: The dependent variable is the corporate sustainable innovation capability (IIP). It is calculated as follows:${\mathit{IIP}}_{i,t}=\mathit{ln}\left[{\displaystyle \frac{{\mathit{IIN}}_{i,t}+{\mathit{IIN}}_{i,t-1}}{{\mathit{IIN}}_{i,t-1}+{\mathit{IIN}}_{i,t-2}}}·\left({\mathit{IIN}}_{i,t}+{\mathit{IIN}}_{i,t-1}\right)+1\right]$, where IIN_i,t denotes the amount of R&D investment by firm i in year t, while IIN_i,t−1 and IIN_i,t−2 represent the firm’s R&D investment in the preceding one and two years, respectively. The index reflects the scale of R&D investment and its intertemporal change, with a higher value generally indicating a greater level of sustained innovation capability. The validity of this measurement approach is further supported by recent empirical studies on ESG and innovation capability. Ma et al. [70] using Chinese A-share listed firms show that ESG performance enhances green new quality productivity, a comprehensive indicator of IIP, through green innovation. Similarly, Tan et al. [71] find that ESG promotes the green transformation of resource-based enterprises primarily through R&D investment and risk-taking channels. Such transformation performance represents a comprehensive rather than a single quantitative measure of innovation output.

(2) Key independent variables: The primary explanatory variable is ESG score, measured by the overall ESG score provided by the Huazheng ESG Rating System. The Huazheng ESG rating has been widely adopted as the principal ESG indicator in empirical studies of Chinese A-share listed firms [14,63]. To further investigate the heterogeneous effects of different ESG dimensions, the overall ESG score is decomposed into three sub-dimensions: E_score (Environmental performance); S_score (Social performance); and G_score (Governance performance). This parallel examination of the overall ESG score and its three constituent dimensions follows the approach of Niu and Wang [65], who separately examined the effects of the environmental, social, and governance dimensions on green innovation and green total factor productivity. It is also consistent with Du et al. [72], who emphasized the heterogeneous effects of individual ESG dimensions when developing a new ESG evaluation framework for Chinese firms.

(3) Mediating variables: The first mediating variable is AI, representing the level of artificial intelligence application. AI adoption is measured by identifying AI-related technologies from patent texts and International Patent Classification (IPC) codes. It is then measured as the natural logarithm of one plus the number of AI-related patent applications filed by a firm in a given year: ${\mathit{AI}}_{i,t}=\ln \left(1+N_{i,t}^{\mathit{AI}}\right),$ where $N_{i,t}^{\mathit{AI}}$ denotes the number of AI-related patent applications filed by firm i in year t. This measurement approach has been widely adopted in recent studies using patent data from Chinese listed firms [73]. Zhai and Liu [18] identified more than 43,000 AI patents of Chinese listed firms based on China’s Strategic Emerging Industries Classification and IPC codes and examined the effect of AI technological innovation on total factor productivity. Su et al. [67] similarly measured AI innovation intensity using the number of AI patent applications filed by state-owned enterprises. Zu et al. [74] further constructed a firm-level AI innovation indicator based on Chinese patent data to investigate its impact on corporate environmental sustainability. The second mediating variable is TD, is calculated based on the distribution of a firm’s patents across different primary IPC classes in a given year. Specifically, it is constructed as: $\mathit{TD}=-{\sum }_{j=1}^n{p}_j\mathit{ln}{p}_j,$ where p_i denotes the proportion of a firm’s total patents in a given year that belong to primary classification j. This entropy-based approach has recently been extended to the Chinese context. For example, Wang and Lyu [69] employed measures of knowledge depth and breadth diversification as mediating variables when examining the effects of data intelligence on firms’ knowledge recombination capability. Kaneko and Kajikawa [68] further proposed an industry-adjusted measure of relative technological diversification, providing a more comparable metric across industries.

(4) Control variables: Following the research design, ten control variables are included in all empirical models. Ownership concentration (Big1), measured by the shareholding ratio of the largest shareholder. It has been shown that heterogeneous ownership structures significantly influence innovation decisions in Chinese state-owned enterprises [75]. Cash ratio (Cash), measured as cash and cash equivalents divided by total assets. Cash holdings alleviate financing constraints and affect firms’ R&D investment decisions [76]. CEO duality (Duality), coded as 1 if the CEO simultaneously serves as board chair and 0 otherwise. It has been found that CEO duality moderates the relationship between CEO social capital and corporate innovation [77]. Fixed asset ratio (Fixratio), measured as fixed assets divided by total assets. It has been shown that firms’ fixed asset intensity significantly affects R&D investment and corporate performance [78]. Sales growth rate (Growth), measured as the ratio of current to previous operating revenue. This variable is widely used to control for firms’ growth opportunities [63,79]. Industry Herfindahl–Hirschman Index (HHI), measuring industry concentration. It has been found that market power and peer effects significantly influence green innovation among Chinese listed firms [80]. Proportion of independent directors (Indirector): it has been reported that the presence of independent directors, particularly those with technical expertise, significantly affects firms’ innovation output [81]. Financial leverage (Leverage), measured as total liabilities divided by total assets. It has been demonstrated the direct impact of leverage on corporate innovation investment [82]. Managerial ownership (Manager), measured by the proportion of shares held by management. It has been shown that managerial ownership and institutional ownership jointly influence corporate innovation [83]. Market-to-book ratio (MTB), reflecting firms’ growth opportunities. It has been found that growth opportunities are an important determinant of corporate investment decisions and are moderated by leverage [84].

3.2.2. Descriptive Statistics of the Variables

As shown in Table 1, several noteworthy distributional characteristics emerge from the sample (the correlations among the variables are reported in Appendix A Table A2). First, the AI variable exhibits a pronounced zero-inflated and right-skewed distribution. Its median, 25th percentile (P25), and 75th percentile (P75) are all equal to zero, while the mean (0.294) is substantially smaller than the standard deviation (0.720). These statistics indicate that approximately two-thirds of the firm-year observations report no AI-related patent output. This distributional feature provides the primary motivation to further examine the AI mediation pathway. Second, the distribution of the IIP is relatively symmetric. The mean (18.771) is very close to the median (18.699), and the standard deviation (1.507) is modest relative to the sample mean, suggesting the absence of severe skewness. A similar pattern is observed for ESG_score (mean = 73.185; median = 73.300) and the three ESG sub-dimensions (E, S, and G), indicating that the Huazheng ESG rating system provides satisfactory discriminatory power and measurement stability within the sample period. Third, the knowledge diversification index (TD) also exhibits a relatively balanced distribution, with a mean of 1.856 and a median of 1.835. However, the considerable gap between the first quartile (1.040) and the minimum value (0) suggests substantial heterogeneity in firms’ technological knowledge structures. While some firms possess highly concentrated patent portfolios, others exhibit much more diversified knowledge bases. This cross-firm variation provides sufficient statistical support for identifying the mediating role of knowledge diversification. Fourth, the control variables also display several notable characteristics. The standard deviation of Manager (managerial ownership) is 20.143, substantially exceeding its median (1.655), while the mean (14.861) lies between the two. This pattern indicates a highly polarized ownership structure among Chinese listed firms: managerial ownership is extremely low for most firms but exceptionally high for a relatively small number of firms. Meanwhile, the average value of Big1 (33.865%) reflects the well-documented phenomenon of relatively concentrated ownership among Chinese listed firms.

Table 1. Descriptive Statistics of the Main Variables.

Variable	Mean	SD	Min	P25	Median	P75	Max
IIP	18.771	1.507	14.780	17.796	18.699	19.665	23.031
ESG_score	73.185	5.852	55.917	69.960	73.300	76.660	89.163
E_score	61.757	7.564	45.800	56.996	61.020	66.139	83.043
S_score	75.601	8.512	48.507	71.170	76.530	80.970	95.120
G_score	79.414	6.228	56.007	76.750	80.579	83.460	90.950
AI	0.294	0.720	0.000	0.000	0.000	0.000	3.526
TD	1.856	1.163	0.000	1.040	1.835	2.616	5.039
Big1	33.865	14.936	8.226	22.316	31.566	43.874	74.401
Cash	0.172	0.136	0.010	0.075	0.132	0.228	0.665
Duality	0.312	0.463	0.000	0.000	0.000	1.000	1.000
Fixratio	0.209	0.155	0.002	0.087	0.180	0.300	0.679
Growth	1.137	0.368	0.437	0.954	1.084	1.236	3.189
HHI	0.105	0.123	0.016	0.044	0.066	0.118	0.812
Indirector	37.701	5.307	33.330	33.330	36.360	42.860	57.140
Leverage	0.414	0.204	0.054	0.251	0.404	0.561	0.918
Manager	14.861	20.143	0.000	0.002	1.654	28.182	69.184
MTB	0.297	0.155	0.024	0.185	0.273	0.382	0.788

3.3. Model Specification

Baseline Regression Model (Two-Way Fixed Effects):

$${\mathit{IIP}}_{\mathit{it}}={\alpha }_0+{\alpha }_1\mathit{ESG}\text{\_}{\mathit{score}}_{\mathit{it}}+{\sum }_k{\gamma }_k{X}_{\mathit{kit}}+{\mu }_i+{\lambda }_t+{\epsilon }_{\mathit{it}}$$

(1)

ESG Sub-Dimension Model:

$${\mathit{IIP}}_{\mathit{it}}={\beta }_0+{\beta }_1{E}_{\mathit{it}}+{\beta }_2{S}_{\mathit{it}}+{\beta }_3{G}_{\mathit{it}}+{\sum }_k{\delta }_k{X}_{\mathit{kit}}+{\mu }_i+{\lambda }_t+v_{\mathit{it}}$$

(2)

where μ_i denotes firm fixed effects, estimated through the within-transformation, and λ_t represents year fixed effects, implemented using year dummy variables. Heteroskedasticity-robust HC1 standard errors are reported. In selected robustness analyses, two-way clustered standard errors at the firm and year levels are employed.

A two-way fixed-effects (TWFE) specification is adopted as the baseline model rather than a random-effects model because firms possess numerous unobservable, time-invariant characteristics that may simultaneously influence both ESG performance and IIP, such as founders’ values, industry-specific endowments, and location advantages. Year fixed effects are included instead of a linear time trend because they provide a more flexible way to absorb nonlinear macroeconomic shocks occurring during the sample period, including the global financial crisis, trade disputes, and the COVID-19 pandemic. This specification helps avoid estimation bias arising from incorrect functional form assumptions regarding time trends. The estimation method and corresponding standard errors are explicitly reported for all regression models to ensure transparency, and comparability of the empirical results. The empirical strategy also follows recent advances in panel data econometrics. Han and Lee [85] point out that fixed-effects estimation in short-panel settings may suffer from finite-sample bias. To alleviate this concern, the present study reports HC1 heteroskedasticity-robust standard errors and, where appropriate, two-way clustered standard errors. Furthermore, Hernandez et al. [86] emphasize that studies involving network effects should pay particular attention to reverse causality and simultaneity bias in causal identification.

4. Empirical Analysis

4.1. Main Regression Results

Table 2 reports the baseline regression results. First, overall ESG performance is positively and significantly associated with IIP. Column (1) shows that the coefficient on ESG_score is 0.017 and statistically significant at the 1% level, indicating a highly significant positive relationship between the two variables. Therefore, Hypothesis H1 is supported. From an economic perspective, a one-standard-deviation increase in ESG performance (5.852 points) is associated with an average increase of approximately 0.099 points in the IIP. This corresponds to approximately 6.60% of the sample standard deviation of IIP (1.507), representing a moderately large economic effect relative to those commonly reported in firm-level innovation studies. Second, all three ESG dimensions are positively associated with IIP, with the social dimension having the largest unstandardized coefficient. Columns (2)–(4) report the regression results for the three ESG dimensions. When estimated separately, E_score, S_score, and G_score all exhibit positive coefficients that are statistically significant at the 1% level. Specifically, the estimated coefficients for E_score, S_score, and G_score are 0.008, 0.011 and 0.008, respectively. More importantly, after simultaneously controlling for all three ESG dimensions in Column (5), the coefficients remain positive and statistically significant. However, their marginal effects differ substantially, following the ordering: S_score (0.010) > G_score (0.007) > E_score (0.005). This finding indicates that each ESG dimension contributes positively to sustainable innovation capability. The largest marginal effect of the social (S) dimension indicates comparing with environmental initiatives, which often rely on engineering-oriented technological improvements, investments in the social dimension enhance sustainable innovation through broader mechanisms such as human capital accumulation, organizational learning, and knowledge collaboration networks. Third, the control variables show varying associations with IIP. The coefficient of Growth is significantly positive, suggesting that firms experiencing rapid business expansion tend to invest more actively in innovation activities. The coefficient of Fixratio is significantly negative, implying that firms with a higher proportion of fixed assets have lower flexibility in resource allocation, thereby reducing their ability to dynamically adjust innovation investments. The proportion of independent directors (Indirector) also exhibits a significantly negative coefficient. The result may instead reflect the relatively formalistic nature of independent director governance mechanisms in the Chinese institutional context, a possibility that deserves further investigation. Both Leverage and MTB have significantly positive coefficients, suggesting that debt financing may provide beneficial governance discipline, while stronger market growth expectations encourage firms to undertake greater innovation investment. Manager also exhibits a significantly positive coefficient, indicating a positive association between managerial characteristics captured by this variable and corporate sustainable innovation capability. Finally, Big1, Cash, Duality, and HHI are statistically insignificant in the baseline regression.

Table 2. Baseline Regression Results.

Variables	IIP	IIP	IIP	IIP	IIP
	(1)	(2)	(3)	(4)	(5)
ESG_score	0.017 ***
	(0.001)
E_score		0.008 ***			0.005 ***
		(0.001)			(0.001)
S_score			0.011 ***		0.010 ***
			(0.001)		(0.001)
G_score				0.008 ***	0.007 ***
				(0.001)	(0.001)
Big1	0.001	0.001	0.001	0.001	0.001
	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)
Cash	−0.029	−0.016	−0.018	−0.035	−0.033
	(0.051)	(0.052)	(0.052)	(0.052)	(0.051)
Duality	0.004	0.006	0.002	0.003	0.002
	(0.013)	(0.013)	(0.013)	(0.013)	(0.013)
Fixratio	−0.356 ***	−0.391 ***	−0.369 ***	−0.369 ***	−0.351 ***
	(0.081)	(0.081)	(0.081)	(0.081)	(0.081)
Growth	0.237 ***	0.237 ***	0.228 ***	0.233 ***	0.232 ***
	(0.016)	(0.016)	(0.016)	(0.016)	(0.016)
HHI	−0.058	−0.041	−0.076	−0.049	−0.069
	(0.116)	(0.117)	(0.116)	(0.116)	(0.116)
Indirector	−0.005 ***	−0.004 **	−0.004 **	−0.005 ***	−0.004 ***
	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)
Leverage	0.712 ***	0.642 ***	0.638 ***	0.718 ***	0.688 ***
	(0.057)	(0.057)	(0.056)	(0.058)	(0.057)
Manager	0.004 ***	0.004 ***	0.004 ***	0.004 ***	0.004 ***
	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)
MTB	0.576 ***	0.593 ***	0.608 ***	0.611 ***	0.588 ***
	(0.050)	(0.051)	(0.050)	(0.051)	(0.050)
Firm FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
N	33,960	33,960	33,960	33,960	33,960
R2	0.320	0.315	0.319	0.314	0.321

Notes: HC1 heteroskedasticity-robust standard errors are reported in parentheses. *** and ** denote statistical significance at the 1% and 5% levels, respectively.

4.2. Robustness Tests

To ensure that the baseline finding regarding the positive relationship between ESG performance and IIP is not driven by specific data treatments, sample periods, model specifications, or potential selection issues, this study conducts five additional robustness tests. The results are reported in Table 3.

Table 3. Results of Robustness Test.

Variable	0.5% Winsorization	L.ESG	Excluding COVID-19	System GMM	Heckman
	(1)	(2)	(3)	(4)	(5)
ESG_score	0.017 ***		0.018 ***	0.018 ***	0.075 ***
	(0.001)		(0.001)	(0.002)	(0.001)
L.ESG		0.014 ***
		(0.001)
L1.IIP				0.888 ***
				(0.021)
IMR					−0.000 **
					(0.000)
Controls	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Dynamic Panel	No
Year FE	Yes	Yes	Yes	Yes	Yes
N	33,960	32,642	27,901	33,960	33,013
R2	0.320	0.302	0.335	—	—

Note: Standard errors are reported in parentheses. Column (1) applies two-sided winsorization at the 0.5% level. Column (2) uses the one-period lagged ESG score. Column (3) excludes observations from 2020–2021. Column (4) employs a two-step System GMM estimator, including the one-period lagged IIP as a dynamic term; the second and third lags of IIP and ESG are used as GMM-style instruments, and the collapse option is applied to limit instrument proliferation. Column (5) employs the Heckman two-step procedure, where IMR denotes the inverse Mills ratio. *** and ** denote statistical significance at the 1% and 5% levels, respectively.

4.2.1. More Stringent Winsorization

The first robustness test examines whether the baseline relationship between ESG performance and IIP is disproportionately influenced by extreme observations. Although the continuous variables in the baseline analysis have already been treated to mitigate the influence of outliers, the estimated relationship could still be sensitive to a small number of observations located at the tails of the distributions. This concern is particularly relevant for firm-level financial and innovation data, where substantial heterogeneity across firms may generate unusually large or small values. To address this issue, this study adopts a more stringent data-treatment procedure by winsorizing all continuous variables at the 0.5% level at both the upper and lower tails and subsequently re-estimating the baseline model using the resulting sample.

As reported in Column (1) of Table 3, the coefficient of ESG performance remains positive and highly statistically significant. Importantly, both the sign and magnitude of the coefficient remain highly consistent with those obtained from the baseline specification. These findings suggest that the positive association between ESG performance and IIP is not attributable to a small number of influential observations or extreme values in the sample. Instead, the estimated ESG effect represents a systematic relationship that persists after reducing the potential influence of distributional tails. Therefore, the main empirical conclusion is robust to alternative winsorization thresholds and does not depend critically on the particular treatment of outliers adopted in the baseline analysis.

4.2.2. One-Period Lagged ESG Variable

The second robustness test addresses potential concerns regarding contemporaneous reverse causality between ESG performance and IIP. Although the baseline results indicate a significant positive relationship between ESG performance and IIP, the contemporaneous specification may not fully rule out the possibility that firms with greater IIP simultaneously improve their ESG performance. Such reverse causality could potentially bias the estimated effect of ESG performance. To mitigate this concern and establish a clearer temporal ordering between the explanatory and dependent variables, this study replaces the contemporaneous ESG score with its one-period lagged value ESG_i,t−1 and re-estimates the baseline fixed-effects model.

As reported in Column (2) of Table 3, the coefficient of the one-period lagged ESG score is positive and highly statistically significant. The estimated coefficient is only slightly smaller than that obtained from the baseline specification, while its direction and statistical significance remain unchanged. These results indicate that ESG performance continues to exert a positive effect on subsequent IIP even after introducing a temporal lag between ESG performance and the investment outcome. Therefore, the baseline finding is unlikely to be driven solely by contemporaneous reverse causality. More broadly, the persistence of the positive ESG coefficient under the lagged specification provides additional evidence supporting the robustness of the positive relationship between ESG performance and IIP.

4.2.3. Excluding the COVID-19 Period

The third robustness test addresses the possibility that the baseline estimates may be influenced by the extraordinary economic and operational disruptions associated with the COVID-19 pandemic. During 2020 and 2021, firms faced substantial uncertainty arising from supply-chain disruptions, mobility restrictions, changes in market demand, financing constraints, and government intervention. These exceptional circumstances may have simultaneously affected firms’ ESG performance and their willingness or ability to undertake IIP. Consequently, including the pandemic period could potentially introduce temporary shocks that are not representative of firms’ normal ESG and innovation behavior. To ensure that the baseline relationship is not driven by these unusual observations, this study excludes all firm-year observations corresponding to 2020 and 2021 and re-estimates the baseline fixed-effects specification.

As shown in Column (3) of Table 3, after removing the COVID-19 period, the coefficient of ESG performance remains positive and statistically significant at the 1% level. The estimated coefficient is also very close to the baseline estimate and to that obtained from the alternative winsorization test. These results demonstrate that the positive relationship between ESG performance and IIP is not an artifact of pandemic-specific conditions or temporary policy and market disturbances during 2020–2021. Rather, the relationship remains evident when the analysis is restricted to periods characterized by relatively more conventional economic conditions. This evidence therefore strengthens the external temporal validity of the baseline results and suggests that the positive ESG–innovation relationship reflects a more persistent corporate behavior rather than a temporary response to an exceptional macroeconomic shock.

4.2.4. Dynamic Panel Estimation Using System GMM

The fourth robustness test explicitly considers the dynamic nature of IIP and potential endogeneity in the ESG–IIP relationship. Innovation investment decisions are typically characterized by substantial persistence because ongoing innovation projects often span multiple periods, require continued resource commitments, and generate organizational capabilities that influence subsequent investment decisions. Consequently, current IIP may depend strongly on its own historical level. Ignoring such dynamic persistence could lead to model misspecification and potentially bias the estimated ESG coefficient. Furthermore, ESG performance may be endogenous because of reverse causality, omitted time-varying factors, or simultaneity between corporate sustainability strategies and innovation decisions. To address these concerns, this study estimates a dynamic panel model using the two-step System GMM approach.

Specifically, the first-order lag of sustainable innovation investment is incorporated into the model to capture dynamic persistence. The second- and third-order lags of IIP and ESG performance are employed as GMM-style instruments, while the control variables enter as standard instruments. In addition, the collapse option is adopted to restrict the number of instruments and mitigate the potential problem of instrument proliferation. As reported in Column (4), the coefficient of ESG performance remains positive and statistically significant. Notably, this coefficient is highly consistent with the estimates obtained from the fixed-effects specifications, providing additional evidence that the positive ESG effect is not sensitive to the use of a dynamic panel estimator. Meanwhile, the coefficient of lagged IIP is positive and strongly significant, revealing substantial persistence in firms’ sustainable innovation investment behavior. This result confirms the importance of accounting for the dynamic structure of the dependent variable. Taken together, the System GMM estimates reinforce the baseline conclusion by showing that the positive relationship between ESG performance and IIP holds even after explicitly incorporating dynamic adjustment and addressing potential sources of endogeneity through internal instruments.

4.2.5. Heckman Two-Step Correction for Sample Selection Bias

The final robustness test examines whether potential sample-selection bias affects the estimated relationship between ESG performance and IIP. A non-negligible proportion of firm-year observations may report zero sustainable innovation capability, implying that the process determining whether a firm engages in such investment could differ from the process determining the magnitude of investment among firms that actually invest. If the decision to undertake sustainable innovation investment is systematically related to unobserved firm characteristics, estimates based solely on the observed outcome process may be subject to non-random selection bias. To address this concern, this study employs the Heckman two-step selection correction procedure and explicitly distinguishes between the participation decision and the subsequent level of IIP.

In the first stage, a Probit selection equation is estimated to model the probability that a firm reports positive IIP. Based on the estimated selection equation, the inverse Mills ratio (IMR) is calculated and incorporated into the second-stage outcome equation to control for potential non-random sample selection. Industry and year fixed effects are also included to account for systematic heterogeneity across industries and common temporal shocks. As reported in Column (5), the coefficient of ESG performance remains positive and highly statistically significant after correcting for sample selection. Moreover, the estimated IMR is negative and statistically significant at the 5% level, suggesting that the selection process is not entirely random and that correcting for potential selection bias is empirically relevant. Nevertheless, after explicitly incorporating this correction, the estimated effect of ESG performance remains strongly positive. The Heckman results therefore provide additional evidence that the baseline conclusion is not driven by systematic differences between observations with zero and positive IIP. More broadly, the persistence of the positive ESG coefficient after controlling for potential sample-selection effects further strengthens the robustness of the central empirical finding.

Taken together, the five robustness tests provide consistent evidence supporting the baseline results. The positive relationship between ESG performance and IIP remains statistically significant after applying a stricter treatment of extreme observations, introducing a temporal lag in ESG performance, excluding the COVID-19 period, accounting for the dynamic persistence of innovation investment through System GMM, and correcting for potential sample-selection bias using the Heckman two-step procedure. Although the magnitude of the estimated ESG coefficient varies across econometric specifications because of differences in model structure and estimation samples, its sign and statistical significance remain remarkably stable. Overall, these findings provide strong evidence that the positive association between ESG performance and IIP is not dependent on a particular sample construction, variable treatment, or estimation methodology, thereby reinforcing the reliability of the main empirical conclusion.

4.3. Endogeneity Analysis

The relationship between ESG performance and IIP may be subject to two major sources of endogeneity. First, reverse causality may exist, as firms with stronger innovation capabilities are more likely to obtain higher ESG ratings. For example, superior technological capabilities and green patent outputs may indirectly improve firms’ environmental (E) scores. Second, omitted-variable bias may arise from unobservable factors such as entrepreneurial orientation, strategic transformation capability, or innovation culture that simultaneously affect both ESG performance and IIP. Although the two-way fixed-effects model controls for time-invariant firm heterogeneity, potential time-varying endogeneity still requires additional treatment.

4.3.1. Instrumental Variable Estimation (IV-2SLS)

To further address potential endogeneity in the relationship between ESG performance and IIP, this study employs an instrumental variable approach based on two-stage least squares (IV-2SLS). Specifically, the ESG score of the same firm in the immediately preceding calendar year, ESG_i,t−1, is used as an instrument for the contemporaneous ESG score, namely, ESG_it.

The use of lagged ESG is motivated by the strong persistence of firms’ ESG practices over time. Corporate ESG performance generally reflects relatively stable organizational characteristics, including governance structures, environmental management systems, stakeholder relationships, and long-term strategic orientation. Consequently, a firm’s ESG performance in year t − 1 is expected to be strongly correlated with its ESG performance in year t, satisfying the relevance condition required for an instrumental variable. Importantly, the lagged ESG variable is constructed using a strict calendar-year matching procedure. For each firm-year observation, ESG_i,t−1 corresponds specifically to the ESG score of the same firm in the immediately preceding year. Thus, observations with discontinuous years are not mechanically treated as valid lags. This construction ensures that the instrument represents a genuine one-year lag rather than the previous available observation.

The first-stage regression can be expressed as:

ESG_it = π₀ + π₁ESG_i,t−1 + γ′X_it + μ_i + λ_t + u_it

(3)

where X_it denotes the vector of control variables, μ_i represents firm fixed effects, and λ_t captures year fixed effects. In the second stage, the predicted component of contemporaneous ESG is used to estimate its effect on IIP:

IIP_it = α₀ + βESG_it + δ′X_it + μ_i + λ_t + ε_it

(4)

Table 4 reports the IV-2SLS estimation results. The first-stage results provide strong evidence for instrument relevance. The coefficient of ESG_t−1 is 0.477 and is statistically significant at the 1% level, with a t-statistic of 67.081. More importantly, the partial first-stage F-statistic is 4499.842, which is substantially above the conventional threshold of 10. These results indicate that lagged ESG has exceptionally strong explanatory power for current ESG performance and that weak-instrument concerns are unlikely to affect the estimation.

Table 4. IV-2SLS Endogeneity Test.

Variable	ESG_Score	IIP
L. ESG_score	0.477 ***
	(67.081)
ESG_score		0.028 ***
		(12.636)
Partial first-stage F statistic	4499.842
Partial first-stage F p-value	0.000
Wu–Hausman statistic		43.119
Wu–Hausman p-value		0.000
N	32,642	32,642
Adj. R2	0.221	0.301
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes

Note: The ESG_score column reports the first-stage regression, while the IIP column reports the second-stage regression. L.ESG denotes the ESG score of the same firm in the immediately preceding calendar year (t − 1) and is used as the instrumental variable for current ESG_score. Robust t-statistics are reported in brackets. The partial first-stage F statistic is based on a heteroskedasticity-robust Wald test. *** indicates significance at the 1% level.

The second-stage results show that the IV estimate of the ESG coefficient is positive and statistically significant. The estimated coefficient of ESG is 0.028, with a t-statistic of 12.636, and is significant at the 1% level. The positive association between ESG performance and corporate IIP persists when lagged ESG is used as an instrument for contemporaneous ESG. The Wu–Hausman test further supports the relevance of explicitly addressing endogeneity. The test statistic is 43.119 with a p-value below 0.001, leading to a strong rejection of the null hypothesis that contemporaneous ESG can be treated as exogenous. This result indicates that conventional fixed-effects estimates may be affected by endogeneity and provides justification for the IV-2SLS specification.

The magnitude of the IV coefficient is also noteworthy. The IV estimate of 0.028 remains positive and statistically significant, reinforcing the conclusion that ESG performance contributes positively to sustainable innovation capability. The fact that the estimated effect persists after instrumenting with contemporaneous ESG suggests that the baseline relationship is unlikely to be explained solely by reverse causality or contemporaneous omitted-variable bias.

Overall, the IV-2SLS results provide strong supplementary evidence supporting the positive relationship between ESG performance and corporate IIP. The extremely strong first-stage relationship, the statistically significant positive second-stage coefficient, and the rejection of exogeneity by the Wu–Hausman test jointly suggest that explicitly accounting for endogeneity does not overturn the central empirical conclusion. Instead, the positive coefficient is retained under this alternative specification.

4.3.2. Propensity Score Matching Methods

This study employs Propensity Score Matching (PSM) to further mitigate potential selection bias arising from systematic differences between firms with relatively high and low ESG performance. Although the baseline fixed-effects model controls for observed firm characteristics and time-invariant firm heterogeneity, firms with stronger ESG performance may differ systematically from other firms in terms of size, financial conditions, governance characteristics, growth opportunities, and other observable attributes. Such differences may simultaneously affect IIP, thereby confounding the estimated relationship between ESG performance and sustainable innovation capability. PSM helps alleviate this concern by constructing a comparison group of firms with similar observable characteristics but different levels of ESG performance.

Specifically, a binary treatment variable, (HighESG_it), is constructed based on the median ESG score within each industry-year group: ${\mathit{HighESG}}_{\mathit{it}}=\left\{\begin{array}{cc}1,& {\mathit{ESG}}_{\mathit{it}}>\mathrm{Median}{\left(\mathit{ESG}\right)}_{\mathit{jt}},\\ 0,& {\mathit{ESG}}_{\mathit{it}}\le \mathrm{Median}{\left(\mathit{ESG}\right)}_{\mathit{jt}}\end{array}\right.$ where j denotes the industry to which firm i belongs. Defining treatment status relative to the industry-year median allows the classification of high-ESG firms to account for systematic differences in ESG performance across industries and over time.

The propensity score is subsequently estimated as the conditional probability that a firm belongs to the high-ESG group given its observed characteristics: p(X_it) = P(HighESG_it = 1 | X_it) where X_it represents the full set of control variables used in the baseline specification. Based on the estimated propensity scores, each treated observation is matched with a control observation using one-to-one nearest-neighbor matching without replacement. Matching is conducted within each year and subject to a caliper of 0.05, thereby restricting matches to treated and control firms with sufficiently similar estimated propensity scores.

The average treatment effect on the treated (ATT) can be expressed as: ATT = E[IIP_it(1) − IIP_it(0)|HighESG_it = 1], where IIP_it(1) represents the sustainable innovation capability of a high-ESG firm and IIP_it(0) denotes the counterfactual sustainable innovation capability that the same firm would have exhibited had it not belonged to the high-ESG group. Because this counterfactual outcome is unobservable, PSM approximates it using matched low-ESG firms with similar observable characteristics.

Table 5 reports the PSM estimation results. The matching procedure generates 14,210 matched pairs, corresponding to 28,420 firm-year observations. Column (1) reports the paired ATT estimate. The estimated coefficient of High ESG is 0.600 and is statistically significant at the 1% level, with a standard error of 0.017. This result indicates that, among firms with comparable observed characteristics, firms in the high-ESG group exhibit significantly greater IIP than their matched low-ESG counterparts.

Table 5. Propensity Score Matching Results.

Variable	PSM-ATT	PSM + FE
	(1)	(2)
High ESG	0.600 ***	0.109 ***
	(0.017)	(0.010)
Controls	No	Yes
Firm FE	No	Yes
Year FE	Exact matching	Yes
Matched pairs	14,210	14,210
N	28,420	28,420
R2		0.330
Mean |SMD| before	0.102
Mean |SMD| after	0.046

Note: High ESG equals one when a firm’s ESG score exceeds the median ESG score of its industry-year group. Propensity scores are estimated using the control variables. One-to-one nearest-neighbor matching without replacement is performed within each year using a caliper of 0.05. Column (1) reports the paired ATT, and Column (2) reports the firm and year fixed-effects estimate using the matched sample. Standard errors are reported in parentheses. *** indicates significance at the 1% level.

An important prerequisite for the credibility of PSM is that the matching procedure substantially improves covariate balance between the treatment and control groups. The balance tests reported in Table 6 indicate a considerable improvement after matching. Specifically, the mean absolute standardized mean difference (Mean |SMD|) across the matching covariates declines from 0.102 before matching to 0.046 after matching. The substantial reduction in covariate imbalance suggests that the matching procedure successfully constructs treatment and control groups that are considerably more comparable in terms of observed firm characteristics.

Table 6. Covariate Balance Test.

Variable	SMD Before	SMD After
Big1	0.146	−0.081
Cash	0.199	−0.099
Duality	−0.012	0.008
Fixratio	−0.068	0.019
Growth	−0.012	0.005
HHI	−0.016	−0.001
Indirector	0.106	−0.055
Leverage	−0.230	0.105
Manager	0.081	−0.029
MTB	0.153	−0.056

To further strengthen the analysis, this study re-estimates the baseline fixed-effects specification using only the matched sample. The corresponding model can be expressed as: IIP_it = α₀ + βHighESG_it + δ′X_it + μ_i + λ_t + ε_it, where X_it denotes the vector of control variables, μ_i represents firm fixed effects, and λ_t captures year fixed effects. Column (2) of Table 5 shows that the coefficient of High ESG remains positive and statistically significant at the 1% level after introducing the full set of controls and firm and year fixed effects. This result demonstrates that the positive relationship between ESG performance and sustainable innovation investment persists even after restricting the analysis to firms with substantially improved observable comparability and further controlling for unobserved time-invariant firm heterogeneity.

The PSM results therefore complement the baseline fixed-effects and IV-2SLS findings from a different perspective. Whereas the fixed-effects specification addresses time-invariant firm heterogeneity and the IV strategy is designed to address potential endogeneity of ESG performance, PSM focuses primarily on reducing observable selection bias by improving the comparability of firms with high and low ESG performance. The persistence of a positive ESG effect across these alternative specifications strengthens the robustness of the central empirical finding.

Taken together, the PSM results provide additional evidence supporting the positive relationship between ESG performance and IIP. The substantial improvement in covariate balance, the positive and statistically significant ATT estimate, and the persistence of the positive coefficient in the matched-sample fixed-effects regression jointly indicate that the main conclusion is not driven solely by systematic observable differences between high- and low-ESG firms.

4.3.3. Double Machine Learning

To further examine whether the estimated effect of ESG performance on IIP is sensitive to functional-form assumptions and potentially complex confounding structures, this study employs the Double Machine Learning (DML) framework as an additional robustness and endogeneity analysis. Unlike conventional linear regression models, DML allows the relationships between the control variables, ESG performance, and IIP to be estimated flexibly using machine-learning algorithms, thereby reducing potential bias arising from nonlinearities, higher-order interactions, and model misspecification. Related sustainability research has also employed explainable machine learning to characterize nonlinear relationships between urban form and built-environment carbon emissions [87].

Specifically, the DML procedure adopts a partialling-out framework. Firm fixed effects are first removed through within-firm transformation. Random Forest is then employed to separately estimate the conditional expectations of IIP and ESG performance as functions of the full set of control variables and year fixed effects. The corresponding residuals can be expressed as: ${\widetilde{\mathit{IIP}}}_{\mathit{it}}={\mathit{IIP}}_{\mathit{it}}-\widehat{\mathbb{E}}\left({\mathit{IIP}}_{\mathit{it}}\mid X_{\mathit{it}}\right)$ and ${\widetilde{\mathit{ESG}}}_{\mathit{it}}={\mathit{ESG}}_{\mathit{it}}-\widehat{\mathbb{E}}\left({\mathit{ESG}}_{\mathit{it}}\mid X_{\mathit{it}}\right)$ where X_it contains the control variables and year fixed effects after accounting for firm-level heterogeneity. The parameter of interest is subsequently estimated by regressing the residualized sustainable innovation investment on the residualized ESG score: ${\widetilde{\mathit{IIP}}}_{\mathit{it}}=\theta {\widetilde{\mathit{ESG}}}_{\mathit{it}}+{\epsilon }_{\mathit{it}}$.

An important feature of the DML framework is cross-fitting, which reduces overfitting and regularization bias arising from the use of flexible machine-learning algorithms. This study implements five-fold cross-fitting at the firm level. Specifically, GroupKFold is employed so that all observations belonging to the same firm are assigned to the same fold. Consequently, observations from a given firm do not simultaneously enter the training and validation samples within a particular cross-fitting iteration. This design is particularly appropriate for panel data because it reduces potential information leakage across observations belonging to the same firm. Random Forest is used as the machine-learning learner for both nuisance functions because of its ability to flexibly capture nonlinear relationships and interactions among firm characteristics without imposing a predetermined functional form. Standard errors in the final estimation are clustered at the firm level to account for within-firm dependence over time.

Table 7 reports the DML estimation results. The estimated coefficient of ESG performance is 0.017 and is statistically significant at the 1% level, with a firm-clustered standard error of 0.001 and a t-statistic of 11.980. The corresponding 95% confidence interval ranges from 0.014 to 0.020 and therefore lies entirely above zero. These results indicate that the positive effect of ESG performance on IIP remains highly robust after flexibly accounting for potentially nonlinear and complex relationships between the observed covariates, ESG performance, and sustainable innovation investment. In particular, after removing the components of both ESG performance and IIP that can be predicted by the observed firm characteristics and common year effects, the residual variation in ESG performance continues to exhibit a strong positive association with IIP.

Table 7. Double Machine Learning Results.

Variable	DML
ESG_score	0.017 ***
	[11.980]
Clustered SE	0.001
95% CI	[0.014, 0.020]
N	33,960
Firms	4510
Cross-fitting	5-fold
ML learner	Random Forest
Controls	Yes
Firm FE	Yes
Year FE	Yes

Note: This table reports the Double Machine Learning (DML) estimate of the effect of ESG performance on IIP. Random forests are employed to estimate the nuisance functions using five-fold cross-fitting grouped at the firm level. Firm and year fixed effects and all control variables are included. Standard errors are clustered at the firm level. Robust t-statistics are reported in brackets. *** indicates significance at the 1% level.

Taken together, the DML results provide additional evidence supporting the robustness of the positive relationship between ESG performance and IIP. The positive coefficient of 0.017, its narrow confidence interval, and its high statistical significance indicate that the main conclusion remains stable even when observable confounding relationships are modeled using a flexible machine-learning framework.

5. Mechanism Analysis

This study examines two potential mediating pathways: the technological capability pathway: ESG → AI adoption → corporate sustainable innovation capability; and the knowledge structure pathway: ESG → knowledge diversification → corporate sustainable innovation capability. This specification allows us to examine whether the two mechanisms make independent and non-substitutable contributions to the relationship between ESG performance and IIP.

5.1. Method for Testing the Mediating Effects

To examine the mediating roles of AI adoption and knowledge diversification (TD) in the relationship between ESG performance and IIP, this study employs a stepwise mediation framework. Specifically, the following mediation models are estimated separately for the AI and TD pathways.

For the AI application pathway, the mediation equations are specified as follows:

$${\mathit{AI}}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{ESG}\text{\_}{\mathit{score}}_{\mathit{it}}+\delta X_{\mathit{it}}+{\mu }_i+{\lambda }_t+{\epsilon }_{\mathit{it}}$$

(5)

$${\mathit{IIP}}_{\mathit{it}}={\beta }_0+{\beta }_2\mathit{ESG}\text{\_}{\mathit{score}}_{\mathit{it}}+{\beta }_3{\mathit{AI}}_{\mathit{it}}+\delta X_{\mathit{it}}+{\mu }_i+{\lambda }_t+{\epsilon }_{\mathit{it}}$$

(6)

Similarly, for the knowledge diversification pathway, the corresponding mediation equations are specified as:

$${\mathit{TD}}_{\mathit{it}}={\beta }_0+{\beta }_4\mathit{ESG}\text{\_}{\mathit{score}}_{\mathit{it}}+\delta X_{\mathit{it}}+{\mu }_i+{\lambda }_t+{\epsilon }_{\mathit{it}}$$

(7)

$${\mathit{IIP}}_{\mathit{it}}={\beta }_0+{\beta }_5\mathit{ESG}\text{\_}{\mathit{score}}_{\mathit{it}}+{\beta }_6{\mathit{TD}}_{\mathit{it}}+\delta X_{\mathit{it}}+{\mu }_i+{\lambda }_t+{\epsilon }_{\mathit{it}}$$

(8)

where AI_it denotes the level of artificial intelligence application of firm i in year t, TD_it represents knowledge diversification, and IIP_it denotes corporate sustainable innovation. ESG_score_it is the core explanatory variable measuring firm-level ESG performance, X_it represents the vector of control variables, μ_i denotes firm fixed effects, λ_t denotes year fixed effects, and ε_it is the error term.

5.2. Results of the Mediation Tests

Table 8 reports the regression-based mediation results for the two proposed mechanisms, namely AI adoption and knowledge diversification (TD) (supplementary mediation effect tests are reported in Appendix A Table A3). Column (1) first establishes the total effect of ESG performance on IIP. The coefficient of ESG_score is positive and statistically significant, confirming that firms with stronger ESG performance exhibit significantly higher IIP. This result provides the necessary empirical basis for examining the proposed mediating mechanisms (sample sizes across regression equations are reported in Appendix A Table A4).

Table 8. Mediation Model: IIP on AI, TD, ESG, and Controls.

Variables	IIP	AI	IIP	TD	IIP
	(1)	(2)	(3)	(4)	(5)
ESG_score	0.017 ***	0.004 ***	0.019 ***	0.010 ***	0.016 ***
	(0.001)	(0.001)	(0.002)	(0.002)	(0.002)
AI			0.084 ***
			(0.010)
TD					0.088 ***
					(0.011)
Controls	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
N	33,960	14,766	11,518	9531	8122
R2	0.320	0.054	0.356	0.145	0.378

Note: *** indicates significance at the 1%levels.

AI Adoption Pathway: Columns (2) and (3) examine the mediating role of corporate AI adoption. As shown in Column (2), ESG performance has a positive and highly significant effect on AI application. This finding indicates that firms with stronger ESG performance tend to exhibit a higher level of AI adoption, consistent with the argument that ESG-oriented firms are more likely to develop the resource base and digital governance capabilities necessary for adopting advanced technologies. Column (3) further introduces AI adoption into the sustainable innovation equation. The coefficient of AI is positive and highly significant, suggesting that greater AI application significantly enhances IIP. Meanwhile, the coefficient of ESG_score remains positive and significant. Taken together, these results provide regression-based evidence that AI application partially mediates the positive relationship between ESG performance and IIP. Therefore, Hypothesis H2 is supported.

Knowledge Diversification Pathway: Columns (4) and (5) examine knowledge diversification as the second mediating mechanism. Column (4) shows that ESG performance has a positive and statistically significant effect on knowledge diversification. This result suggests that stronger ESG performance facilitates the development of a more diversified technological knowledge base, potentially by encouraging broader knowledge acquisition, cross-domain learning, and the integration of heterogeneous technological resources. In Column (5), the coefficient of TD is positive and highly significant, demonstrating that greater knowledge diversification contributes significantly to sustainable innovation performance. At the same time, ESG_score remains positive and significant. These findings indicate that knowledge diversification also partially mediates the relationship between ESG performance and IIP; the evidence provides strong support for Hypothesis H3.

5.3. The Joint Contributions of AI and TD

To examine whether AI adoption and knowledge diversification make independent and non-substitutable contributions, this study estimates the outcome equation of a parallel mediation model in which AI, TD, and ESG_score are simultaneously included in the sustainable innovation equation. The results are reported in Table 9.

Table 9. Outcome Equation of the Parallel Mediation Model.

Variable	IIP
	(1)
ESG_score	0.016 ***
	(0.002)
AI	0.040 ***
	(0.012)
TD	0.075 ***
	(0.012)
Controls	Yes
Firm FE	Yes
Year FE	Yes
N	8122
R2	0.379

Note: HC1 heteroskedasticity-robust standard errors are reported in parentheses. AI technology adoption and knowledge diversity (TD) are simultaneously included in the model. The model includes control variables and controls for firm and year fixed effects. *** denotes statistical significance at the 1% level.

In the subsample in which complete observations are available for both mediators and all model variables, both AI adoption and knowledge diversification remain independently significant. The direct effect of ESG_score also remains positive and robust. This result suggests that AI and TD capture distinct incremental information concerning the process through which ESG enhances IIP. The results jointly support the proposed dual-engine transmission framework. Technological capability, represented by AI adoption, and knowledge structure, represented by knowledge diversification, constitute two independent and complementary channels through which ESG promotes IIP. AI adoption remains positively and significantly associated with IIP after controlling for knowledge diversification, while knowledge diversification also retains a positive and significant coefficient in the joint specification. Thus, the AI effect is not absorbed or replaced by the knowledge diversification pathway.

5.4. Overall Conclusions of the Mechanism Analysis

The mechanism analysis reveals a clear dual-engine, partial-mediation structure. ESG performance is transformed into IIP through two independent and statistically significant pathways: increased application of artificial intelligence technologies and greater diversification of the firm’s technological knowledge base. At the same time, a significant direct ESG effect remains that cannot be fully explained by these two channels. This finding directly addresses the central research question raised in the introduction: ESG performance can indeed be converted into substantive innovation productivity, and this transformation occurs through multiple, parallel, and partially observable mechanisms.

From the perspective of theoretical dialogue, the dual-engine mechanism identified in this study contributes to the existing ESG–innovation literature in two important ways. First, prior studies generally explain the effect of ESG on innovation through mechanisms relatively external to firms’ internal knowledge-production processes, including the alleviation of financing constraints and the accumulation of reputational assets [33,34]. In contrast, AI adoption and knowledge diversification represent capability-based mechanisms that are more deeply embedded in firms’ internal knowledge-production processes. The present study therefore extends the mechanism literature by introducing explanatory pathways grounded in the knowledge-based view [55].

This finding implies that future research examining the increasingly prominent proposition that “digital transformation leads to innovation empowerment” should distinguish between the adoption of individual digital technologies and the broader knowledge ecosystem within which those technologies are embedded. Treating technological adoption and knowledge-system transformation as equivalent may obscure the more fundamental organizational and knowledge-based foundations of sustainable innovation.

6. Extended Analysis

This part systematically examines the boundary conditions of the innovation-enhancing effect of ESG across five dimensions: R&D intensity, environmental performance, ownership type, firm size, and industry type.

6.1. Heterogeneity Analysis

6.1.1. Heterogeneity by R&D Intensity

The heterogeneity analysis based on R&D intensity provides the most direct and compelling evidence for the absorptive capacity mechanism and is reported in Table 10. Among firms with R&D intensity above the sample median, the effect of ESG score on IIP is positive and highly significant. In contrast, among firms with below-median R&D intensity, the estimated effect is statistically indistinguishable from zero. This disparity further suggests that the variation in IIP among low-R&D firms is largely driven by factors outside the present model and that ESG-related resources cannot be effectively transformed into innovation output when firms lack an adequate R&D foundation.

Table 10. Results of Heterogeneity Analysis by R&D Intensity.

Variables	High-R&D Group	Low-R&D Group
	(1)	(2)
ESG_score	0.015 ***	0.004
	(0.001)	(0.003)
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes
N	29,125	4835
R2	0.359	0.032

Note: *** indicates significance at the 1%levels.

This finding has an important policy implication: ESG investment is not, by itself, a sufficient condition for innovation capability. Firms must also possess a corresponding base of R&D investment and technological absorptive capacity in order to convert ESG-related resources into IIP.

6.1.2. Heterogeneity in the E_Score–IIP Relationship: High- and Low-E_Score Groups

The effect of environmental performance on sustainable innovation capability also exhibits a pronounced threshold pattern. Only among firms with environmental scores above the median does E_score have a significantly positive effect on IIP. For firms in the low-E_score group, the estimated effect is nearly zero and statistically insignificant. This pattern where the effect emerges only at higher levels of environmental performance suggests that environmental management capability must exceed a certain threshold before it can generate technological spillovers and contribute meaningfully to sustainable innovation. Fragmented or preliminary environmental compliance investments appear insufficient to establish the technological capability base required for innovation. Together with the R&D intensity threshold identified in Section 6.1.1, this finding of Table 11 provides dual empirical support for the proposed absorptive capacity constraint.

Table 11. Heterogeneity by E_score: Subsamples Split at the Median.

Variables	High E_Score	Low E_Score
	(1)	(2)
E_score	0.013 ***	0.001
	(0.001)	(0.002)
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes
N	17,707	16,253
R2	0.309	0.295

Note: *** indicates significance at the 1%levels.

6.1.3. Ownership Heterogeneity

The present study finds that the ESG coefficients are significantly positive for both state-owned and non-state-owned firms. The difference lies mainly in magnitude: the coefficient for non-state-owned enterprises is slightly larger than that for state-owned enterprises (0.019 versus 0.013). These results of Table 12 indicates that the innovation-enhancing effect of ESG is broadly applicable across ownership types. Unlike findings reported in some studies of supply-chain outcomes or short-term financial performance, no ownership-based reversal in the direction of the ESG performance effect is observed. Nevertheless, ESG performance driven by stronger market competition and efficiency incentives among non-state-owned firms appears to exhibit somewhat greater innovation conversion efficiency.

Table 12. Ownership Heterogeneity: State-Owned and Non-State-Owned Enterprises.

Variable	SOEs	Non-SOEs
	(1)	(2)
ESG_score	0.013 ***	0.019 ***
	(0.002)	(0.001)
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes
N	10,291	23,669
R2	0.315	0.334

Note: *** indicates significance at the 1%levels.

6.1.4. Firm Size Heterogeneity

Table 13 examines whether the effect of ESG on sustainable innovation varies with firm size. The subgroup estimates show that the coefficients of ESG_score are positive and statistically significant for both large and small firms, although the estimated ESG effect is slightly larger among small firms. This finding suggests that the innovation-enhancing effect of ESG is relatively homogeneous across firm size. In contrast to the pronounced heterogeneity associated with R&D intensity documented in Section 6.1.1, the results further suggest that absorptive capacity, rather than resource scale alone, constitutes a more important boundary condition for converting ESG performance into IIP.

Table 13. Results of Heterogeneity Analysis by Firm Size.

Variables	Large Firms	Small Firms
	(1)	(2)
ESG_score	0.013 ***	0.014 ***
	(0.001)	(0.001)
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes
N	18,152	15,808
R2	0.277	0.185

Note: *** indicates significance at the 1%levels.

6.1.5. Industry Heterogeneity: Manufacturing Versus Non-Manufacturing Firms

Table 14 examines whether the innovation-enhancing effect of ESG differs between manufacturing and non-manufacturing firms. The subgroup estimates show that the coefficients of ESG_score are positive and statistically significant in both manufacturing and non-manufacturing firms, indicating that ESG performance promotes sustainable innovation capability across both industry groups. These results suggest that the positive effect of ESG on sustainable innovation is not confined to manufacturing firms but exhibits broad cross-industry applicability.

Table 14. Results of Industry Heterogeneity.

Variables	Manufacturing Firms	Non-Manufacturing Firms
	(1)	(2)
ESG_score	0.017 ***	0.016 ***
	(0.001)	(0.002)
Controls	Yes	Yes
Firm FE	Yes	Yes
Year FE	Yes	Yes
N	25,360	8600
R2	0.363	0.233

Note: *** indicates significance at the 1%levels.

6.2. Additional Analyses

To further examine whether the innovation-enhancing effect of ESG varies with changes in the external business environment, this study conducts a set of expanded analyses focusing on external shocks, temporal heterogeneity, and changes in ESG performance. First, a composite external-shock indicator is constructed as manufacturing firm × post-2018 period, capturing the relatively stronger exposure of Chinese manufacturing firms to the China–U.S. trade tensions and the associated restructuring of global supply chains. The interaction term between this shock indicator and the mean-centered ESG score (ESG_shock) is then introduced into the baseline model. In addition, the sample period is divided into five subperiods—2009–2012, 2013–2015, 2016–2018, 2019–2021, and 2022–2025—to investigate whether the ESG–sustainable innovation relationship changes over time. Finally, the level of ESG performance and its year-to-year change (ΔESG) are simultaneously included to distinguish the effect of accumulated ESG capabilities from that of short-term adjustments in ESG performance.

The results reported in Table 15 provide evidence that the effect of ESG performance on IIP is conditional on the external environment. In the external-shock specification, the coefficient of the mean-centered ESG score remains positive and highly significant, whereas the coefficient of the shock indicator itself is negative but statistically insignificant. More importantly, the coefficient on ESG_shock is positive and statistically significant. This result indicates that the positive marginal effect of ESG on sustainable innovation becomes stronger for manufacturing firms after 2018, when they were more intensively exposed to trade frictions, supply-chain disruptions, and heightened external uncertainty. Rather than merely generating compliance or reputational benefits, ESG-related resources therefore appear to acquire greater strategic value when firms face substantial external pressure. Firms with stronger ESG capabilities may be better positioned to mobilize stakeholder relationships, environmental technologies, organizational resources, and risk-management capabilities to sustain innovation under adverse conditions.

Table 15. Expanded Analysis Results.

Variable	ESG × Manufacturing × Post-2018	Period-Specific ESG Coefficient	ESG Level + ΔESG
ESG_score	0.015 ***
	(0.001)
shock	−0.035
	(0.023)
ESG_shock	0.003 **
	(0.002)
ESG_score (2009–2012)		0.008
		(0.006)
ESG_score (2013–2015)		0.019 ***
		(0.003)
ESG_score (2016–2018)		0.005 **
		(0.002)
ESG_score (2019–2021)		0.000
		(0.001)
ESG_score (2022–2025)		0.005 ***
		(0.001)
ESG_score			0.021 ***
			(0.001)
ΔESG			−0.007 ***
			(0.001)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
Year FE	Yes	Yes	Yes

Note: Column (1) examines the interaction effect between ESG performance and the manufacturing × post-2018 shock. The middle section reports the results from period-specific regressions. The final column simultaneously includes the level of ESG performance and the change in ESG (ΔESG). HC1 heteroskedasticity-robust standard errors are reported in parentheses. All models include control variables and control for firm and year fixed effects. *** and ** denote statistical significance at the 1% and 5% levels, respectively.

The subperiod regressions further reveal substantial temporal heterogeneity in the ESG–innovation relationship. The coefficient of ESG_score is positive but statistically insignificant during 2009–2012, becomes strongly positive during 2013–2015, remains positive and significant during 2016–2018, and becomes statistically insignificant during 2019–2021. The latter period coincides with an unusually turbulent external environment characterized by escalating trade tensions, the COVID-19 pandemic, and severe disruptions to global production and supply chains. Under such extreme uncertainty, firms may have temporarily redirected financial and managerial resources toward liquidity preservation, operational continuity, and short-term risk management, thereby weakening the immediate conversion of ESG advantages into sustainable innovation investment. By contrast, the ESG coefficient becomes positive and highly significant again during 2022–2025, suggesting that the innovation-enhancing role of ESG re-emerged as firms adapted to the new institutional and competitive environment.

The final specification provides an additional distinction between the level and change effects of ESG. When ESG performance and its annual change are simultaneously included, the coefficient of ESG_score remains strongly positive, whereas the coefficient of ΔESG is significantly negative. This contrast suggests that sustainable innovation benefits primarily from the cumulative stock of ESG-related capabilities rather than from abrupt short-term improvements in ESG ratings. Building environmental technologies, stakeholder relationships, organizational routines, and governance capabilities requires time and persistent investment. Rapid changes in ESG performance may instead entail adjustment costs, resource reallocation, or short-term compliance expenditures that temporarily crowd out innovation resources. Therefore, a high and sustained level of ESG performance appears more conducive to sustainable innovation than short-lived increases in ESG scores.

Taken together, these additional analyses show that the relationship between ESG performance and IIP is neither mechanically constant over time nor independent of the external environment. ESG generates stronger innovation benefits when firms face meaningful external competitive pressure, but its effectiveness may be temporarily weakened under exceptionally severe systemic disruptions. At the same time, the results emphasize the importance of persistent ESG capability accumulation rather than short-term ESG adjustments. These findings extend the baseline results by demonstrating that ESG functions as a strategic organizational resource whose innovation value depends on both the external environment and the maturity of firms‘ accumulated ESG capabilities.

7. Conclusions and Policy Implications

7.1. Main Conclusions

Using panel data of Chinese A-share listed firms from 2009 to 2025, this study systematically examines the relationship between ESG performance and corporate sustainable innovation capability and explores the underlying transmission mechanisms through artificial intelligence (AI) adoption and knowledge diversification. Several major conclusions are obtained. These findings are not independent of one another; rather, they collectively address the central research question of whether ESG is associated with sustainable innovation capability, through which mechanisms this transformation occurs, and under what conditions such transformation becomes more effective. Together, they constitute a coherent chain of evidence, progressing from identifying the overall effect, to uncovering transmission mechanisms, and finally to clarifying boundary conditions.

First, overall ESG performance has a robust and significantly positive effect on corporate sustainable innovation capability. This finding remains consistent across robustness checks employing more stringent winsorization, a one-period lagged ESG variable, exclusion of the COVID-19 period, dynamic panel estimation using system GMM, and the Heckman two-step correction for sample selection bias.

Second, the environmental (E), social (S), and governance (G) dimensions all contribute significantly and positively to sustainable innovation capability, indicating that no “aggregation fallacy” exists whereby individual dimensions offset one another. However, their marginal effects differ substantially, following the ranking: Social > Governance > Environmental. The social dimension contributes more fundamentally and broadly to sustainable innovation through human capital accumulation and knowledge collaboration networks than the environmental dimension, which primarily promotes engineering and technological breakthroughs.

Third, AI adoption and knowledge diversification operate as two parallel and complementary mediating channels. Both mechanisms remain independently significant in the joint model; AI adoption and knowledge diversification exhibit a relatively mediating role, suggesting that ESG enhances innovation not only through digital technology adoption but also, more fundamentally, through the restructuring and expansion of firms’ knowledge bases.

Fourth, the innovation-enhancing effect of ESG shows patterns potentially consistent with an absorptive-capacity interpretation. ESG exerts a strong and significant positive impact among firms with R&D intensity, whereas its effect is statistically indistinguishable from zero among firms below the median. A similar threshold effect is observed for environmental performance, with firms possessing higher E scores exhibiting significantly stronger effects than their lower-scoring counterparts. In contrast, the ESG effect remains broadly consistent across ownership types, firm sizes, and industries, highlighting absorptive capacity as a particularly important boundary condition for converting ESG-related resources into innovation outcomes.

Fifth, the results suggest that short-term changes in ESG performance may be associated with adjustment costs. The coefficient on the annual change in ESG performance (ΔESG) is significantly negative. Furthermore, it identifies a complete dynamic trajectory in which innovation performance temporarily declines before ESG improvement events but subsequently accelerates. The consistency between these two independent approaches indicates that the transformation of ESG into sustainable innovation capability is not immediate but instead involves real resource reallocation costs and should therefore be evaluated from a medium- to long-term perspective. Additional analyses further reveal temporal variation in the ESG–innovation relationship, suggesting that its magnitude may depend on broader institutional and macroeconomic conditions.

7.2. Policy Implications

Drawing on the empirical findings, this study offers several policy implications for regulators and firms to improve the effectiveness of ESG performance and strengthen its transformation into corporate sustainable innovation capability.

For regulators: to begin with, strengthen firms‘ absorptive capacity as a central pillar for coordinating ESG and innovation policies. One of the most policy-relevant findings of this study is that the innovation-enhancing effect of ESG may vary with firms’ R&D capabilities, proxied by R&D intensity. Accordingly, when designing ESG disclosure and sustainability policies, regulators should complement such initiatives with innovation-support measures targeted at firms with relatively weak R&D capabilities, particularly firms with limited R&D resources. Differentiated R&D tax incentives, innovation vouchers, and support for technology upgrading could help these firms develop the capabilities required to transform ESG-related resources into substantive innovation outcomes. Such measures would reduce the risk of emphasizing ESG compliance without simultaneously strengthening firms’ capacity to convert ESG investments into productive innovation.

Additionally, promote differentiated ESG strategies that account for firms’ innovation and technological capabilities. The findings suggest that firms with relatively limited innovation capabilities may benefit more from strengthening the social foundations of innovation, including human capital development, employee training, and collaborative networks, whereas firms with stronger technological and absorptive capacities may be better positioned to translate environmental investments into innovation outcomes. Industry associations and ESG rating agencies should therefore provide more differentiated guidance that recognizes heterogeneity in firms’ technological foundations rather than relying exclusively on uniform ESG priorities.

Finally, promote stronger coordination between ESG initiatives and firms’ digital and knowledge-development strategies. The mediation results show that AI application and knowledge diversification constitute two complementary channels through which ESG contributes to sustainable innovation. Policymakers and industry associations should therefore encourage firms to integrate ESG initiatives with digital transformation, knowledge management, university–industry collaboration, cross-industry technology transfer, and open innovation programs. Such coordination can help firms transform ESG-related resources into both technological capabilities and broader knowledge bases, thereby strengthening the innovation returns to ESG performance.

For firms: first, adopt a long-term approach to ESG performance building rather than pursuing campaign-style improvements in ESG ratings. The results indicate that rapid short-term improvements in ESG performance may impose temporary adjustment costs on innovation, whereas sustained ESG performance is associated with more stable innovation benefits over time. Firms should therefore integrate ESG objectives into long-term strategic planning and avoid reallocating resources excessively toward short-term rating improvements. A gradual approach to ESG capability building can provide sufficient time for organizational learning and efficient resource reallocation while reducing potential crowding-out effects on sustainable innovation capability.

Second, strengthen internal R&D and knowledge-integration capabilities alongside ESG investment. The heterogeneity results indicate that ESG alone is insufficient to generate innovation benefits when firms lack an adequate R&D foundation. Firms should therefore treat R&D investment, knowledge acquisition, and cross-functional knowledge integration as complementary capabilities that enable ESG-related resources to be transformed into sustainable innovation. In particular, firms with relatively weak absorptive capacity should prioritize strengthening their technological and knowledge foundations before expecting substantial innovation returns from additional ESG performance.

Third, differentiate ESG dimensions and their associated technological pathways rather than adopting indiscriminate digitalization strategies. Firms should therefore avoid assuming that AI adoption will generate equally strong innovation benefits across all ESG activities. Instead, they may prioritize AI-enabled applications in areas closely related to the social dimension, such as digital employee training, human-capital management, and digital supply-chain collaboration, while adopting technology strategies tailored to the distinct characteristics of environmental and governance activities.

7.3. Limitations and Future Research

The paper still has several limitations: (1) Although the Integrated Innovation Performance (IIP) provides a more comprehensive measure of sustainable innovation than simple patent counts by capturing both innovation efficiency and sustainability, it cannot distinguish among different innovation types, such as incremental versus radical innovation. Future research may combine patent text mining techniques to provide a more refined classification of innovation activities. (2) The measurement of AI adoption is constrained by the availability and scope of firm-level AI data. The current measure primarily captures observable AI-related technological activities and may not fully reflect AI applications embedded in firms‘ production processes, organizational management, and decision-making systems. Consequently, the estimated mediating effect may not capture the full extent to which AI contributes to the transformation of ESG performance into sustainable innovation capability. (3) The identification of short-term ESG adjustment costs relies primarily on indirect evidence derived from the event study and the ΔESG coefficient. More sophisticated dynamic structural models capable of estimating the magnitude and persistence of adjustment costs remain to be developed in future research.

Future research may proceed in three directions. First, patent text mining techniques can be employed to distinguish incremental from radical sustainable innovation and examine whether different ESG dimensions exert heterogeneous effects across innovation types. Second, as firm-level AI application data become increasingly available, future studies may investigate the differential roles of specific AI technologies, such as generative AI, predictive AI, and automation AI, within the ESG–innovation transmission framework. Third, the analytical framework developed in this study may be extended to cross-country settings to examine the generalizability and boundary conditions of the absorptive capacity threshold mechanism under different institutional environments, including the United States, the European Union, and other emerging economies.