When AI Amplifies Negative Echoes: CEO–TMT Faultlines, Eco-Attention, and the Hindrance of Green Innovation

Abstract

Digital technologies are increasingly reshaping how top management teams (TMTs) make strategic decisions regarding green innovation. Although prior research has examined the roles of TMT characteristics and artificial intelligence (AI), it remains unclear how TMT internal structures influence green innovation through organizational attention and how AI shapes this process. In particular, CEO–TMT faultlines, reflecting divisions in experiences, roles, and authority, may affect how environmental issues are recognized and prioritized, especially in AI-enabled contexts where complex information processing can amplify internal divisions. Drawing on the attention-based view (ABV), this study examines the cognitive mechanism linking CEO–TMT faultlines to green innovation. Using a panel dataset of 35,347 firm-year observations from 2010 to 2023, we find that CEO–TMT faultlines negatively affect green innovation through reduced eco-attention. Moreover, AI technology strengthens the negative relationship between CEO–TMT faultlines and eco-attention, thereby deepening the negative indirect effect on green innovation. These findings highlight the role of organizational attention in linking TMT structures to strategic outcomes and suggest that AI adoption may reinforce, rather than mitigate, the challenges arising from internal divisions.

1. Introduction

Green innovation is a critical strategic choice for firms to achieve sustainable development [1,2], but it entails considerable challenges. Compared with conventional innovation, green innovation involves higher risks and investment while requiring firms to constantly balance economic performance and environmental responsibility [3,4]. This makes the process of determining whether and how to implement green innovation strategies particularly complex [5], as it has been shown to be highly susceptible to the quality of interactions within the top management team (TMT) [6,7]. However, existing research has mainly examined the CEO as an individual or the TMT as a unitary entity [8,9,10], overlooking the critical differences between the CEO and other TMT members in terms of demographics, role, and authority, as well as how these shape strategic decisions [11,12,13]. These differences reflect divergent cognitions [14], giving rise to CEO–TMT faultlines, defined as the degree of differentiation between the CEO and other TMT members across multiple characteristics [7,15]. Accordingly, this study addresses the following research question: How do CEO–TMT faultlines affect corporate green innovation?

Although existing research suggests that CEO–TMT differences shape their interactions and strategic decision-making [16,17], the psychological and cognitive mechanisms underlying these shaping processes remain underexplored [18,19]. Faced with resource constraints, firms cannot attend to all strategic issues simultaneously and must instead selectively allocate attention to those deemed most consequential [20]. Prior research has established that this allocation process is far from uniform across TMT members, as differences in experiential backgrounds and functional expertise systematically shape what executives notice and prioritize [21,22]. Therefore, building on the attention-based view (ABV), we propose that eco-attention [23], defined as the degree to which organizations focus on and prioritize environmental issues, serves as a crucial mediating mechanism through which CEO–TMT faultlines influence green innovation. In doing so, our central research question can be more precisely articulated as follows: How do CEO–TMT faultlines shape corporate green innovation through the attention-allocation mechanism?

The rapid advancement of digital technologies has further shaped this decision-making process. Artificial intelligence (AI), in particular, has become a vital tool in strategic decision-making, reshaping how organizations allocate attention [24,25,26]. The rise in AI technology offers a new contextual condition for understanding the relationship between CEO–TMT faultlines and green innovation. From a socio-technical perspective, AI is not merely a neutral analytical tool, as its effects largely depend on the organizational context in which it is embedded [27,28]. On the one hand, AI technology rapidly analyzes vast amounts of data and provides structured analysis [29,30], thereby alleviating information barriers caused by CEO–TMT faultlines and expanding the cognitive scope. This capability can redirect managerial attention to environmental issues. On the other hand, AI algorithms based on historical preferences may reinforce existing cognitive frames [31,32], deepening the cognitive divides between subgroups and reducing the identification of environmental opportunities and threats. We argue that under conditions of strong CEO–TMT faultlines, the latter tendency is more likely to prevail. Rather than bridging cognitive divides, AI amplifies existing fragmentation, ultimately suppressing eco-attention and weakening green innovation. Therefore, this boundary condition warrants critical inquiry.

In sum, drawing on the ABV, this study explores how CEO–TMT faultlines influence green innovation through eco-attention as well as the moderating role of AI technology. Our research makes three important contributions. First, this study enhances the understanding of TMTs’ green innovation decision-making processes and identifies the mediating role of eco-attention by revealing the cognitive mechanisms underlying how CEO–TMT faultlines affect green innovation. Prior studies have offered limited insights into how CEO–TMT interactions shape strategy decision-making processes [17,18,19], particularly regarding green innovation. By introducing eco-attention as a mediating variable, we deepen our understanding of strategy decision-making processes and extend the application of ABV in sustainability strategies. Second, we identify AI technology as an important contextual factor shaping green innovation decisions in the digital era and highlight the need for a contextualized understanding of the influence of AI on organizational strategy. While prior studies have highlighted the positive role of AI in enhancing strategic decision-making, they have often overlooked the fact that the impact of AI is highly contingent upon its users. Our study examines the interaction between AI technology and CEO–TMT faultlines, showing that AI’s effects are not universally beneficial but depend on users’ cognitive characteristics and interactions [33]. This finding challenges the dominant assumption of AI’s uniformly positive role and underscores the importance of considering the context in which it is embedded [34,35,36]. Third, this study integrates AI technology into the attention-based framework, deepening the understanding of how digital technologies reshape managerial cognition under conditions of team faultlines. Specifically, we respond to the call by Laamanen, Weiser, von Krogh and Ocasio [26] to investigate whether team dynamics and potential trust deficits amplify or distort the influence of AI on organizational attention. By integrating CEO–TMT faultlines, eco-attention, AI technology, and green innovation into a unified framework, this research not only advances the theoretical boundaries of ABV but also provides a novel perspective on how AI reshapes decision-making processes in sustainability strategy.

2. Literature Review and Hypothesis Development

2.1. Literature Review

Prior research on green innovation has accumulated substantial evidence examining individual CEO characteristics such as age and gender [10,37], as well as aggregate TMT attributes such as diversity [6,9]. Yet a closer examination reveals a key conceptual limitation in these streams of research. Treating the CEO as the sole decision-maker overlooks the collective influence of other TMT members [11,17], while treating the TMT as a homogeneous group obscures meaningful differences in authority and role between the CEO and other TMT members [38]. This has prompted a shift in focus toward the interaction between the CEO and the TMT [7].

This shift is theoretically well-grounded. The CEO occupies a structurally distinct position, simultaneously relying on other TMT members for information and expertise while bearing the ultimate responsibility for strategic outcomes [39]. This duality positions the CEO as both separate from and embedded within the broader TMT. Therefore, CEO–TMT interaction offers a critical lens for understanding strategic leadership [11,40]. Compared to conventional strategic decisions, green innovation places greater demands on intra-TMT communication, as it requires both cross-disciplinary knowledge integration [41] and the coordination of diverse stakeholder interests [5,42,43]. CEO–TMT faultlines capture both the CEO’s leadership position and the TMT’s collective influence, offering a useful conceptual tool for examining this interaction [15]. Although existing work has provided initial evidence linking CEO–TMT faultlines to green innovation [7], the underlying mechanisms remain insufficiently understood.

Furthermore, the effect of TMT heterogeneity on firm strategic decisions remains contested [14,44]. Early research treated team heterogeneity as a cognitive resource, arguing that differences in members’ characteristics broaden information sources and challenge cognitive inertia, improving strategic decision quality [45,46,47]. Yet heterogeneity can equally fracture team cohesion [48]. Differences among members may give rise to informal subgroups, intensify affective conflict, and undermine the trust and collaboration necessary for effective strategy execution [49,50]. This tension extends to green technology innovation. Even within the same type of heterogeneity, findings diverge. On gender diversity, He and Jiang [51] find that a higher proportion of female directors promotes green product innovation but has no significant effect on green process innovation. Liu, Liao, Ma and Dong [6] further reveal an inverted U-shaped relationship between female executive representation and the green strategy: moderate gender diversity enhances environmental awareness, but beyond a certain threshold, the associated communication and coordination costs outweigh the informational benefits. Similarly, resource-based heterogeneity within TMTs may intensify coordination friction and impede green technology innovation [52]. We argue that one source of these inconsistencies lies in the failure to distinguish between the CEO and other TMT members, a gap that underscores the necessity of the present study.

Attention allocation serves as a pivotal cognitive mechanism through which TMTs shape corporate strategy [20,53]. Rooted in the ABV, organizational attention is conceptualized as the process by which decision-makers selectively focus on specific issues while ignoring others, a process deeply embedded in managerial cognition [20]. Within this tradition, eco-attention represents a domain-specific form of strategic attention, reflecting the extent to which environmental issues are prioritized in managerial cognition and organizational decision-making [23]. The structural configuration of the TMT serves as the social architecture that channels this attention. Specifically, the structure and interaction of TMTs shape team communication, which in turn influences the formation of strategic priorities [54,55] and determines how organizational attention is distributed [45]. Just as TMTs focused on sustainability can pivot a firm’s attention toward corporate social responsibility [56], while TMTs focused on digitalization can shift organizational attention toward digital innovation [57]. Therefore, we posit that the CEO–TMT faultlines influence green innovation by altering eco-attention.

AI has increasingly attracted scholarly attention as both a data-analytic tool and a cognitive technology that shapes organizational decision-making processes [58,59,60]. From a socio-technical perspective, the effects of AI are understood to depend on the organizational contexts in which it is embedded, as well as the social processes and relational dynamics surrounding its use [61,62]. Building on this perspective, prior studies suggest that AI interacts with human cognition to influence how information is processed, interpreted, and utilized in decision-making [63]. In particular, AI’s capacity to process large volumes of data and generate structured insights has been shown to reshape the scope of issues that organizations attend to and how these issues are evaluated [25,28], thereby affecting patterns of attention allocation. Consistent with this view, a growing body of research finds that AI adoption can facilitate corporate green innovation by enhancing information processing capabilities and decision quality [64,65,66]. At the same time, emerging studies highlight that the effects of AI are not uniformly positive. Potential challenges associated with AI use include information overload [67], algorithmic biases [68,69], and the risk of inappropriate reliance or misuse by decision makers [70]. These findings suggest that the outcomes of AI adoption depend on how decision makers interpret and integrate AI-generated insights [26]. Taken together, existing research indicates that AI plays a complex role in shaping organizational attention and strategic decision-making, with its effects contingent on both technological characteristics and social contexts.

2.2. CEO–TMT Faultlines and Green Innovation

CEOs and other TMT members differ significantly in power and formal status [13], which amplifies the potential conflicts and tensions that may arise from their characteristic differences. While CEO–TMT faultlines can enrich information sources and expand skill ranges through increased member diversity, this positive effect is undermined by asymmetry in authority and status [71]. The CEO’s unique leadership position further reinforces this imbalance, making the CEO–TMT faultlines more likely to exert negative effects on corporate strategy.

CEO–TMT faultlines create a sense of in-group and out-group membership within the TMT, where other TMT members tend to form a subgroup, whereas the CEO is perceived as an individual separate from this subgroup [72]. Subgroup polarization fosters in-group favoritism and out-group antagonism [73], eroding team cohesion and inter-member trust [74,75]. This adversarial climate drives strategic myopia, because CEOs may prioritize short-term, low-risk objectives to override potential threats from other TMT members, which could increase their downside risks [76]. Given that green innovation is inherently high-risk and long-term [77], CEO–TMT faultlines reduce the likelihood of TMTs’ consensus on green innovation.

Additionally, CEO–TMT faultlines can engender emotional conflict and impede effective communication and information integration [14]. While CEOs strive to coordinate communication within fragmented teams [14,72], other TMT members withhold critical insights into green innovation because of distrust [49]. The inherent risks and uncertain returns of green innovation heighten risk aversion within TMTs, thereby inhibiting creative expression. Although faultlines introduce enriched viewpoints and insights, the resultant communication inefficiencies outweigh the potential benefits, particularly in resolving complex technological challenges [78]. The complex interdisciplinary knowledge required for green innovation further amplifies these challenges [41]. Moreover, the dual externalities of green innovation strengthen conservative tendencies within TMTs, restricting their inclination to innovation. Therefore, we propose the following hypotheses:

Hypothesis 1.

CEO–TMT faultlines negatively influence green innovation.

2.3. Mediating Role of Eco-Attention

The implementation of green innovation depends on whether firms prioritize environmental issues as a strategic concern. Given its high risks, substantial investments, and partially non-internalizable costs, green innovation cannot gain adequate resource support without paying attention to environmental issues [79,80]. Thus, we posit that eco-attention provides a critical mechanism for understanding how CEO–TMT faultlines influence green innovation.

Strong CEO–TMT faultlines can hinder the integration of environmental issues into the TMT’s strategic focus. Given the limited cognitive capacity of decision makers, CEOs rely on the collaborative input of other TMT members for comprehensive analysis and decision support [39,40]. However, in the presence of strong CEO–TMT faultlines, emotional conflict and lack of trust hinder communication and erode mutual identification [74,81]. CEOs disproportionately rely on their own judgments and diminish their dependence on inputs from other TMT members [82]. Environmental issues are particularly vulnerable to this narrowing process. Unlike financial performance indicators, which are salient, quantifiable, and closely tied to short-term incentives, environmental issues are typically long-term [5], externally oriented [80], and harder to evaluate within conventional performance frameworks. When collective communication weakens, such issues are more likely to be pushed out of the strategic agenda, as they lack the immediacy and measurability needed to attract managerial attention under conditions of limited information sharing. Furthermore, the presence of affective conflict leads TMT members to avoid greater strategic responsibility and risk-taking [49]. TMTs tend to prioritize issues that enhance financial performance over environmental concerns. In this context, a reduction in shared information processing and collaborative deliberation decreases the likelihood of the TMT devoting attention to environmental issues.

Furthermore, the CEO–TMT faultlines impede the acknowledgment of the strategic significance of environmental issues. When CEO–TMT faultlines are weak, boundaries between CEOs and other TMT members become less salient, fostering frequent communication and smoother information exchange [72,83]. Such interactions enable the TMT to better identify environmental opportunities and threats and develop a more comprehensive environmental understanding [84], thereby shifting corporate goals from a narrow focus on economic performance to a multidimensional balance that incorporates economic, environmental, and social performance [85]. Conversely, as CEO–TMT faultlines deepen, recognition of the long-term value and strategic importance of environmental issues tends to diminish, reducing the likelihood that such issues will be integrated into the core strategy [86].

We argue that organizations that prioritize environmental issues are more likely to focus on environmental concerns, allocate greater eco-attention, and treat green innovation as a key strategy to address environmental challenges. However, the CEO–TMT faultlines weaken the allocation of eco-attention within firms. Therefore, we propose the following hypothesis:

Hypothesis 2.

Eco-attention mediates the relationship between CEO–TMT faultlines and green innovation.

2.4. Moderating Role of AI Technology

With the continuous advancement of digital technologies, AI technology has become increasingly influential in firms’ strategic decision-making, reshaping how organizations allocate attention [29,87]. Although AI enhances information processing efficiency, its actual effect on corporate strategy depends largely on the organizational context in which it is embedded [24,88]. We argue that in teams with strong CEO–TMT faultlines, the power asymmetry between the CEO and other TMT members undermines the potential benefits of AI. AI collectively suppresses eco-attention through three distinct mechanisms: selective information adoption, information overload, and reduced collaborative deliberation.

First, in the presence of strong CEO–TMT faultlines, AI technology encourages selective information adoption, distorting how strategic information is interpreted and prioritized within the team. Although AI technology provides standardized and objective analytical outputs, it does not ensure that TMTs interpret and use such information objectively [89]. In teams with CEO–TMT faultlines, distinct subgroups may interpret AI-generated information through different strategic perspectives, leading to various priorities [16]. Under these conditions, AI technology outputs are unlikely to serve as neutral arbiters of competing viewpoints. Instead, subgroups may invoke the perceived authority of AI analyses to support their own positions. Executives are therefore more likely to attend to AI-generated insights that align with their existing preferences while discounting information that challenges their views [90]. Such selective use of seemingly objective data can reinforce cognitive divisions between the CEO and other TMT members and narrow the scope of strategic discussions. As a result, issues that deviate from dominant strategic priorities, such as environmental concerns, are less likely to enter the core strategic agenda and thus receive limited attention.

Second, AI technology may reduce collaboration in teams characterized by strong CEO–TMT faultlines by enabling greater decision-making autonomy within subgroups. In such contexts, AI technology can inadvertently weaken collective decision-making by substituting for the integrative processes that typically support strategic consensus [91]. AI technology provides each subgroup with a credible and internally consistent analytical basis for forming independent judgments, thereby reducing the need for cross-subgroup consultation. When interpersonal trust is limited and communication barriers exist, TMT members may increasingly depend on AI technology for independent decision-making [69] rather than collaborative decision-making. This independent mode of decision-making exacerbates the communication barriers created by CEO–TMT faultlines, making it more difficult for TMTs to integrate viewpoints and establish strategic consensus [59]. As a result, the top management team’s ability to allocate attention broadly across competing issues is weakened. Consistent with this view, Glickman and Sharot [31] suggest that human–AI interactions can reinforce existing cognitive biases. Therefore, AI technology exacerbates strategic judgment discrepancies within the team caused by CEO–TMT faultlines, making it more difficult for TMTs to reconcile perspectives and develop strategic priorities on complex, long-term issues, such as green innovation. Strategic issues such as environmental concerns, which are complex and long-term in nature and typically require integrative deliberation to gain support, are therefore further marginalized.

Finally, AI technology generates information overload that fragmented teams cannot collectively process, further diverting attention away from environmental issues. AI technology substantially increases both the volume and complexity of information available to top management teams [24]. When robust communicative and integrative mechanisms exist within the team, teams may be able to collectively process and absorb this information, supporting the identification of strategic priorities. However, CEO–TMT faultlines weaken these integration mechanisms by creating communication barriers and cognitive divergences among team members [14]. Given the constraints of communication pressures and limited cognitive resources, the increased analytical outputs generated by AI technology may lead to information overload rather than enhanced collective understanding [67,92]. Faced with competing informational signals and limited team integration, executives may prioritize short-term operational concerns over complex and long-term issues. As a result, environmental issues, due to their long-term and external nature, are more likely to be marginalized and fail to receive strategic priority. Therefore, AI technology weakens the ability of teams with CEO–TMT faultlines to identify and focus on environmental issues. Therefore, we propose the following hypothesis:

Hypothesis 3.

AI technology negatively moderates the relationship between CEO–TMT faultlines and eco-attention.

Based on the analyses above, we build a theoretical model in Figure 1.

3. Method and Data

3.1. Data

To test the proposed hypotheses, we employed a sample of Chinese A-share listed firms covering the period from 2010 to 2023. To account for the time-lagged nature of green innovation and to alleviate potential reverse causality concerns, we adopt a one-year lead structure in which independent and control variables are measured at time t and the dependent variable at time t + 1. Accordingly, the independent variables cover the period from 2010 to 2023, while the dependent variable spans 2011 to 2024. The data used in this study were collected from multiple sources. The Chinese Research Data Services Platform (CNRDS) database provides the information required to calculate firms’ green innovation. Data on CEO–TMT faultlines were constructed using executive biographical information retrieved from the China Stock Market and Accounting Research (CSMAR) database and were supplemented with manual collection from corporate disclosures to address missing records. Data on eco-attention were obtained through textual analysis of the Management Discussion and Analysis (MD&A) sections of firms’ annual reports. Data on AI technology and all control variables were obtained from the CSMAR database. We used Stata 18.0 to process and analyze the data.

Based on previous research [93], we employed exclusion criteria for sample selection to enhance clarity and minimize potential bias. Specifically, we excluded the following types of firms: (a) those labeled as “ST” in the respective year; (b) firms operating in the financial industry; (c) firms that did not publish Corporate Social Responsibility (CSR) reports in the respective year; and (d) firms with missing data for the key variables. To avoid the influence of extreme values on the results, we winsorized the continuous variables at 1% and 99% [9]. The final sample consisted of 35,347 firm-year observations.

3.2. Variable Measurement

3.2.1. CEO–TMT Faultlines

The independent variable in our study is the CEO–TMT faultlines (Faultlines). In this study, following the theoretical conceptualization of the CEO as a structurally distinct actor, the CEO is treated as one subgroup while the remaining TMT members constitute the other [11,15]. This study measures CEO–TMT faultlines by using the Faultline Strength algorithm [48] as utilized in Zhang, Ayoko and Liang [15]. This algorithm measures the degree to which team members can be divided into distinct subgroups based on the alignment of multiple demographic attributes simultaneously, with higher values indicating greater demographic separation between the CEO and other TMT members. Importantly, this algorithm requires all demographic dimensions to be coded as categorical variables, which informed our operationalization of each characteristic [48]. Drawing on existing research [49,81,94], we operationalize faultlines using five demographic characteristics known to have a significant impact on CEO–TMT interactions: gender, age, tenure, education, and functional background. To meet the algorithm’s requirement, all demographic dimensions are coded as categorical variables [48]. Specifically, gender was represented by a categorical variable coded 1 for males and 2 for females. Age was categorized into five groups: 30 and below, 31–40, 41–50, 51–60, and over 60 years. Tenure was classified into four categories: less than two years, two to four years, five to six years, and more than six years. Educational background was classified as secondary school and below, college degree, bachelor’s degree, master’s degree, doctoral degree, or others. Functional background was classified as production and technical, marketing and sales, financial, legal, operations, and others. These categorical coding choices were adopted directly from prior published research on CEO–TMT faultlines and upper echelons theory [15,45,94], ensuring comparability with existing literature rather than reflecting arbitrary choices.

The measurement of CEO–TMT faultlines proceeds in two sequential steps, involving the calculation of an internal subgroup alignment score (IA) and a cross-subgroup alignment index (CGAI).

$${\mathrm{IA}}_{\mathrm{TMT}/\mathrm{age}}={\sum \left(\right({\mathrm{O}}_{\mathrm{TMTi}}-{\mathrm{E}}_{\mathrm{TMTi}})}^2/{\mathrm{E}}_{\mathrm{TMTi}})$$

(1)

where ${\mathrm{IA}}_{\mathrm{TMT}/\mathrm{age}}$ denotes the observed internal alignment index for other TMT members across age categories. ${\mathrm{O}}_{\mathrm{TMTi}}$ represents the observed number of TMT members in the ith age category. ${\mathrm{E}}_{\mathrm{TMTi}}$ represents the expected number of TMT members in the ith age category under the assumption of a random distribution. The overall internal subgroup alignment index is then computed by averaging the internal alignment scores across all remaining attributes.

In the second step, we calculate the CGAI, which captures the degree of dissimilarity between the CEO and other TMT members. This index is derived using a cross-product approach based on frequency counts of subgroup members within each attribute category, providing a measure of the extent to which the CEO and other TMT members diverge from one another across categories. Again, taking age as an example, the cross-subgroup alignment index is computed as follows:

$${\mathrm{CGAI}}_{\mathrm{CEO}/\mathrm{TMT}/\mathrm{age}}=\sum \left({\mathrm{N}}_{\mathrm{CEOi}} + {\mathrm{N}}_{\mathrm{TMTi}}\right)/({\mathrm{N}}_{\mathrm{TMT}} \times { \mathrm{N}}_{\mathrm{CEO}})$$

(2)

where ${\mathrm{CGAI}}_{\mathrm{CEO}/\mathrm{TMT}/\mathrm{age}}$ denotes the cross-subgroup alignment index between the CEO and other TMT members on the age attribute. ${\mathrm{N}}_{\mathrm{CEOi}}$ represents the number of CEOs in the ith age category. ${\mathrm{N}}_{\mathrm{TMTi}}$ represents the number of other TMT members in the ith age category. ${\mathrm{N}}_{\mathrm{TMT}}$ denotes the total number of other TMT members. ${\mathrm{N}}_{\mathrm{CEO}}$ denotes the total number of CEOs, which equals one. The overall cross-subgroup alignment index is similarly computed by averaging across all remaining attributes.

Finally, CEO–TMT faultline strength is computed as follows:

$${\mathrm{FLS}}_{\mathrm{position}} ={ \mathrm{IA}}_{\mathrm{position}} \times (1 -{ \mathrm{CGAI}}_{\mathrm{position}})$$

(3)

In the above formula, ${\mathrm{IA} }_{\mathrm{position}}$ refers to the subgroup’s internal alignment in terms of the relevant attributes (including gender, age, tenure, educational background, and functional background) between the position of the CEO and other TMT members. ${\mathrm{CGAI} }_{\mathrm{position}}$ refers to the cross-subgroup alignment in terms of relevant attributes (including gender, age, tenure, educational background, and functional background) between the position of the CEO and other TMT members. Specifically, we choose the natural logarithm of the number of CEO–TMT faultlines plus one as a proxy variable.

3.2.2. Green Innovation

The dependent variable is Green Innovation (Green Innovation). Although various indicators have been used to measure green innovation, we employ the natural logarithm of the number of green patent applications plus one as a proxy. Green patent applications provide an objective reflection of firms’ green innovation outcomes, and capture their capabilities and investments in environmental R&D. Compared with alternative measurement approaches, green patent applications represent the most widely adopted proxy for corporate green innovation in the literature [9,93]. In addition, logarithmic transformation effectively mitigates the impact of data distribution skewness on the analytical results. Given that green innovation involves considerable technological complexity and extended R&D cycles, we therefore adopt green patent applications in period t + 1 as the dependent variable, which better reflects the lagged nature of innovation outcomes and helps alleviate potential simultaneity concerns in estimation.

3.2.3. Eco-Attention

The mediating variable in our study is eco-attention (Eco-attention), operationalized through textual analysis of corporate disclosures. Text-based measurement has been widely employed to capture managerial cognition and attention allocation [53,95], including in the context of environmental issues. The measurement procedure follows three steps. First, we identified environment-related seed words drawing on the existing literature, policy documents, and CSR reports and refined the final lexicon through iterative comparison with synonym databases. We followed Wu and Hua [96] and used their environmental issue keyword table (see

Table A1. The keywords of the environmental issue.

Environmental Issue Keywords

safe production (anquan shengchan); protection (baohu); exceeding the standard (chaobiao); ozone layer (chouyangceng); dust removal (chuchen); atmosphere (daqi); low carbon (ditan); carbon dioxide (eryanghuatan); prevention and control (fangzhi); waste gas (feiqi); discard (feiqi); wastewater (feishui); waste material (feiwu); waste residue (feizha); dust/particulate matter (fenchen); wind energy (fengneng); boiler (guolu); filtration (guolv); environmental protection (huanbao); environment (huanjing); recycling (huishou); methane (jiawan); emission reduction (jianpai); consumption reduction (jianghao); degradation (jiangjie); noise reduction (jiangzao); energy saving (jieneng); conservation/saving (jieyue); purification (jinghua); sustainable development (kechixu fazhan); renewable (kezaisheng); air (kongqi); garbage/waste (laji); waste (langfei); process redesign (liuchengzaizao); afforestation/greening (lvhua); green (lvse); energy consumption (nenghao); energy (nengyuan); emission/discharge (paifang); exhaust (paiqi); sewage discharge (paiwu); destruction/damage (pohuai); habitat cleaning (qixidiqingjie); fuel (ranliao); three industrial wastes (sanfei); ecology (shengtai); biomass (shengwuzhi); water treatment (shuichuli); acidity (suanxing); solar energy (taiyangneng); natural gas (tianranqi); soil (turang); desulfurization (tuoliu); denitrification (tuoxiao); tail gas/exhaust (weiqi); greenhouse gas (wenshiqiti); pollution (wuran); sewage (wushui); harmless (wuhai); paperless (wuzhihua); species (wuzhong); consumption (xiaohao); recycling/circulation (xunhuan); smoke and dust (yanchen); flue gas (yanqi); liquefied gas (yehuaqi); toxic (youdu); organic matter (youjiwu); residual heat (yure); reuse (zailiyong); noise (zaosheng); heavy metals (zhongjinshu); natural resources (ziranziyuan)

Note: Pinyin pronunciations are provided in parentheses.

in Appendix A). Second, we applied this lexicon to the management discussion and analysis (MD&A) sections of firms’ annual reports, counting the frequency of environment-related keywords within each disclosure. Finally, Eco-attention is measured as the ratio of the frequency of environment-related keywords to the total word count in the MD&A section [96]. For ease of interpretation, the resulting value is multiplied by 100. In addition, we constructed an alternative keyword list using a Word2Vec-based approach (see

Table A2. The keywords of the environmental issue (Word2Vec).

Environmental Issue Keywords

environmental strategy (huanjing zhanlüe); environmental impact (yingxiang huanjing); natural environment (ziran huanjing); environmental standards (huanjing biaozhun); environmental regulation (huanjing jianguan); environmental organization (huanjing zuzhi); natural resources (ziran ziyuan); environmental law (huanjing fa); environmental factors (huanjing yinsu); environmental risk (huanjing fengxian); environmental damage (huanjing pohuai); environmental governance (huanjing zhili); environmental policy (huanjing zhengce); environmental protection (huanjing baohu); environmental protection (huanbao); protect the environment (baohu huanjing); environmental monitoring (huanjing jiance); environmental management (huanjing guanli); environmental performance (huanjing jixiao); environmentally friendly (huanjing youhao); environmental issues (huanjing yiti); environmental investment (huanjing touzi); environmental indicators (huanjing zhibiao); environmental problems (huanjing wenti); environmental impact (huanjing yingxiang); environmental prevention and control (huanjing fangzhi); environmental quality (huanjing zhiliang); environmental restoration (huanjing xiufu); ecological recovery (huanjing huifu); species protection (wuzhong baohu); environmental harm (huanjing sunhai); destroy the environment (pohuai huanjing); environmental responsibility (huanjing zeren); environmental certification (huanjing renzheng); carbon dioxide (eryanghuatan); greenhouse effect (wenshixiaoying); material saving (jiecai); pollution (wuran); pollutants (wuwu); pollution control (zhiwu); pollution prevention (fangwu); dust (fenchen); floating dust (fuchen); emission reduction (jianpai); waste reduction (jianfei); pollution reduction (jianwu); energy saving (jieneng); comprehensive management (zonghe zhili); comprehensive utilization (zonghe liyong); solid waste (gufei); three wastes (sanfei); clear waters and green mountains (lüshui qingshan); green mountains and clear waters (qingshan lüshui); windbreak and sand fixation (fangfeng gusa); green (lüse); coal liquefaction (meiye); sulfides (liuhuawu); fossil fuels (huashi ranliao); low sulfur (diiliu); low carbon (ditan); carbon reduction (jiangtan); clean (qingjie); petroleum gas (shiyouqi); coalbed methane (meicengqi); combined heat and power (redianlianchan); biomass (shengwuzhi); recycling (xunhuan); low consumption (didihao); waste (feiwu); wastewater (feishui); hazardous waste (weihai feiwu); waste material (feiliao); waste liquid (feiye); noise (zaosheng); solid waste (feiqiwu); waste oil (feiyu); iron filings (tiexie); waste residue (feizha); coal slag (meizha); waste rock (feishi); tailings (weikuang); coal ash (meihui); scrap steel (feigan); fly ash (fenmeihuai); fugitive dust (yangchen); defective products (feipin); waste paper (feizhi); hazardous materials (weixianwu); waste gas (feiqi); energy consumption (nenghao); recycling (zaisheng); turning waste into treasure (bianfei weiibao); reuse (chongfu shiyong); repair and reuse (xiujiu lifei); low energy consumption (di nenghao); geothermal energy (direneng); recycling (huishou); wind energy (fengneng); solar energy (taiyangneng); photovoltaic (guangfu); hydropower (shuineng); electrical energy (dianneng); wind power generation (fengli fadian); mineral resources (kuangchan ziyuan); natural gas (tianranqi); bioenergy (shengwuneng); tidal energy (chaoxineng); clean coal (jiejingmei); energy consumption (haoneng); waste (langfei); consumption (xiaohao); energy efficiency (nengxiao); emission/discharge (paifang); sewage discharge (paiwu); conservation (jieyue); ecology (shengtai); sustainable (kechixu); consumption reduction (jianghao); water conservation (jieshui); dust removal (chuchen); wastewater treatment (feishui chuli); desulfurization (tuoliu); sewage treatment (wushui chuli); noise reduction (jiangzao); alternative energy (tidai nengyuan); clean energy (qingjie nengyuan); new energy (xin nengyuan); nature reserve (ziran baohuqu); habitat (qixidi); simplified packaging (jianhua baozhuang); degradation (jiangjie); paperless (wuzhihua); ISO14000

Note: Pinyin pronunciations are provided in parentheses.

in Appendix A) and used it to re-measure eco-attention in the robustness checks to ensure the reliability of our conclusions.

3.2.4. AI Technology

The moderating variable in this study is AI technology (AI Technology), measured as the natural logarithm of one plus the ratio of a firm’s AI-related assets to its total assets. Existing literature employs various approaches to capture firms’ AI capabilities, primarily by calculating the proportion of AI keyword frequencies in annual reports [65,97] or counting the number of AI patents [98]. However, keyword-based text analysis predominantly reflects managerial or strategic attention toward artificial intelligence rather than actual technological capability. Similarly, while AI patent counts offer insight into a firm’s research and development capacity, they capture inventive output rather than the extent to which AI technology has been substantively deployed and integrated within the firm. By contrast, the share of AI-related assets in total assets provides a more direct and objective reflection of the degree to which a firm has committed to and embedded AI technology within its operational structure. We therefore adopt the natural logarithm of one plus this ratio as our proxy for AI technology. The underlying data were sourced from the CSMAR database.

3.2.5. Control Variables

Our model controlled for a list of governance-related variables. TMT size (TMT size) was calculated as the natural logarithm of the total TMT plus one. Duality (Duality) was coded as one when the CEO also served as the firm’s board chairman. Board independence (Independent Ratio) was calculated as the ratio of independent directors to total directors. We also incorporate several firm-level control variables to account for potential confounding factors. Firm age (Firm Age) was calculated as the natural logarithm of the years the firm has been in business plus one. Firm size (Firm Size) was calculated as the natural logarithm of the total number of a firm’s employees plus one. Debt (Debt) was calculated as the ratio of total liabilities to total assets. ROA (ROA) was calculated as the ratio of net income to total assets. SOE (Ownership) was coded as one if the firm’s ultimate controlling shareholder is the state, and zero otherwise. Subsidy (Subsidy) was measured as the total amount of government subsidies received by the focal firm. Environmental regulation (Regulation) was measured as the ratio of regional environmental spending to gross domestic product (GDP). Institutional Investor (Institutional Investor) was measured as the total number of shares held by institutional investors divided by the firm’s total outstanding shares. Furthermore, we controlled for industry and year effects to account for variations specific to each industry and period. The symbol, definitions, and sources of all variables are summarized in

Table 1. Variable Measurement and Source.

Variable	Symbol	Definition	Source
CEO–TMT faultlines	Faultlines	The degree to which team members can be divided into distinct subgroups based on the alignment of multiple demographic attributes, including gender, age, tenure, education, and functional background.	CSMAR, Corporate disclosures
Green innovation	Green Innovation	The natural logarithm of the number of green patent applications plus one.	CNRDS
eco-attention	Eco-attention	The ratio of the frequency of environment-related keywords to the total word count in the MD&A section.	Firms’ annual reports
AI technology	AI Technology	The natural logarithm of one plus the share of AI-related assets in total assets.	CSMAR
TMT size	TMT Size	The natural logarithm of the total TMT plus one.	CSMAR
Duality	Duality	It was coded as one when the CEO also served as the firm’s board chairman, and zero otherwise.	CSMAR
Board independence	Independent Ratio	The ratio of independent directors to total directors.	CSMAR
Firm age	Firm Age	The natural logarithm of the years the firm has been in business plus one.	CSMAR
Firm size	Firm Size	The natural logarithm of the total number of a firm’s employees plus one.	CSMAR
Debt	Debt	The ratio of total liabilities to total assets	CSMAR
ROA	ROA	The ratio of net income to total assets.	CSMAR
SOE	Ownership	It was coded as one if the firm’s ultimate controlling shareholder is the state, and zero otherwise.	CSMAR
Subsidy	Subsidy	The total amount of government subsidies received by the focus firm.	CSMAR
Environmental regulation	Regulation	The ratio of regional environmental spending to GDP.	CSMAR
Institutional investor	Institutional Investor	The total number of shares held by institutional investors divided by the firm’s total outstanding shares.	CSMAR

.

3.3. Research Models

To investigate the mechanism of CEO–TMT faultlines in green innovation, we constructed the following models:

$${\mathrm{Green} \mathrm{Innovation}}_{\mathrm{i},\mathrm{t}+1} = {\mathrm{\alpha }}_0 + {\mathrm{\beta }}_1{\mathrm{Faultlines}}_{\mathrm{i},\mathrm{t}} + \sum {\mathrm{\beta }}_{\mathrm{j}}{\mathrm{Control}}_{\mathrm{i},\mathrm{t}} +{ \mathrm{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(4)

$${\mathrm{Eco}\text{-}\mathrm{attention}}_{\mathrm{i},\mathrm{t}}={\mathrm{\alpha }}_0+{\mathrm{\beta }}_1{\mathrm{Faultlines}}_{\mathrm{i},\mathrm{t}}+\sum {\mathrm{\beta }}_{\mathrm{j}}{\mathrm{Control}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(5)

$${\mathrm{Green} \mathrm{Innovation}}_{\mathrm{i},\mathrm{t}+1}={ \mathrm{\alpha }}_0+{ \mathrm{\beta }}_1{\mathrm{Faultlines}}_{\mathrm{i},\mathrm{t}}+{ \mathrm{\beta }}_2{\mathrm{Eco}\text{-}\mathrm{attention}}_{\mathrm{i},\mathrm{t}}+\sum {\mathrm{\beta }}_{\mathrm{j}}{\mathrm{Control}}_{\mathrm{i},\mathrm{t}}+{ \mathrm{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(6)

$$\begin{array}{c}{\mathrm{Eco}\text{-}\mathrm{attention}}_{\mathrm{i},\mathrm{t}}=\hfill \\ {\mathrm{\alpha }}_0+{\mathrm{\beta }}_1{\mathrm{Faultlines}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\beta }}_2{\mathrm{Faultlines}}_{\mathrm{i},\mathrm{t}}\times {\mathrm{AI} \mathrm{Technology}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\beta }}_3{\mathrm{AI} \mathrm{Technology}}_{\mathrm{i},\mathrm{t}}\hfill \\ \hfill +\sum {\mathrm{\beta }}_{\mathrm{j}}{\mathrm{Control}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\epsilon }}_{\mathrm{i},\mathrm{t}}\hfill \end{array}$$

(7)

where ${\mathrm{Green} \mathrm{Innovation}}_{\mathrm{i},\mathrm{t}+1}$ denotes green innovation of firm_i in year_t+1. ${\mathrm{Faultline}}_{\mathrm{i},\mathrm{t}}$ is CEO–TMT faultlines of firm_i in year_t. ${\mathrm{Eco}\text{-}\mathrm{attention}}_{\mathrm{i},\mathrm{t}}$ is eco-attention of firm_i in the year_t. ${\mathrm{AI} \mathrm{Technology}}_{\mathrm{i},\mathrm{t}}$ is AI technology ability of firm_i in year_t. ${\mathrm{Controls}}_{\mathrm{i},\mathrm{t}}$ refers to control variables of firm_i in year_t, including industry and year effects. Where ε is the random error term. All models are estimated using Ordinary Least Squares (OLS) regression with robust standard errors to account for potential heteroscedasticity and within-firm correlation.

This approach effectively captures the underlying relationship between CEO–TMT faultlines and green innovation. To test H1, Model (4) examines the statistical significance of the negative impact of CEO–TMT faultlines on green innovation. To test H2, Model (5) investigates the significant influence of CEO–TMT faultlines on corporate eco-attention. Model (6) assesses the mediating role of eco-attention. To test H3, Model (7) examines the moderating effect of AI technology. Additionally, to account for the long-term nature of green innovation and to mitigate potential reverse causality concerns, we use green innovation at t + 1 as the dependent variable.

4. Results

4.1. Descriptive Statistics

Table 2. Summary Statistics of Variables.

Variable	N	Mean	SD	Min	Median	Max
Green Innovation	35,347	0.440	0.844	0.000	0.000	3.829
Faultlines	35,347	0.409	0.047	0.290	0.412	0.510
Eco-attention	35,347	0.260	0.313	0.000	0.143	1.739
AI Technology	35,347	0.005	0.008	0.000	0.002	0.052
TMT Size	35,347	7.258	2.684	3.000	7.000	16.000
Duality	35,347	0.315	0.465	0.000	0.000	1.000
Independent Ratio	35,347	37.662	5.275	33.330	36.360	57.140
Firm Age	35,347	2.932	0.339	1.946	2.996	3.611
Firm Size	35,347	6.057	1.246	3.896	5.865	9.930
Debt	35,347	0.407	0.203	0.050	0.397	0.873
ROA	35,347	0.044	0.062	−0.192	0.042	0.222
Ownership	35,347	0.309	0.462	0.000	0.000	1.000
Subsidy	35,347	0.005	0.003	0.002	0.004	0.016
Regulation	35,347	2.653	1.537	−1.812	2.662	6.576
Institutional Investor	35,347	42.868	25.012	0.350	44.102	90.751

presents the descriptive statistics for each variable. The mean value of green innovation is 0.440. The mean eco-attention value is 0.260. The mean value of the CEO–TMT faultlines is 0.409.

4.2. Hypothesis Testing

Table 3. Correlation matrix.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Green Innovation	1.000
Faultlines	−0.119	1.000
Eco-attention	0.218	−0.059	1.000
AI Technology	0.066	−0.020	−0.105	1.000
TMT Size	0.155	−0.550	0.058	0.044	1.000
Duality	−0.004	0.045	−0.064	0.040	−0.055	1.000
Independent Ratio	0.006	0.082	−0.047	0.013	−0.054	0.116	1.000
Firm Age	−0.034	0.083	0.066	−0.035	0.021	−0.103	−0.006	1.000
Firm Size	0.204	−0.180	0.128	−0.080	0.284	−0.190	−0.015	0.231
Debt	0.094	−0.090	0.140	−0.078	0.164	−0.147	−0.020	0.193
ROA	0.049	−0.045	−0.054	−0.047	−0.033	0.041	−0.009	−0.141
Ownership	0.051	−0.179	0.061	−0.030	0.160	−0.312	−0.079	0.170
Subsidy	0.254	−0.154	0.064	0.065	0.229	−0.059	0.004	0.088
Regulation	−0.019	−0.053	−0.016	0.031	0.081	−0.078	0.003	−0.023
Institutional Investor	0.059	−0.110	0.045	−0.055	0.130	−0.187	−0.080	0.027
	(9)	(10)	(11)	(12)	(13)	(14)	(15)
Firm Size	1.000
Debt	0.525	1.000
ROA	−0.035	−0.372	1.000
Ownership	0.360	0.291	−0.099	1.000
Subsidy	0.590	0.240	0.057	0.149	1.000
Regulation	0.082	0.046	−0.029	0.149	0.023	1.000
Institutional Investor	0.414	0.189	0.108	0.404	0.226	0.061	1.000

presents the correlations between the variables in this study. All correlation coefficients are below the 0.7 threshold. We also calculate the variance inflation factor (VIF). The maximum VIF is 2.50 and is well below 10, indicating that there is no multicollinearity problem between the variables.

Table 4. Regression results.

	(1)	(2)	(3)	(4)	(5)
	Green Innovation	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines		−0.371 ***	−0.064 **	−0.341 ***	−0.004
		(0.108)	(0.032)	(0.107)	(0.037)
Eco-attention				0.468 ***
				(0.021)
Faultlines * AI					−14.515 ***
					(3.055)
AI Technology	3.384 ***	3.382 ***	−1.090 ***	3.892 ***	4.792 ***
	(0.624)	(0.624)	(0.135)	(0.621)	(1.261)
TMT Size	0.025 ***	0.022 ***	0.002 ***	0.021 ***	0.002 ***
	(0.002)	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.025 ***	0.024 ***	−0.013 ***	0.030 ***	−0.013 ***
	(0.009)	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.000	0.001	−0.002 ***	0.001	−0.002 ***
	(0.001)	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.162 ***	−0.158 ***	−0.000	−0.158 ***	−0.001
	(0.015)	(0.015)	(0.004)	(0.015)	(0.004)
Firm Size	0.129 ***	0.129 ***	0.009 ***	0.124 ***	0.009 ***
	(0.006)	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.072 ***	0.072 ***	0.086 ***	0.032	0.086 ***
	(0.027)	(0.027)	(0.009)	(0.026)	(0.009)
ROA	0.769 ***	0.755 ***	−0.094 ***	0.799 ***	−0.094 ***
	(0.072)	(0.072)	(0.022)	(0.071)	(0.022)
Ownership	0.100 ***	0.096 ***	−0.012 ***	0.102 ***	−0.011 ***
	(0.011)	(0.011)	(0.004)	(0.011)	(0.004)
Subsidy	0.045 ***	0.044 ***	0.000	0.044 ***	0.000
	(0.004)	(0.004)	(0.001)	(0.003)	(0.001)
Regulation	−9.578 ***	−9.501 ***	−0.181	−9.416 ***	−0.178
	(1.529)	(1.529)	(0.485)	(1.520)	(0.484)
Institutional Investor	0.000	0.000	0.000 **	0.000	0.000 **
	(0.000)	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.253 ***	−0.089	0.250 ***	−0.206 ***	0.227 ***
	(0.061)	(0.078)	(0.023)	(0.078)	(0.024)
Year	Yes	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes	Yes
R2	0.248	0.249	0.464	0.265	0.465
F	192.853	181.266	39.278	202.334	40.577
N	35,347	35,347	35,347	35,347	35,347

Note: Robust standard errors in parentheses. ** p < 0.05; *** p < 0.01.

presents the results of the regression analysis. Column (1) shows the baseline model with control variables only. Column (2) reports the effects of CEO–TMT faultlines on green innovation. The coefficient estimate of Faultlines is −0.371 and is statistically significant at the 1% level. The results show that, after controlling for other factors, the CEO–TMT faultlines negatively affect green innovation. In other words, the stronger the CEO–TMT faultlines, the less green innovation firms have. Thus, H1 is supported.

Columns (3) and (4) examine the mediating role of eco-attention in the relationship between CEO–TMT faultlines and green innovation. Column (3) tests the effect of CEO–TMT faultlines on eco-attention. The coefficient of the Faultlines is negative and statistically significant (β = −0.064, p < 0.05). The result indicates that the CEO–TMT faultlines decrease the allocation of eco-attention. The result of Column (4) suggests that the coefficient of Faultlines on green innovation is −0.341, which is negative and significant (β = −0.341, p < 0.01). The coefficient of Eco-attention on green innovation is 0.468, which is significantly positive (β = 0.468, p < 0.01), indicating that eco-attention plays a mediating role between CEO–TMT faultlines and green innovation. We conducted robustness tests using bootstrapping methods to validate the mediating role of eco-attention, which confirmed our results. This finding suggests that the CEO–TMT faultlines inhibit green innovation by reducing eco-attention. Thus, H2 is supported.

Column (5) examines the moderating role of AI technology in the relationship between CEO–TMT faultlines and eco-attention. The coefficient of the interaction variable Faultlines * AI Technology is negative and significant (β = −14.515, p < 0.01), indicating that AI technology amplifies the negative relationship between CEO–TMT faultlines and eco-attention. H3 is supported.

4.3. Robustness Testing

To ensure the robustness of the results, we check these hypotheses with several additional tests. All robustness tests consistently support our hypotheses.

First, we used two alternative measures of green innovation. To assess whether the results are sensitive to this choice, we re-estimated the models using two alternative proxies, namely the natural logarithm of one plus the number of green invention patent applications, which impose a higher threshold for technological novelty, and the natural logarithm of one plus the number of granted green patents, which capture innovation outcomes that have cleared formal examination. The results, reported in

Table A3. Robustness result—alternative measures of green innovation (green invention patent applications).

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.200 **	−0.064 **	−0.179 **	−0.003
	(0.090)	(0.032)	(0.089)	(0.037)
Eco-attention			0.317 ***
			(0.018)
Faultlines * AI				−14.499 ***
				(3.055)
AI Technology	3.617 ***	−1.090 ***	3.962 ***	4.785 ***
	(0.561)	(0.135)	(0.560)	(1.261)
TMT Size	0.018 ***	0.002 ***	0.018 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.026 ***	−0.013 ***	0.030 ***	−0.013 ***
	(0.008)	(0.003)	(0.008)	(0.003)
Independent Ratio	−0.000	−0.002 ***	0.000	−0.002 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.107 ***	−0.000	−0.107 ***	−0.001
	(0.013)	(0.004)	(0.013)	(0.004)
Firm Size	0.108 ***	0.009 ***	0.105 ***	0.009 ***
	(0.005)	(0.002)	(0.005)	(0.002)
Debt	0.018	0.086 ***	−0.009	0.086 ***
	(0.022)	(0.009)	(0.022)	(0.009)
ROA	0.563 ***	−0.094 ***	0.592 ***	−0.094 ***
	(0.060)	(0.022)	(0.059)	(0.022)
Ownership	0.101 ***	−0.011 ***	0.105 ***	−0.011 ***
	(0.010)	(0.004)	(0.009)	(0.004)
Subsidy	0.042 ***	0.000	0.042 ***	0.000
	(0.003)	(0.001)	(0.003)	(0.001)
Regulation	−7.314 ***	−0.167	−7.261 ***	−0.165
	(1.274)	(0.485)	(1.270)	(0.485)
Institutional Investor	0.000	0.000 **	−0.000	0.000 **
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.241 ***	0.250 ***	−0.320 ***	0.227 ***
	(0.066)	(0.023)	(0.066)	(0.024)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.223	0.464	0.234	0.465
F	167.005	39.274	175.593	40.570
N	35,340	35,340	35,340	35,340

Note: Robust standard errors in parentheses. ** p < 0.05; *** p < 0.01.

and

Table A4. Robustness result—alternative measures of green innovation (granted green patents).

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.380 ***	−0.062 **	−0.352 ***	−0.006
	(0.093)	(0.031)	(0.092)	(0.036)
Eco-attention			0.442 ***
			(0.019)
Faultlines * AI				−13.475 ***
				(3.004)
AI Technology	1.638 ***	−1.102 ***	2.125 ***	4.353 ***
	(0.512)	(0.133)	(0.511)	(1.239)
TMT Size	0.017 ***	0.002 ***	0.015 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.014 *	−0.013 ***	0.019 **	−0.013 ***
	(0.008)	(0.003)	(0.008)	(0.003)
Independent Ratio	0.001 *	−0.002 ***	0.002 ***	−0.002 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.137 ***	−0.000	−0.137 ***	−0.000
	(0.013)	(0.004)	(0.013)	(0.004)
Firm Size	0.107 ***	0.009 ***	0.103 ***	0.009 ***
	(0.005)	(0.002)	(0.005)	(0.002)
Debt	0.047 **	0.081 ***	0.011	0.081 ***
	(0.022)	(0.008)	(0.022)	(0.008)
ROA	0.354 ***	−0.072 ***	0.386 ***	−0.072 ***
	(0.057)	(0.020)	(0.056)	(0.020)
Ownership	0.052 ***	−0.011 ***	0.057 ***	−0.011 ***
	(0.010)	(0.004)	(0.009)	(0.004)
Subsidy	0.036 ***	0.000	0.036 ***	0.000
	(0.003)	(0.001)	(0.003)	(0.001)
Regulation	−7.384 ***	−0.180	−7.304 ***	−0.176
	(1.271)	(0.480)	(1.264)	(0.480)
Institutional Investor	−0.000	0.000 **	−0.000	0.000 **
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.015	0.247 ***	−0.124 *	0.226 ***
	(0.067)	(0.022)	(0.067)	(0.024)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.239	0.463	0.259	0.463
F	151.429	38.793	177.373	39.838
N	35,945	35,945	35,945	35,945

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

, are consistent with the baseline findings, suggesting that our conclusions are not sensitive to the measurement of green innovation. H1, H2, and H3 are supported.

Second, we employed two alternative measures of eco-attention. Specifically, we first measured eco-attention using the total number of environmental keywords appearing in the MD&A sections of firms’ annual reports. The results reported in

Table A5. Robustness result—alternative measure of eco-attention (total number of environmental keywords).

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.374 ***	−62.626 ***	−0.186 *	−42.166 ***
	(0.108)	(12.231)	(0.105)	(13.781)
Eco-attention			0.001 ***
			(0.000)
Faultlines * AI				−4920.841 ***
				(1450.766)
AI Technology	3.395 ***	−92.879	3.366 ***	1900.952 ***
	(0.624)	(61.481)	(0.593)	(591.971)
TMT Size	0.022 ***	1.428 ***	0.020 ***	1.419 ***
	(0.002)	(0.218)	(0.002)	(0.218)
Duality	0.024 ***	−1.058	0.033 ***	−1.023
	(0.009)	(1.084)	(0.009)	(1.084)
Independent Ratio	0.001	−0.527 ***	0.000	−0.529 ***
	(0.001)	(0.090)	(0.001)	(0.090)
Firm Age	−0.157 ***	−9.041 ***	−0.142 ***	−9.117 ***
	(0.015)	(1.653)	(0.015)	(1.653)
Firm Size	0.129 ***	10.033 ***	0.088 ***	10.030 ***
	(0.006)	(0.646)	(0.006)	(0.646)
Debt	0.072 ***	8.226 **	0.080 ***	8.361 ***
	(0.027)	(3.218)	(0.025)	(3.218)
ROA	0.754 ***	−81.509 ***	0.788 ***	−81.462 ***
	(0.072)	(8.954)	(0.070)	(8.952)
Ownership	0.096 ***	−10.567 ***	0.095 ***	−10.505 ***
	(0.011)	(1.285)	(0.011)	(1.285)
Subsidy	0.045 ***	2.840 ***	0.033 ***	2.853 ***
	(0.004)	(0.432)	(0.003)	(0.432)
Regulation	−9.477 ***	102.291	−7.819 ***	103.128
	(1.530)	(184.545)	(1.464)	(184.500)
Institutional Investor	0.000	−0.072 ***	−0.000	−0.073 ***
	(0.000)	(0.023)	(0.000)	(0.023)
Cons	−0.091	294.574 ***	−0.254 ***	286.539 ***
	(0.078)	(8.613)	(0.079)	(8.967)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.249	0.443	0.213	0.443
F	181.657	104.102	128.797	97.731
N	35,357	35,357	35,357	35,357

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

remain consistent with the baseline findings. In addition, we constructed an expanded keyword dictionary using a Word2Vec approach (see Table A2). Following prior studies, seed words related to environmental issues were first identified based on CSR reports, policy documents, and the relevant literature and were then expanded using the Word2Vec model to capture semantically related terms. Eco-attention was subsequently measured as the ratio of environmental keywords to the total number of words in the MD&A sections. The results reported in

Table A6. Robustness result—alternative measure of eco-attention (Word2Vec).

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.374 ***	−0.061 **	−0.214 **	0.012
	(0.108)	(0.031)	(0.105)	(0.035)
Eco-attention			0.462 ***
			(0.023)
Faultlines * AI				−17.558 ***
				(3.338)
AI Technology	3.395 ***	−0.921 ***	3.708 ***	6.193 ***
	(0.624)	(0.146)	(0.595)	(1.381)
TMT Size	0.022 ***	0.001	0.021 ***	0.001
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.024 ***	−0.004	0.034 ***	−0.004
	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.001	−0.001 ***	0.000	−0.001 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.157 ***	−0.014 ***	−0.144 ***	−0.014 ***
	(0.015)	(0.004)	(0.014)	(0.004)
Firm Size	0.129 ***	0.013 ***	0.091 ***	0.013 ***
	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.072 ***	0.124 ***	0.030	0.124 ***
	(0.027)	(0.008)	(0.025)	(0.008)
ROA	0.754 ***	−0.108 ***	0.764 ***	−0.108 ***
	(0.072)	(0.022)	(0.070)	(0.022)
Ownership	0.096 ***	−0.023 ***	0.096 ***	−0.022 ***
	(0.011)	(0.003)	(0.011)	(0.003)
Subsidy	0.045 ***	0.002 *	0.035 ***	0.002 *
	(0.004)	(0.001)	(0.003)	(0.001)
Regulation	−9.477 ***	−0.557	−7.469 ***	−0.554
	(1.530)	(0.458)	(1.460)	(0.458)
Institutional Investor	0.000	0.000	−0.000	−0.000
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.091	0.224 ***	−0.091	0.196 ***
	(0.078)	(0.022)	(0.078)	(0.023)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.249	0.431	0.220	0.431
F	181.657	63.007	131.063	61.267
N	35,357	35,357	35,357	35,357

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

are also consistent with the baseline findings.

Third, we replaced the measurement of AI technology. As an alternative specification, we aggregate firms’ AI-related intangible assets and fixed assets and use the natural logarithm of their sum plus one as a proxy for AI technology. In columns (1)–(3) of

Table A7. Robustness result—alternative measure of AI technology.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.371 ***	−0.063 **	−0.341 ***	−0.098
	(0.108)	(0.032)	(0.107)	(0.150)
Eco-attention			0.471 ***
			(0.021)
Faultlines * AI				0.002
				(0.009)
AI Technology	0.016 ***	−0.009 ***	0.021 ***	−0.010 **
	(0.003)	(0.001)	(0.002)	(0.004)
TMT Size	0.022 ***	0.002 ***	0.021 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.024 ***	−0.013 ***	0.030 ***	−0.013 ***
	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.001	−0.002 ***	0.001	−0.002 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.155 ***	−0.001	−0.155 ***	−0.001
	(0.015)	(0.004)	(0.015)	(0.004)
Firm Size	0.114 ***	0.016 ***	0.106 ***	0.016 ***
	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.068 **	0.087 ***	0.027	0.087 ***
	(0.027)	(0.009)	(0.026)	(0.009)
ROA	0.741 ***	−0.094 ***	0.786 ***	−0.094 ***
	(0.072)	(0.021)	(0.071)	(0.021)
Ownership	0.097 ***	−0.012 ***	0.102 ***	−0.012 ***
	(0.011)	(0.004)	(0.011)	(0.004)
Subsidy	0.044 ***	0.001	0.043 ***	0.001
	(0.004)	(0.001)	(0.003)	(0.001)
Regulation	−9.311 ***	−0.241	−9.198 ***	−0.239
	(1.530)	(0.484)	(1.522)	(0.484)
Institutional Investor	0.000	0.000 ***	−0.000	0.000 ***
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.241 ***	0.336 ***	−0.399 ***	0.350 ***
	(0.082)	(0.024)	(0.082)	(0.064)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.249	0.465	0.265	0.465
F	182.540	42.283	204.483	39.264
N	35,353	35,353	35,353	35,353

Note: Robust standard errors in parentheses. ** p < 0.05; *** p < 0.01.

, the results show that both the main effect and the mediating effect remain consistent with the baseline findings. However, in column (4) of Table A7, the coefficient of Faultlines × AI becomes statistically insignificant under this specification. One possible explanation is that this measure captures the absolute scale of AI investment rather than the relative intensity of AI assets embedded within the firm. Consequently, it tends to covary with overall firm size, which may obscure the moderating mechanism proposed in our theoretical framework. Importantly, this pattern is broadly consistent with our theoretical argument. If the moderating role of AI technology operates through its structural integration into firms’ information processing and decision-making processes, then a measure that primarily reflects investment scale may be less suited to capturing this mechanism. From this perspective, the sensitivity of the moderating effect to measurement specification highlights the conceptual distinction between AI asset intensity and the overall scale of AI investment.

Fourth, considering that firms’ provincial locations may influence both the formation of CEO–TMT faultlines and their effects on green innovation, we additionally control for provincial fixed effects. Firms’ geographic locations may shape both the formation of CEO–TMT faultlines and their innovation behaviors due to differences in regional institutions, environmental regulations, and economic development. After including provincial fixed effects, the results remain robust in

Table A8. Robustness result—including provincial fixed effects.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.294 ***	−0.059 *	−0.266 **	0.001
	(0.108)	(0.032)	(0.107)	(0.037)
Eco-attention			0.477 ***
			(0.021)
Faultlines * AI				−14.335 ***
				(3.077)
AI Technology	3.205 ***	−1.061 ***	3.711 ***	4.746 ***
	(0.621)	(0.135)	(0.619)	(1.270)
TMT Size	0.022 ***	0.002 ***	0.021 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.022 **	−0.011 ***	0.027 ***	−0.011 ***
	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.001	−0.001 ***	0.001	−0.001 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.155 ***	0.004	−0.157 ***	0.003
	(0.015)	(0.004)	(0.015)	(0.004)
Firm Size	0.127 ***	0.006 ***	0.123 ***	0.006 ***
	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.072 ***	0.087 ***	0.031	0.087 ***
	(0.027)	(0.009)	(0.026)	(0.009)
ROA	0.719 ***	−0.108 ***	0.771 ***	−0.108 ***
	(0.072)	(0.022)	(0.071)	(0.022)
Ownership	0.099 ***	−0.010 ***	0.104 ***	−0.010 ***
	(0.011)	(0.004)	(0.011)	(0.004)
Subsidy	0.045 ***	0.002 *	0.044 ***	0.002 *
	(0.004)	(0.001)	(0.003)	(0.001)
Regulation	−1.584	−0.245	−1.467	−0.232
	(3.165)	(0.955)	(3.122)	(0.954)
Institutional Investor	0.000	0.000 **	0.000	0.000 **
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.166 **	0.241 ***	−0.281 ***	0.217 ***
	(0.080)	(0.023)	(0.079)	(0.024)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
Province	Yes	Yes	Yes	Yes
R2	0.255	0.471	0.272	0.471
F	174.918	37.796	197.688	38.940
N	35,351	35,351	35,351	35,351

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

.

Fifth, we employed the Tobit regression model as an alternative specification to test the robustness of our results. Since green patent counts are left-censored at zero, OLS estimates may be subject to censoring bias. We therefore re-estimated the models using Tobit regression, with the untransformed count of green patent applications as the dependent variable. As shown in

Table A9. Robustness result—tobit regression model.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−25.723 ***	−0.061 *	−23.903 ***	−0.005
	(−3.80)	(−1.94)	(−3.54)	(−0.12)
Eco-attention			18.250 ***
			(18.65)
Faultlines * AI				−15.499 ***
				(−3.84)
AI Technology	128.256 ***	−1.089 ***	152.669 ***	5.151 ***
	(3.87)	(−8.08)	(4.61)	(3.13)
TMT Size	0.919 ***	0.002 ***	0.878 ***	0.002 ***
	(7.93)	(3.74)	(7.59)	(3.38)
Duality	2.883 ***	−0.013 ***	3.267 ***	−0.013 ***
	(4.84)	(−4.53)	(5.49)	(−4.36)
Independent Ratio	−0.088 *	−0.002 ***	−0.044	−0.002 ***
	(−1.77)	(−6.43)	(−0.88)	(−6.88)
Firm Age	−9.619 ***	−0.000	−9.673 ***	0.001
	(−10.44)	(−0.02)	(−10.51)	(0.17)
Firm Size	8.284 ***	0.009 ***	8.161 ***	0.009 ***
	(22.01)	(5.10)	(21.70)	(4.95)
Debt	0.446	0.086 ***	−1.833	0.085 ***
	(0.24)	(9.99)	(−0.98)	(9.97)
ROA	47.815 ***	−0.095 ***	50.896 ***	−0.087 ***
	(9.24)	(−4.40)	(9.81)	(−3.72)
Ownership	5.006 ***	−0.012 ***	5.431 ***	−0.011 ***
	(6.80)	(−3.17)	(7.39)	(−3.25)
Subsidy	2.645 ***	0.000	2.692 ***	0.000
	(10.05)	(0.22)	(10.23)	(0.14)
Regulation	−597.859 ***	−0.194	−605.651 ***	−0.192
	(−5.59)	(−0.40)	(−5.67)	(−0.39)
Institutional Investor	−0.029 **	0.000 **	−0.033 ***	0.000 ***
	(−2.30)	(2.34)	(−2.61)	(2.58)
Cons	−232.995	0.249	−241.138	0.231
	(−1.00)	(11.0)	(−0.00)	(9.66)
R2	-	0.464	-	0.448
F	-	39.318	-	41.505
N	35,521	35,357	35,521	35,504

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

, the results are consistent with those reported in the baseline.

Finally, we tested the robustness of the mediating effect using bootstrapping and performed 500 resamples. The bootstrapped confidence intervals confirm that the mediating effect remains statistically significant, further supporting the proposed mechanism. The results in

Table A10. Robustness result—bootstrap mediation test.

	Coefficient	SE	Z	P	LLCI	ULCI
Indirect effect	−0.0299	0.0145	−2.06	0.039	−0.0582	−0.0015
Direct effect	−0.3409	0.1023	−3.33	0.001	−0.5416	−0.1403

indicate that the mediating effect is robust.

4.4. Endogeneity Testing

To address potential endogeneity concerns, we conducted several additional analyses, including propensity score matching (PSM), Heckman’s sample selection model, and seemingly unrelated regression (SUR). The results consistently support our main findings, confirming their robustness after accounting for endogeneity issues.

First, we used PSM to mitigate potential self-selection bias. CEO–TMT faultlines may not be randomly distributed across firms but rather systematically associated with firm characteristics, raising concerns about self-selection bias. We divided firms into high and low CEO–TMT faultlines groups. Firms with above-average CEO–TMT faultlines were classified into the high faultlines group, while those below average were assigned to the low faultlines group. We then performed several matching procedures, including 1:1 matching, 1:2 matching, nearest neighbor matching, and radius matching, using TMT Size, AI Technology, and ROA as matching variables. The results reported in

Table 5. PSM result—1:1 matching.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.682 ***	−0.117 ***	−0.635 ***	−0.079 *
	(0.131)	(0.039)	(0.129)	(0.046)
Eco-attention			0.403 ***
			(0.025)
Faultlines * AI				−9.338 **
				(3.909)
AI Technology	3.226 ***	−1.501 ***	3.830 ***	2.383
	(0.732)	(0.154)	(0.730)	(1.635)
TMT Size	0.018 ***	0.002 ***	0.017 ***	0.002 ***
	(0.003)	(0.001)	(0.003)	(0.001)
Duality	0.020 *	−0.014 ***	0.026 **	−0.014 ***
	(0.011)	(0.003)	(0.011)	(0.003)
Independent Ratio	0.001	−0.001 ***	0.001	−0.001 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.175 ***	0.007	−0.178 ***	0.007
	(0.018)	(0.005)	(0.018)	(0.005)
Firm Size	0.113 ***	0.007 ***	0.110 ***	0.007 ***
	(0.007)	(0.002)	(0.007)	(0.002)
Debt	0.064 **	0.081 ***	0.032	0.081 ***
	(0.030)	(0.010)	(0.029)	(0.010)
ROA	0.656 ***	−0.114 ***	0.702 ***	−0.113 ***
	(0.083)	(0.026)	(0.082)	(0.026)
Ownership	0.073 ***	−0.014 ***	0.079 ***	−0.013 ***
	(0.013)	(0.004)	(0.013)	(0.004)
Subsidy	0.045 ***	0.002	0.045 ***	0.002
	(0.004)	(0.001)	(0.004)	(0.001)
Regulation	−9.981 ***	0.155	−10.044 ***	0.168
	(1.721)	(0.586)	(1.716)	(0.586)
Institutional Investor	−0.000	0.000	−0.000	0.000
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	0.227 **	0.243 ***	0.129	0.228 ***
	(0.091)	(0.027)	(0.090)	(0.029)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.236	0.464	0.249	0.464
F	103.267	31.155	116.005	30.176
N	24,973	24,973	24,973	24,973

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

,

Table 6. PSM result—1:2 matching.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.561 ***	−0.091 **	−0.523 ***	−0.036
	(0.119)	(0.035)	(0.118)	(0.041)
Eco-attention			0.419 ***
			(0.024)
Faultlines * AI				−13.356 ***
				(3.483)
AI Technology	3.364 ***	−1.299 ***	3.909 ***	4.192 ***
	(0.685)	(0.144)	(0.682)	(1.444)
TMT Size	0.019 ***	0.002 ***	0.018 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.018 *	−0.015 ***	0.024 **	−0.015 ***
	(0.010)	(0.003)	(0.010)	(0.003)
Independent Ratio	0.001	−0.001 ***	0.001	−0.001 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.165 ***	0.001	−0.165 ***	0.001
	(0.017)	(0.005)	(0.017)	(0.005)
Firm Size	0.120 ***	0.009 ***	0.117 ***	0.009 ***
	(0.007)	(0.002)	(0.006)	(0.002)
Debt	0.056 **	0.079 ***	0.023	0.079 ***
	(0.029)	(0.009)	(0.028)	(0.009)
ROA	0.702 ***	−0.104 ***	0.746 ***	−0.103 ***
	(0.078)	(0.024)	(0.077)	(0.024)
Ownership	0.080 ***	−0.011 ***	0.085 ***	−0.011 ***
	(0.012)	(0.004)	(0.012)	(0.004)
Subsidy	0.045 ***	0.001	0.044 ***	0.001
	(0.004)	(0.001)	(0.004)	(0.001)
Regulation	−9.482 ***	−0.116	−9.433 ***	−0.104
	(1.625)	(0.544)	(1.621)	(0.544)
Institutional Investor	0.000	0.000	0.000	0.000
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	0.089	0.249 ***	−0.016	0.228 ***
	(0.085)	(0.025)	(0.085)	(0.026)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.238	0.463	0.251	0.463
F	125.756	33.240	139.805	33.266
N	28,523	28,523	28,523	28,523

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

,

Table 7. PSM result—nearest neighbor matching.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.491 ***	−0.078 **	−0.456 ***	−0.027
	(0.112)	(0.033)	(0.111)	(0.039)
Eco-attention			0.439 ***
			(0.022)
Faultlines * AI				−12.269 ***
				(3.131)
AI Technology	3.428 ***	−1.252 ***	3.977 ***	3.743 ***
	(0.651)	(0.138)	(0.649)	(1.291)
TMT Size	0.021 ***	0.002 ***	0.020 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.020 **	−0.014 ***	0.026 ***	−0.013 ***
	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.000	−0.001 ***	0.001	−0.001 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.159 ***	−0.001	−0.158 ***	−0.001
	(0.016)	(0.005)	(0.016)	(0.005)
Firm Size	0.123 ***	0.009 ***	0.119 ***	0.009 ***
	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.066 **	0.085 ***	0.029	0.085 ***
	(0.027)	(0.009)	(0.027)	(0.009)
ROA	0.740 ***	−0.102 ***	0.784 ***	−0.102 ***
	(0.076)	(0.023)	(0.075)	(0.023)
Ownership	0.090 ***	−0.012 ***	0.095 ***	−0.011 ***
	(0.012)	(0.004)	(0.012)	(0.004)
Subsidy	0.044 ***	0.001	0.044 ***	0.001
	(0.004)	(0.001)	(0.004)	(0.001)
Regulation	−9.106 ***	−0.166	−9.034 ***	−0.160
	(1.573)	(0.512)	(1.566)	(0.512)
Institutional Investor	0.000	0.000 *	0.000	0.000 *
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	0.014	0.253 ***	−0.097	0.233 ***
	(0.081)	(0.024)	(0.081)	(0.025)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.245	0.462	0.259	0.462
F	149.512	37.616	166.209	37.595
N	31,894	31,894	31,894	31,894

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

and

Table 8. PSM result—radius matching.

	(1)	(2)	(3)	(4)
	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Faultlines	−0.412 ***	−0.060 *	−0.384 ***	0.001
	(0.109)	(0.032)	(0.107)	(0.037)
Eco-attention			0.468 ***
			(0.022)
Faultlines * AI				−14.796 ***
				(3.146)
AI Technology	3.405 ***	−1.093 ***	3.916 ***	4.890 ***
	(0.632)	(0.138)	(0.629)	(1.295)
TMT Size	0.022 ***	0.002 ***	0.021 ***	0.002 ***
	(0.002)	(0.001)	(0.002)	(0.001)
Duality	0.025 ***	−0.013 ***	0.031 ***	−0.013 ***
	(0.009)	(0.003)	(0.009)	(0.003)
Independent Ratio	0.000	−0.002 ***	0.001	−0.002 ***
	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.157 ***	−0.000	−0.157 ***	−0.000
	(0.015)	(0.004)	(0.015)	(0.004)
Firm Size	0.128 ***	0.009 ***	0.124 ***	0.009 ***
	(0.006)	(0.002)	(0.006)	(0.002)
Debt	0.070 ***	0.086 ***	0.030	0.087 ***
	(0.027)	(0.009)	(0.026)	(0.009)
ROA	0.785 ***	−0.104 ***	0.834 ***	−0.104 ***
	(0.074)	(0.022)	(0.073)	(0.022)
Ownership	0.096 ***	−0.012 ***	0.102 ***	−0.012 ***
	(0.011)	(0.004)	(0.011)	(0.004)
Subsidy	0.045 ***	0.000	0.045 ***	0.000
	(0.004)	(0.001)	(0.004)	(0.001)
Regulation	−9.614 ***	−0.205	−9.518 ***	−0.204
	(1.546)	(0.491)	(1.536)	(0.491)
Institutional Investor	0.000	0.000 **	0.000	0.000 **
	(0.000)	(0.000)	(0.000)	(0.000)
Cons	−0.077	0.250 ***	−0.194 **	0.226 ***
	(0.079)	(0.023)	(0.078)	(0.024)
Year	Yes	Yes	Yes	Yes
Industry	Yes	Yes	Yes	Yes
R2	0.2492	0.464	0.265	0.464
F	176.798	38.803	198.104	39.910
N	34,892	34,892	34,892	34,892

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

remain consistent with the baseline findings, indicating that our conclusions are robust after controlling for potential self-selection bias.

Second, we employed Heckman’s two-stage sample selection model to address potential selection bias. Some firms may not engage in green innovation activities, which may lead to a non-random sample of firms with observable green innovation. To account for this potential selection issue, we applied Heckman’s two-stage sample selection, using the provincial average level of green innovation in the prior year as an instrumental variable. This instrumental variable affects the likelihood of engaging in green innovation but is unlikely to directly influence the focal relationships examined in this study. In the first stage, we estimated the probability that a firm engages in green innovation and calculated the inverse Mills ratio (IMR). In the second stage, we assessed the robustness of our results by including the IMR in the model. The results reported in

Table 9. Heckman’s two-stage sample selection.

	(1)	(2)	(3)	(4)	(5)
	Green Innovation	Green Innovation	Eco-Attention	Green Innovation	Eco-Attention
Main	0.373 ***
Greenmean	(0.136)
Faultlines		−0.462 ***	−0.064 *	−0.432 ***	−0.005
		(0.128)	(0.037)	(0.126)	(0.043)
Eco-attention				0.464 ***
				(0.024)
Faultlines * AI					−13.400 ***
					(3.407)
AI Technology	7.486 ***	4.886 ***	−1.267 ***	5.473 ***	4.193 ***
	(2.176)	(0.746)	(0.149)	(0.743)	(1.412)
TMT Size	−0.003	0.022 ***	0.003 ***	0.020 ***	0.003 ***
	(0.006)	(0.003)	(0.001)	(0.002)	(0.001)
Duality	−0.019	0.025 **	−0.014 ***	0.032 ***	−0.014 ***
	(0.037)	(0.011)	(0.003)	(0.011)	(0.003)
Independent Ratio	0.003	−0.000	−0.002 ***	0.001	−0.002 ***
	(0.003)	(0.001)	(0.000)	(0.001)	(0.000)
Firm Age	−0.060	−0.177 ***	−0.001	−0.176 ***	−0.001
	(0.059)	(0.018)	(0.005)	(0.018)	(0.005)
Firm Size	0.179 ***	0.131 ***	0.006 ***	0.129 ***	0.006 ***
	(0.020)	(0.007)	(0.002)	(0.007)	(0.002)
Debt	−1.570 ***	0.053	0.124 ***	−0.004	0.125 ***
	(0.085)	(0.035)	(0.011)	(0.035)	(0.011)
ROA	5.361 ***	1.057 ***	−0.160 ***	1.131 ***	−0.159 ***
	(0.198)	(0.103)	(0.028)	(0.102)	(0.028)
Ownership	0.202 ***	0.108 ***	−0.014 ***	0.114 ***	−0.014 ***
	(0.042)	(0.013)	(0.004)	(0.013)	(0.004)
Subsidy	0.105 ***	0.054 ***	−0.002	0.055 ***	−0.002
	(0.013)	(0.004)	(0.001)	(0.004)	(0.001)
Regulation	−20.215 ***	−10.425 ***	−0.348	−10.264 ***	−0.331
	(5.923)	(1.795)	(0.563)	(1.790)	(0.563)
Institutional Investor	−0.004 ***	−0.000	0.000 ***	−0.000	0.000 ***
	(0.001)	(0.000)	(0.000)	(0.000)	(0.000)
IMR		0.251 ***	−0.102 ***	0.298 ***	−0.102 ***
		(0.064)	(0.022)	(0.063)	(0.022)
Cons	0.647	−0.002	0.266 ***	−0.126	0.243 ***
	(0.622)	(0.093)	(0.027)	(0.092)	(0.028)
Industry	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes
R2	-	0.239	0.429	0.255	0.429
F	-	146.086	34.456	159.949	35.298
N	28,489	26,439	26,439	26,439	26,439

Note: Robust standard errors in parentheses. * p < 0.1; ** p < 0.05; *** p < 0.01.

show that the main findings remain consistent, suggesting that sample selection bias does not materially affect our conclusions.

Third, we used SUR to address the potential correlation of error terms across equations. Because eco-attention and green innovation are jointly determined outcomes in our theoretical framework, the error terms of their regression equations may be correlated. We used SUR methods to simultaneously estimate regression equations to correct for such potential bias. The results in

Table 10. SUR results.

	(1)	(2)
	Green Innovation	Eco-Attention
Faultlines	−0.482 ***	−0.268 ***
	(0.112)	(0.043)
AI Technology	6.291 ***	−3.618 ***
	(0.536)	(0.205)
TMT Size	0.024 ***	0.001
	(0.002)	(0.001)
Duality	0.026 ***	−0.021 ***
	(0.010)	(0.004)
Independent Ratio	0.002 *	−0.002 ***
	(0.001)	(0.000)
Firm Age	−0.167 ***	0.029 ***
	(0.013)	(0.005)
Firm Size	0.067 ***	0.014 ***
	(0.005)	(0.002)
Debt	0.130 ***	0.125 ***
	(0.027)	(0.011)
ROA	0.707 ***	−0.120 ***
	(0.078)	(0.030)
Ownership	0.014	−0.006
	(0.011)	(0.004)
Subsidy	0.096 ***	0.002
	(0.004)	(0.001)
Regulation	−11.775 ***	−2.759 ***
	(1.534)	(0.586)
Institutional Investor	−0.001 ***	−0.000
	(0.000)	(0.000)
Cons	0.226 ***	0.268 ***
	(0.071)	(0.027)
Industry	Yes	Yes
Year	Yes	Yes
R2	0.090	0.090
N	35,511	35,511

Note: Robust standard errors in parentheses. * p < 0.1; *** p < 0.01.

show that the CEO–TMT faultlines have significant negative effects on green innovation and eco-attention. These results are consistent with those of the baseline regressions.

5. Conclusions and Discussion

5.1. Conclusions

As environmental challenges continue to grow, the role of AI in fostering green innovation has become increasingly important. Based on ABV, this study investigates how CEO–TMT faultlines influence green innovation by incorporating eco-attention and AI technology concepts. We empirically analyzed data from 35,347 firm-year observations of Chinese A-share listed firms from 2010 to 2023. The findings are as follows:

First, the results confirm that CEO–TMT faultlines have a negative effect on green innovation. Existing research identifies the structure and interaction of TMTs as key factors influencing corporate green innovation [6,7]. This study demonstrates that CEO–TMT faultlines, stemming from characteristics, roles, and power differences, amplify the detrimental effects of member diversity on communication and knowledge integration while failing to harness the benefits of informational variety, ultimately hindering corporate green innovation. Our study examines the relationship between CEO–TMT faultlines and green innovation, deepening the understanding of prior research.

Second, by introducing eco-attention as a mediating variable, the results demonstrate that CEO–TMT faultlines hinder the allocation of eco-attention, ultimately impeding green innovation. Based on ABV, we find that when CEO–TMT faultlines are strong, TMTs become less willing to focus on sustainability strategies. Although scholars have examined the influence of intra-TMT dynamics on green innovation [7], they have largely overlooked the underlying mechanisms. Bromiley and Rau [17] called for further investigation into how CEO–TMT interactions shape team cognition and the intermediate effects on organizational strategic decision-making; however, research in this area remains limited. Neely, Lovelace, Cowen and Hiller [19] further explored how top management teams influence the cognitive processes underlying executive decision making and actions. Our findings extend the existing literature on green innovation.

Third, our findings reveal that AI technology intensifies the negative effects of CEO–TMT faultlines on eco-attention, suggesting that the organizational benefits of AI are not unconditional but are contingent upon the cohesion of the top management team. When CEO–TMT faultlines are pronounced, AI technology appears to reinforce rather than bridge existing cognitive divides, further impeding the allocation of organizational attention toward environmental issues. This finding advances research on the organizational implications of AI by highlighting that its effectiveness is shaped by the intra-team dynamics within which it is embedded. We respond to calls for greater attention to how managerial and organizational contexts condition the outcomes of AI adoption [33] and suggest that realizing the potential of AI for sustainable strategy requires firms to simultaneously address the structural fault lines that fragment executive attention.

5.2. Theoretical Implications

First, this study advances the understanding of how TMT interactions shape green innovation by revealing the cognitive mechanisms through which CEO–TMT faultlines influence green innovation via eco-attention as a mediating variable. Although prior research highlights the importance of both the CEO and TMT for green innovation [6,37,99], the cognitive mechanisms explaining why TMTs differ in prioritizing environmental issues remain underexplored [18]. While some scholars have suggested that distinguishing the CEO from other TMT members is a key factor in addressing this question [7], the underlying mechanisms are still not fully understood. Drawing on the attention-based view (ABV), this study shows that CEO–TMT faultlines weaken organizational eco-attention by disrupting the processes of communication and integration through which environmental issues are recognized and elevated to strategic priorities. Importantly, our findings suggest that this reduction in eco-attention is not merely about the amount of attention directed toward environmental issues. Rather, it reflects a weakening of the shared understanding among TMT members that environmental issues deserve strategic priority. This study responds to calls by Neely, Lovelace, Cowen and Hiller [19] to further investigate the mechanism of TMT interactions in shaping organizational outcomes.

Second, this study examines how AI technology interacts with team structure to shape attention allocation, thereby deepening the understanding of the effect of AI technology on managerial cognition. Prior studies have frequently emphasized the positive role of AI technology in enhancing information processing and decision-making efficiency [65,100]. Recent research has similarly shown that AI can stimulate team interaction and strengthen relational coordination [101]. In contrast, our findings indicate that these beneficial effects depend strongly on the structural and relational characteristics of the team, thereby complementing existing research. Specifically, our results highlight an important boundary condition of this argument. When power asymmetry coexists with low trust between subgroups, AI-generated analyses are unlikely to serve as a shared basis for team-level discussion. Instead, they are more likely to be interpreted by different subgroups in ways that support their pre-existing strategic positions. Under such conditions, AI technology does not necessarily facilitate coordination; rather, it may reinforce existing divisions in attention and interpretation. These findings suggest that the positive effects of AI technology are not inherent properties of the technology itself, but depend on the organizational context in which AI technology is embedded [102]. By identifying CEO–TMT faultlines as a key condition, this study advances our understanding of how AI technology shapes managerial cognition and strategic decision-making processes.

Third, this study extends the boundaries of ABV in the digital era by integrating CEO–TMT faultlines, eco-attention, green innovation, and AI technology into a unified theoretical framework. Our findings suggest that introducing AI technology into strategic decision-making does not automatically improve organizational outcomes. Its impact depends on whether the team has the relational foundations needed to discuss and translate AI-generated information into collective strategic decisions. This insight responds to recent calls to examine how AI shapes strategic decision-making in contexts characterized by intra-team conflict and limited trust [26], contributing to the development of the ABV.

5.3. Practical Implications

This study’s findings have several important implications. First, it provides a theoretical foundation on how firms can enhance green innovation by choosing the CEO and other TMT members. Firms should pay attention to the cognitive differences between the CEO and other TMT members to minimize potential faultlines and communication barriers. In practice, firms can take measures to strengthen TMT cohesion, improve the quality of interactions between the CEO and other TMT members, and enhance their collective focus on environmental issues to drive green innovation. For example, fostering cross-departmental or cross-industry collaboration can help to reduce potential conflicts within the TMT.

Second, this study provides practical guidance for firms to optimize strategic decision making through AI technology. The findings suggest that AI technology reinforces the biases caused by CEO–TMT faultlines and exacerbates their negative impact on eco-attention. It further amplifies emotional conflicts and communication barriers within the TMT, hindering their ability to engage in effective communication. Therefore, when adopting AI, firms should design applications that promote information sharing and the establishment of a common ground, thereby facilitating internal communication. In addition, managers must be aware of the potential double-edged effect of AI, fostering the capacity to manage the risks it may introduce.

Third, governments should recognize the potential negative impacts of technological development on sustainability. As AI advances, its influence on individuals, organizations, and society is increasing, but so are its challenges, such as discrimination and bias. Our study demonstrates that AI technology can amplify the likelihood of unethical behavior within firms. Governments should monitor the corporate use of AI technology and strengthen ethical regulations on AI technology to reduce the risk of firms making harmful or unethical decisions.

5.4. Limitations

However, it is important to acknowledge the inherent limitations of this study. First, our analysis relies on secondary data, which entails inherent measurement constraints. Specifically, CEO–TMT faultlines are operationalized using five commonly adopted demographic attributes, such as gender, age, education, functional background, and tenure. While widely used, these observable characteristics may not fully capture deeper differences in values, cognitive frames, and strategic orientations between the CEO and other TMT members. Such cognitive and psychological heterogeneity remains difficult to observe using archival data. Similarly, our measure of AI technology, based on the natural logarithm of one plus the ratio of a firm’s AI-related assets to its total assets, may not fully reflect firms’ actual AI capabilities or the extent of AI deployment in practice. Moreover, although our firm-level measure captures variation in AI adoption across firms and over time, it reflects an average effect over the sample period (2010–2023). Given the rapid diffusion of AI, particularly following policy initiatives such as China’s New Generation Artificial Intelligence Development Plan in 2017, the role of AI as a contextual moderator may vary across different phases of technological development. While our additional analyses suggest that the estimated effects are robust to temporal variation, future research could further refine these insights by incorporating primary data (e.g., surveys or interviews) and explicitly modeling the dynamic evolution of AI adoption.

Second, firm-level characteristics such as industry type, firm size, geographic region, firm age, and organizational form were not systematically analyzed in this study. These factors may meaningfully influence management systems, resource availability, and strategic decision-making processes. For instance, firms in heavily regulated industries or located in regions with stronger environmental enforcement may respond differently to CEO–TMT faultlines. Similarly, larger firms may have more formalized decision-making structures that moderate the influence of faultlines on eco-attention. Subdividing the sample along these dimensions could therefore yield different results. Future research should conduct more fine-grained subgroup analyses to examine whether the identified effects hold across different firm types and contextual settings.

Third, our dataset is confined to Chinese A-share listed firms. This may limit the generalizability of our findings to other national contexts. China’s cultural nuances and distinctive institutional environment, including the Confucian emphasis on collective harmony, increasingly stringent environmental regulations, and state-driven policy initiatives such as the dual carbon goals, may shape the identified relationships in ways that are specific to the Chinese context. In institutional environments where environmental pressures are less pronounced or where TMT authority structures differ substantially, the mechanisms linking CEO–TMT faultlines to eco-attention and green innovation may operate differently. Future research should therefore integrate data from diverse countries and regions to enhance external validity. Cross-national comparative studies would be particularly valuable for disentangling the cultural and institutional boundary conditions of the identified effects.