Artificial Intelligence Adoption, Innovation Efficiency, and Governance Mechanisms: Evidence from China

Abstract

The rapid development of artificial intelligence (AI) is reshaping innovation models and enterprises’ competitive advantages. However, existing research has not systematically examined whether AI adoption enhances enterprise innovation or identified its influencing factors. Using panel data from Chinese A-share listed companies between 2012 and 2023, this study constructed an index measuring the extent of AI adoption based on annual report texts and analyzed its impact on enterprise innovation efficiency and underlying mechanisms. The results indicate that AI adoption significantly improves innovation efficiency. Further analysis shows that institutional ownership, executives’ digital background (EDB), and media attention positively moderate this effect. This study provides new evidence for the governance boundaries of AI-driven innovation, enriches enterprise innovation theory in the context of digital transformation, and offers managerial insights for enterprises seeking to realize the value of digital technologies.

1. Introduction

The rapid development of artificial intelligence (AI) is profoundly transforming enterprises’ innovation paradigms and competitive advantages, becoming a key driver of global digital transformation. AI adoption improves operational efficiency and reshapes innovation processes in research and development, product design, process management, and market decision-making [1,2]. Studies indicate that establishing AI labs or embedding AI systems can enhance knowledge integration and organizational innovation performance [3]. In the context of the digital economy, AI is considered a new-generation general-purpose technology, whose degree of internal integration is increasingly critical to the evolution of enterprise technologies and the formation of competitive advantages [4].

Despite AI’s enabling potential, post-adoption innovation performance remains highly heterogeneous. Some firms achieve substantial breakthroughs, while others fall into a “technology investment trap”, where high expenditures fail to generate expected innovation returns [5]. This has prompted those in both academic and industry circles to ask the following: Does AI adoption truly improve enterprise innovation efficiency? Under which governance mechanisms and external conditions can AI’s innovation value be realized?

The relationship between AI and enterprise innovation has emerged as a key focus in academic research. Drawing on dynamic capability theory and the resource-based view, some studies suggest that AI enhances innovation performance by improving knowledge absorption, resource allocation, and learning capacity [6,7]. Empirical evidence shows that AI can promote green innovation, improve patent quality, and optimize technology transformation paths [8,9]. However, three main limitations remain in the literature. First, most studies examine AI’s direct effects on innovation quantity or outcomes, neglecting innovation efficiency, a core indicator of the conversion of R&D inputs into outputs [10,11]. Second, boundary conditions influencing AI’s innovation effects are underexplored, particularly the synergistic roles of governance structures, managerial characteristics, and external oversight [12,13]. Third, empirical evidence largely draws from developed countries in Europe and North America, with limited exploration of the mechanisms underlying AI-driven innovation in emerging markets and the external validity of existing theories [14].

This study addresses these gaps by investigating the impact of AI adoption on enterprise innovation efficiency using panel data from Chinese A-share listed companies from 2012 to 2023. It constructs a moderation framework from a three-dimensional governance perspective, examining the roles of institutional ownership, executives’ digital background (EDB), and media attention. Unlike traditional AI input measures, this study applies text mining to extract AI-related keyword frequencies from corporate annual reports, constructing a firm-level AI adoption index that captures the degree of AI integration and depth of technology application. In addition, this study adopts “innovation efficiency”—defined as ln(1 + number of patents)/ln(1 + R&D investment)—as the core dependent variable, providing a precise measure of AI’s transformation of R&D output.

China offers a highly relevant empirical setting for this research. As one of the world’s fastest-growing AI markets, China has issued strategic plans such as the “New Generation Artificial Intelligence Development Plan” and “Made in China 2025” to support large-scale AI applications in manufacturing, finance, and retail sectors. A significant digital divide exists among Chinese enterprises: some are at the forefront of application, while others remain in the exploratory phase [15]. In addition, China’s capital market exhibits unique governance characteristics: institutional investors vary in structure and capacity, executives’ digital competence is unevenly distributed, and media oversight environments differ widely. These factors provide a rich empirical basis for analyzing how governance mechanisms moderate the relationship between AI adoption and innovation efficiency. Therefore, studying China helps us to understand the mechanisms of AI-driven innovation in emerging economies and tests the applicability and scalability of existing theories in diverse institutional contexts.

This study contributes to the literature in four key ways. First, it shifts the focus from AI’s impact on innovation “output volume” to its influence on “innovation efficiency”, enhancing the explanatory power of R&D transformation theories. Second, it develops a three-pronged governance framework—linking internal cognition, ownership structure, and external supervision—to uncover the multidimensional logic behind AI-enabled innovation. Third, by constructing an AI adoption index through text mining, it improves measurement precision and operationalizability, expanding methodological tools for studying digital transformation. Fourth, using China as a representative emerging market, it extends the external validity of research on AI and innovation efficiency and enriches cross-disciplinary discussions in digital governance, strategic decision-making, and enterprise innovation.

In summary, this study aims to clarify the enabling pathways and governance boundaries of AI in real-world enterprise innovation. It constructs a theoretical logic chain linking AI adoption to innovation efficiency, providing theoretical support and practical insights for promoting high-quality enterprise development and implementing digital strategies.

2. Theoretical Background and Hypotheses

2.1. The Degree of AI Adoption and Enterprise Innovation Efficiency

AI adoption has become a core driver of enterprise innovation performance. In the context of the digital economy, AI is regarded as the “general-purpose technology” of a new technological revolution, and its penetration, empowerment, and disruptive impact on innovation activities are increasingly significant [4]. As a capability resource, AI can indirectly improve overall enterprise performance through process optimization [16]. By leveraging technologies such as natural language processing and computer vision, AI can reconstruct R&D models and decision-making logic, systematically enhancing enterprise innovation efficiency.

First, AI directly enhances knowledge creation and integration. It provides a foundation for innovation and strengthens internal capabilities through data processing and knowledge integration mechanisms [17]. AI can optimize cost structures through automation, thereby increasing enterprise value [18]. Research shows that AI adoption significantly improves firms’ abilities in knowledge mining, market perception, and technology forecasting, leading to substantial gains in product development, process improvement, and service innovation [5]. Within the dynamic capability framework, AI acts as a key enabler that enhances firms’ abilities to sense, seize, and reconfigure market opportunities [19].

Second, AI facilitates more efficient allocation of innovation resources. It reduces financing constraints and agency costs, improving R&D productivity [7]. By integrating internal and external knowledge, AI improves product development performance [20]. AI-assisted resource allocation allows firms to evaluate and execute innovation projects more rapidly, increasing both the volume and conversion rate of R&D outputs [1].

Third, AI improves organizational learning capabilities. It promotes knowledge accumulation and innovation diffusion through data-driven feedback loops, thus forming an “intelligent innovation ecosystem” [21]. AI strengthens absorptive capacity and promotes cross-department collaboration and open innovation [14]. These mechanisms collectively transform AI from an auxiliary tool into a strategic innovation asset.

In summary, the extent of AI adoption is a key determinant of enterprise innovation efficiency. By optimizing resource allocation, improving organizational learning, and enabling intelligent innovation, AI adoption significantly boosts R&D productivity. Firms that intensively adopt AI typically exhibit stronger innovation capabilities, faster product iterations, and higher innovation output. Based on this analysis, this study proposes the following hypothesis:

H1. 

AI adoption improves enterprise innovation efficiency.

2.2. Moderating Effect of Institutional Ownership

Institutional investors are key actors in the capital market and exert a profound influence on enterprise innovation. Their shareholding ratio not only reflects market confidence in the enterprise’s long-term value but also plays a dual role in governance: both restraint and incentive [12]. Drawing on enterprise governance theory and agency theory, institutional investors can mitigate the short-term bias of innovation investment by supervising management, reducing agency costs, and optimizing resource allocation, thereby encouraging enterprises to pursue high-risk, high-return, AI-driven innovation [12].

First, institutional investors foster an environment conducive to AI adoption and innovation by improving information transparency and governance efficiency. Research shows that they can promote technology adoption and knowledge sharing through the “information effect” and “governance effect” in enterprise innovation [22]. With strong professional research capabilities and information-processing advantages, institutional investors help companies identify AI adoption scenarios and evaluate the potential risks and benefits of technological innovation projects [23]. Compared with short-term investors, institutional investors are more likely to encourage companies to implement long-term innovation strategies and strengthen the strategic use of AI in product development and process optimization.

Second, institutional investors encourage management to invest in AI innovation by reducing managers’ professional risks. According to the “career concerns model”, a high institutional shareholding ratio can alleviate the psychological pressure on managers to avoid high-risk innovation due to fear of failure, thereby increasing their willingness to innovate [12]. This aligns with the characteristics of AI investment, which typically involves substantial initial investments and uncertain long-term returns. Institutional investors’ long-term investment preferences and stable shareholding behavior provide assurance for risk-sharing and strategic patience in AI innovation projects.

Third, institutional investors strengthen the link between AI adoption and innovation performance through external governance mechanisms. Empirical research indicates that an increase in passive institutional investors significantly improves both the quantity of innovations and the quality of patents [13]. Such investors provide a “stable capital base” in enterprise governance, enabling enterprises to implement AI projects without excessive pressure from stock price fluctuations. Active institutional investors further influence innovation resource allocation through shareholder meetings and voting rights, promoting the use of AI to enhance innovation performance [12].

Moreover, institutional investors not only improve governance structures but also facilitate the formation of an AI-related innovation ecosystem through capital allocation. They exert a significant positive effect on enterprise innovation, particularly by strengthening the relationship between innovation input (e.g., R&D expenditure) and output (e.g., patent quality) [24]. In firms with high institutional ownership, AI adoption enhances innovation performance, sustainable innovation capability, and market competitiveness. Based on this analysis, this study proposes the following hypothesis:

H2. 

Institutional ownership enhances the positive effect of AI adoption on enterprise innovation efficiency.

2.3. Moderating Effect of Executives’ Digital Background

The characteristics of the top management team are key determinants of enterprise innovation strategies and technology adoption behaviors. According to Upper Echelons Theory, executives’ educational background, professional experience, and cognitive characteristics profoundly influence the speed and strategic orientation of enterprises’ responses to external technological changes [25]. In the context of AI adoption, executives’ digital backgrounds are critical for effective adoption and for unlocking innovation potential. Consistent with Upper Echelons Theory, executives’ education and professional experience shape their technical cognition and strategic preferences, influencing enterprises’ understanding and utilization of emerging technologies [26]. Digital Leadership Theory offers a valuable theoretical lens for understanding how EDB strengthens the link between AI adoption and innovation efficiency. Digital leadership emphasizes leaders’ ability to guide organizations through technological transformation by fostering digital-oriented cultures, encouraging exploratory innovation, and leveraging data-driven decision-making. Empirical evidence shows that digital leadership enhances innovation performance by shaping strategic orientation, cultivating a digital organizational culture [27], and promoting organizational agility and innovative behavior [28]. Executives with digital acumen therefore act as catalysts for digital transformation, aligning AI initiatives with broader innovation strategies and enhancing the firm’s capacity to integrate technological change effectively. In AI-enabled innovation, EDB serves as a bridge connecting technical capabilities and innovation outcomes. Managers with experience in information technology, computer science, and data analysis can better understand AI’s logic and potential, reducing technical and cognitive barriers and improving the overall efficiency of AI adoption and innovation transformation [29].

Digital cognition offers several advantages. First, it enables management to recognize AI’s value in innovation. A digital background helps executives understand AI’s functions in knowledge generation, data mining, and product design and encourages the integration of AI into R&D decision-making and innovation processes, rather than treating it as merely an auxiliary operational tool [30,31].

Second, strategic guidance and resource integration are key to strengthening AI-driven innovation. Executives with technical understanding are more likely to adopt data-driven decision-making, optimize resource allocation, and direct funds, personnel, and information toward AI-enabled high-value projects, thereby improving innovation quality [32,33].

Third, shaping organizational culture is an important pathway. Research shows that digital leaders tend to create an open, inclusive, and trial-and-error-oriented innovation atmosphere; enhance employees’ acceptance of AI; and promote cross-departmental collaboration and knowledge sharing. This cultural orientation significantly improves the absorptive capacity and knowledge conversion rate of AI within enterprises, enabling AI advancements to rapidly transform into innovative outputs. Furthermore, EDB exerts a “bridging effect” through cross-departmental communication and knowledge integration, reducing technology silos and accelerating the diffusion of AI innovation achievements within organizations [34]. Based on these mechanisms, this study proposes the following hypothesis:

H3. 

EDB enhances the positive effect of AI adoption on enterprise innovation efficiency.

2.4. Moderating Effect of Media Attention

Media attention, as an external governance mechanism and social supervision force, plays a key moderating role in the relationship between AI adoption and enterprise innovation performance. The media serves not only as a channel for information dissemination and image construction but also influences strategic decision-making and innovation activities through “reputational pressure” and the “information amplification effect” [35].

First, media attention motivates enterprises to accelerate AI adoption and implementation by shaping external perceptions and market expectations. When enterprises receive high media exposure, management is more likely to use AI projects to signal innovation capability and social responsibility, thereby enhancing trust among external capital markets and the public [36]. Positive media coverage strengthens internal innovation culture, encouraging enterprises to take risks in AI experimentation and improving innovation efficiency.

Second, media attention significantly affects external governance. High levels of media attention prompt management to prioritize transparency in information disclosure and ethical AI governance, enhancing the social legitimacy of AI projects and innovation transformation efficiency [37]. As an external governance mechanism, media supervision helps reduce information asymmetry and prevents management from engaging in “symbolic AI adoption” or “formalist innovation” [38]. For example, media reports on generative AI influence public awareness of AI innovation and guide enterprises toward ethical self-discipline and socially responsible AI adoption, promoting a positive feedback loop between AI technology and innovation activities.

Third, media attention indirectly moderates the relationship between AI adoption and innovation performance by influencing capital market response and investor expectations. Empirical evidence shows that innovation announcements from firms with high media visibility trigger more positive market responses, reflecting favorable investor expectations regarding AI innovation capabilities [39]. This mechanism transmits innovation signals through capital markets, attracting greater resources for AI R&D and enhancing innovation efficiency. Based on these mechanisms, this study proposes the following hypothesis:

H4. 

Media attention strengthens the positive effect of AI adoption on innovation efficiency.

Figure 1 is the research model.

3. Methods

3.1. Research Samples and Data Sources

This study selected data on China’s A-share listed companies from 2012 to 2023 and processed it as follows. First, only companies listed before 2012 were retained. Second, samples classified as ST, *ST, or delisted during the period were excluded. Third, samples with missing key variables were removed. After processing, 19,637 observations remained. Data were sourced from the China Stock Market and Accounting Research (CSMAR) database, the CNIPA Patent Publication Database, the Wind database, and the China Research Data Service Platform (CNRDS). AI-related word frequency data were derived from textual analysis of enterprises’ annual reports. To mitigate the influence of extreme values, all continuous variables were winsorized at the 1% level at both tails. Data were processed using Stata 18.0 and Python 3.8.

3.2. Dependent Variable

China’s Patent Law divides patents into three categories: invention, utility model, and design. An invention is a new technical solution related to a product or method. A utility model pertains to improvements in a product’s shape or structure. A design pertains to the aesthetic appeal and industrial applicability of a product’s shape, pattern, or color.

Following Hirshleifer et al. (2013), patent output was measured using the year of application rather than the year of authorization to align with the timing of innovation and avoid biases caused by authorization delays [10]. The specific procedure was as follows: the total number of patent applications in year t was obtained by summing the three types of patents, denoted as Patent, serving as a measure of overall innovation output. Given the right-skewed distribution of patent counts, all patent-related variables were transformed using the natural logarithm of one plus the original value—LnPatent = ln(1 + Patent)—representing enterprise innovation output based on quantity expansion. Similarly, R&D expenditure was log-transformed as lnRD = ln(1 + RD) to indicate the level of R&D investment in patents.

Innovation efficiency (InnoEff) was calculated as the ratio of ln(1 + Patent filings) to ln(1 + R&D expenditure), reflecting R&D productivity. This measure captures the efficiency of converting R&D expenditures into patentable outputs while addressing diminishing returns on innovation.

3.3. Independent Variable

Following Yao et al. [40], annual report text was used to assess the extent of AI adoption. AI-related terms were identified using a specialized dictionary and machine-counted in the annual reports of listed companies. The dictionary contained 73 AI-related terms, covering expressions such as “AI”, “machine learning”, “deep learning”, “neural networks”, “computer vision/image recognition”, “natural language processing/speech recognition”, “recommendation systems”, “data mining”, “big data”, “distributed computing”, “knowledge graph”, “RPA”, “AR/VR”, “generative AI”, “pre-training”, and “Transformer/large model”, along with synonyms and English abbreviations (e.g., ML, DL, NLP, RPA, LLM, and CNN) [40]. The degree of AI adoption was measured as the natural logarithm of the number of AI-related keywords plus one: (ln(words)).

3.4. Moderator Variable

3.4.1. Shareholding Ratio of Institutional Investor Group

To extract a single institutional investor group from the institutional investor network, Crane et al. (2019) proposed a method based on whether two institutional investors jointly held a large number of shares in the same company [41]. This approach enables the construction of an institutional investor network and the identification of institutional investor groups within it. Specifically, if two institutional investor companies (denoted by i and j, respectively) jointly held more than 5% of the total outstanding shares of at least one common company at the end of quarter t, an association was established (X_ij = 1); otherwise, no association was recorded (X_ij = 0). An adjacency matrix A linking pairs of institutional investors was then constructed. Using matrix A, the institutional investor network was built, and institutional investor groups were extracted. After group extraction, the shareholding ratio was calculated using Equation (1). For a given year t, among institutions holding shares in company i, ${\mathrm{CliqueInstitution}}_{\mathrm{j},\mathrm{t}}$ was computed as

$${\mathrm{CliqueOwnership}}_{\mathrm{i},\mathrm{t}}={\sum }_{\mathrm{j}=1}^{\mathrm{N}}{\mathsf{\lambda }}_{\mathrm{i},\mathrm{j},\mathrm{t}}\hspace{1em}1\left({\mathrm{CliqueInstitution}}_{\mathrm{j},\mathrm{t}}\right)$$

(1)

where ${\mathsf{\lambda }}_{\mathrm{i},\mathrm{j},\mathrm{t}}$ represents the proportion of shares of company $\mathrm{i}$ held by organization $\mathrm{j}$ in quarter $\mathrm{t}$ relative to the company’s outstanding shares, and 1$({\mathrm{CliqueInstitution}}_{\mathrm{j},\mathrm{t}})$ indicates a dummy variable equal to 1 if organization j belongs to a group member, and 0 otherwise.

3.4.2. Executives’ Digital Background

EDB influences digital decision-making. This study followed Yuan [42] to construct a dummy variable for EDB based on professional information of directors and supervisors. Executives with academic majors in “information, intelligence, software, electronics, communications, systems, networks, automation, wireless, or computers” were defined as having a digital background. If any executive in a listed company had such a background, the indicator was assigned a value of 1; otherwise, it was 0 [42].

3.4.3. Media Supervision

Following Li et al. (2022), media supervision was measured using the number of positive, negative, and neutral reports from the CNRDS financial database and the Janis–Fadner coefficient (J-F), as shown in Equation (2) [36]:

$$\mathrm{J}-\mathrm{F} \mathrm{coefficient}\left\{\begin{array}{ll}{\displaystyle \frac{{\mathrm{e}}^2 - \mathrm{ec}}{{\mathrm{t}}^2}}& \mathrm{if} \mathrm{e} > \mathrm{c}\\ {\displaystyle \frac{\mathrm{ec} -{ \mathrm{c}}^2}{{\mathrm{t}}^2}}& \mathrm{if} \mathrm{e} < \mathrm{c}\\ 0& \mathrm{if} \mathrm{e} = \mathrm{c}\end{array}\right.$$

(2)

where e is the number of positive media reports, c is the number of negative media reports, and t is the sum of positive and negative reports. The J-F coefficient ranges from −1 to 1. The closer the coefficient is to 1, the more positive reports there are and the less media supervision pressure enterprises face. Conversely, if negative reports dominate, the coefficient approaches −1, indicating greater pressure from media supervision.

3.5. Control Variable

To account for factors that might affect enterprise innovation efficiency, this study included several control variables identified in recent research [11,43]. These were enterprise size (Size), asset–liability ratio (Lev), return on assets (ROA), largest shareholder’s shareholding ratio (Top1), operating income growth rate (Growth), enterprise age (Age), board size (Board), independent director ratio (Indep), total asset turnover (ATO) [44], and inventory ratio (INV) [45]. Industry (Ind) and year (Year) dummy variables were also included. If a company belonged to the specified industry or year, the variable was set to 1; otherwise, it was set to 0.

Table 1. Variable definition.

Variable Type	Variable Name	Definition	Meaning and Description
Dependent Variable	Enterprise innovation efficiency	InnoEff	ln(1 + Patent)/ln(1 + RD)
Independent Variable	The degree of AI adoption	AI	Take the logarithm of the frequency of AI keywords appearing in the annual report.
Moderator Variable	Institutional ownership	Inst	After constructing the institutional investor network and identifying the institutional investor groups within the network, calculate using the equation.
	Board independence	Indep	The number of independent directors/The total number of board members.
	Executives’ digital background	EDB	Use a dummy variable. If any of the company’s senior executives has a digital background, the indicator is 1; otherwise, it is 0.
	Media attention	Media	Use the Janis–Fadner coefficient.
Control Variable	Enterprise size	Size	The natural logarithm of the total ending assets of enterprises.
	Asset-liability ratio	Lev	The ratio of the company’s total liabilities to total assets at the end of the period.
	Return on assets	ROA	The company’s net profit/total assets.
	Largest shareholder’s shareholding ratio	Top1	The ratio of the shareholding of the largest shareholder to the total share value.
	Board size	Board	The total number of board members (including independent directors and non-independent directors) at the end of the reporting period.
	Independent director ratio	Indep	The number of independent directors divided by the number of board members, expressed as a percentage.
	Total asset turnover	ATO	Operating income/Average total assets.
	Inventory ratio	INV	The balance of loans at the end of the period/The balance of deposits at the end of the period.
	Operating income growth rate	Growth	Operating revenue of this year/Operating revenue of last year.
	Enterprise age	Age	The number of years since the company was established is equal to the current year plus 1 minus the year the company was founded.

is the definition of all variables.

Table 2. Descriptive statistics.

Variables	Sample Size	Mean Value	Standard Deviation	Minimum Value	Median	Maximum Value
InnoEff	19,421	0.039	0.084	0.000	0.000	0.303
AI	19,421	1.121	1.328	0.000	0.693	4.828
Inst	19,421	0.401	0.259	0.003	0.396	0.918
EDB	19,421	0.247	0.431	0.000	0.000	1.000
Media	19,421	0.169	0.209	−0.399	0.167	0.659
Size	19,421	22.163	1.284	20.065	21.939	26.434
Lev	19,421	0.381	0.190	0.049	0.371	0.826
ROA	19,421	0.047	0.064	−0.208	0.045	0.227
Top1	19,421	0.327	0.145	0.081	0.302	0.725
Age	19,421	2.925	0.322	1.946	2.944	3.555
Growth	19,421	0.156	0.325	−0.475	0.107	1.727
Board	19,421	2.104	0.196	1.609	2.197	2.639
Indep	19,421	0.378	0.054	0.333	0.364	0.571
ATO	19,421	0.632	0.384	0.109	0.550	2.394
INV	19,421	0.123	0.090	0.000	0.107	0.472

presents the descriptive statistics of the study.

3.6. Model Design

To empirically test the proposed hypotheses, this study employed a two-way fixed-effects panel regression framework accounting for industry and year effects. Independent variables, control variables, and interaction terms were gradually introduced across Models (1) through (4) to examine baseline and moderating effects.

3.6.1. Benchmark Model

Model (1) tests Hypotheses 1 and 2, where i represents the enterprise and t represents the year. ${\mathrm{InnoEff}}_{\mathrm{i},\mathrm{t}}$ represents the degree of AI adoption and is the core explanatory variable. Controls includes the control variables: enterprise size (Size), asset-liability ratio (Lev), return on assets (ROA), the largest shareholder’s shareholding ratio (Top1), board size (Board), independent director ratio (Indep$\alpha$ is the regression coefficient, ε${\mathrm{\alpha }}_0$ denotes the constant term. To address potential endogeneity and unobserved heterogeneity, a two-way fixed-effects panel model with industry and year fixed effects was used. Firm-level fixed effects were not included because moderator variables such as executives’ digital background (EDB) and institutional ownership exhibit little within-firm variation over time. Including firm fixed effects would absorb these time-invariant or slow-moving variables, making it difficult to identify their moderating roles.

$${\mathrm{InnoEff}}_{\mathrm{i},\mathrm{t}}={\mathrm{\alpha }}_0+{\mathrm{\alpha }}_1{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\alpha }}_2{\mathrm{Controls}}_{\mathrm{i},\mathrm{t}}+\sum \mathrm{year}+\sum \mathrm{ind}+{\mathsf{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(3)

3.6.2. Moderating Effect Model

Models (4)–(6) examine the boundary conditions of AI adoption and test Hypotheses 2–4. The following regression models were developed based on the benchmark model:

$${\mathrm{InnoEff}}_{\mathrm{i},\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\beta }}_2{\mathrm{Inst}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\beta }}_3{{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}*\mathrm{Inst}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\beta }}_4{\mathrm{Controls}}_{\mathrm{i},\mathrm{t}}+\sum \mathrm{year}+\sum \mathrm{ind}+{\mathsf{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(4)

$${\mathrm{InnoEff}}_{\mathrm{i},\mathrm{t}}={\mathsf{\gamma }}_0+{\mathsf{\gamma }}_1{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\gamma }}_2{\mathrm{EDB}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\gamma }}_3{{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}*\mathrm{EDB}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\gamma }}_4{\mathrm{Controls}}_{\mathrm{i},\mathrm{t}}+\sum \mathrm{year}+\sum \mathrm{ind}+{\mathsf{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(5)

$${\mathrm{InnoEff}}_{\mathrm{i},\mathrm{t}}={\mathsf{\rho }}_0+{\mathsf{\rho }}_1{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\rho }}_2{\mathrm{Media}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\rho }}_3{{\mathrm{AI}}_{\mathrm{i},\mathrm{t}}*\mathrm{Media}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\rho }}_4{\mathrm{Controls}}_{\mathrm{i},\mathrm{t}}+\sum \mathrm{year}+\sum \mathrm{ind}+{\mathsf{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(6)

where ${\mathrm{Inst}}_{\mathrm{i},\mathrm{t}}$ is the moderator variable for institutional ownership in Model (4), ${\mathrm{EDB}}_{\mathrm{i},\mathrm{t}}$ is the moderator variable for EDB in Model (5), and ${\mathrm{Media}}_{\mathrm{i},\mathrm{t}}$ is the moderator variable for media attention in Model (6). If the regression coefficients of the core explanatory variable ${\mathrm{AI}}_{\mathrm{i},\mathrm{t}}$ and its interaction term with a moderator are both significantly positive, the corresponding moderator strengthens the effect of AI on enterprise innovation efficiency. Specifically, institutional ownership, the EDB, and media attention are expected to enhance the positive impact of AI adoption on innovation performance.

4. Results

4.1. Descriptive Statistics

Table 2 presents the descriptive statistics of the main variables. Overall, the skewness and kurtosis of each variable satisfy the requirements for a normal distribution, indicating a reasonable data distribution. The mean of enterprise innovation efficiency (InnoEff) is 0.039, with a standard deviation (SD) of 0.084, a minimum of 0, a maximum of 0.303, and a median of 0.000. This indicates substantial variation in innovation efficiency among the sample enterprises. Most enterprises exhibit low innovation efficiency, while only a few demonstrate high innovation capability, consistent with the pronounced stratification observed in the innovation capabilities of Chinese enterprises. The mean level of AI adoption is 1.121, with an SD of 1.328 and a median of 0.693, indicating substantial variation across enterprises. While some enterprises have achieved a high degree of digitalization, many remain in the early stages of AI adoption—a pattern that is consistent with recent scholarly findings on the “polarization phenomenon” in the digital transformation of China’s manufacturing sector [15]. The mean value of institutional ownership (Inst) is 0.401, with an SD of 0.259, indicating that institutional investors generally hold a relatively high stake in the sample enterprises, although there are substantial differences across firms, reflecting varying levels of market attention to corporate governance. The mean of EDB is 0.247, with an SD of 0.431, suggesting that approximately one-quarter of executives in the sample have a digital background. This indicates that Chinese enterprises have room to improve top-level digital capabilities and that digital leadership varies significantly across firms. The mean value of media attention (Media) is 0.169, with an SD of 0.209 and a median of 0.167, indicating that most enterprises in the sample receive relatively low media coverage, and disparities in public opinion and public supervision across firms are evident.

4.2. Correlation Analysis

As shown in

Table 3. Correlation analysis.

Variables	InnoEff	AI	Inst	EDB	Media	Size	Lev	ROA	Top1	Age	Growth	Board	Indep	ATO	INV
InnoEff	1.000
AI	0.063 ***	1.000
Inst	0.092 ***	−0.107 ***	1.000
EDB	0.063 ***	0.321 ***	−0.061 ***	1.000
Media	0.083 ***	0.098 ***	0.070 ***	0.050 ***	1.000
Size	0.067 ***	−0.007	0.437 ***	−0.081 ***	0.128 ***	1.000
Lev	−0.004	−0.025 ***	0.190 ***	−0.078 ***	0.026 ***	0.536 ***	1.000
ROA	0.032 ***	−0.073 ***	0.110 ***	0.016 **	0.216 ***	−0.035 ***	−0.350 ***	1.000
Top1	0.008	−0.164 ***	0.473 ***	−0.077 ***	0.034 ***	0.134 ***	0.024 ***	0.151 ***	1.000
Age	−0.058 ***	0.035 ***	0.030 ***	−0.086 ***	0.033 ***	0.217 ***	0.174 ***	−0.098 ***	−0.078 ***	1.000
Growth	0.030 ***	−0.005	0.045 ***	0.023 ***	0.147 ***	0.041 ***	0.045 ***	0.297 ***	0.000	−0.096 ***	1.000
Board	0.042 ***	−0.084 ***	0.219 ***	−0.042 ***	0.028 ***	0.262 ***	0.134 ***	0.011	−0.018 **	0.060 ***	0.008	1.000
Indep	−0.024 ***	0.057 ***	−0.059 ***	0.034 ***	−0.017 **	0.001	−0.008	−0.005	0.058 ***	0.007	−0.013 *	−0.574 ***	1.000
ATO	−0.011 *	−0.027 ***	0.134 ***	−0.071 ***	0.038 ***	0.110 ***	0.210 ***	0.203 ***	0.103 ***	0.021 ***	0.169 ***	0.027 ***	−0.011	1.000
INV	0.004	−0.072 ***	0.000	−0.032 ***	−0.007	0.045 ***	0.232 ***	−0.039 ***	0.021 ***	0.033 ***	0.027 ***	0.006	−0.008	0.202 ***	1.000

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

, Pearson’s correlation coefficient was used to examine the correlations among the main variables. The correlation between enterprise innovation efficiency (InnoEff) and the extent of AI adoption is 0.063, significant at the 1% level, indicating that AI adoption modestly promotes enterprise innovation efficiency. This supports Hypothesis 1. All correlation coefficients among the variables are below 0.6, and the variance inflation factor (VIF) values are below 3, indicating no serious multicollinearity in the model.

4.3. Benchmark Regression

As shown in

Table 4. Benchmark regression results.

	(1)	(2)
Variables	InnoEff	InnoEff
AI	0.006 ***	0.005 ***
	(8.581)	(7.122)
Size		0.009 ***
		(12.316)
Lev		−0.020 ***
		(−4.436)
ROA		0.005
		(0.475)
Top1		0.010 **
		(2.155)
Age		−0.015 ***
		(−6.835)
Growth		0.004 *
		(1.956)
Board		0.017 ***
		(4.183)
Indep		−0.012
		(−0.867)
ATO		−0.001
		(−0.665)
INV		0.002
		(0.240)
Constant	0.033 ***	−0.141 ***
	(36.388)	(−7.886)
Observed Value	19,421	19,421
R2	0.049	0.065
Industry Fixed Effect	YES	YES
Year Fixed Effect	YES	YES

Note: The T value is in parentheses; *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

, this study constructed two regression models to explore the influence of enterprise AI adoption on innovation efficiency: one without control variables and one including all control variables. This approach verified the robustness of the effect of AI adoption under different model specifications. The first column presents the regression results excluding the control variables. The coefficient of the core explanatory variable, AI, with respect to the dependent variable, InnoEff, is 0.006 and significant at the 1% level. This indicates that a one-unit increase in ln(1 + AI keyword frequency) is associated with a 0.006 increase in the innovation efficiency index, holding other factors constant. The second column shows the regression results obtained after adding control variables. The effect of AI on InnoEff remains positive and statistically significant, with a coefficient of 0.005 at the 1% significance level. These findings suggest that the extent of enterprise AI adoption promotes improvements in innovation efficiency. Overall, the empirical results indicate that AI adoption significantly enhances enterprise innovation efficiency, supporting the conclusions of Li et al. [46].

4.4. Analysis of Moderating Effect

A Hausman test was conducted before the regression analysis to ensure the accuracy of the regression results. The test indicated that the fixed-effects model was more appropriate. Therefore, the subsequent regression analysis employed a two-way fixed-effects model that included both industry and year fixed effects.

As shown in

Table 5. Results of moderating effect.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	InnoEff	InnoEff	InnoEff	InnoEff	InnoEff	InnoEff
AI	0.005 ***	0.005 ***	0.005 ***	0.004 ***	0.005 ***	0.005 ***
	(7.425)	(7.700)	(6.707)	(6.299)	(6.796)	(6.654)
Inst	0.035 ***	0.035 ***
	(11.906)	(12.002)
AI*Inst		0.008 ***
		(4.286)
EDB			0.006 ***	0.005 ***
			(4.016)	(3.385)
AI*EDB				0.002 **
				(2.107)
Media					0.022 ***	0.023 ***
					(7.743)	(7.842)
AI*Media						0.006 ***
						(2.742)
Size	0.006 ***	0.006 ***	0.009 ***	0.009 ***	0.008 ***	0.008 ***
	(8.434)	(8.422)	(12.366)	(12.376)	(11.587)	(11.508)
Lev	−0.020 ***	−0.020 ***	−0.019 ***	−0.019 ***	−0.021 ***	−0.021 ***
	(−4.460)	(−4.441)	(−4.368)	(−4.366)	(−4.713)	(−4.689)
ROA	0.000	−0.003	0.004	0.004	−0.012	−0.013
	(0.032)	(−0.241)	(0.354)	(0.352)	(−1.076)	(−1.117)
Top1	−0.016 ***	−0.015 ***	0.010 **	0.010 **	0.009 **	0.009 **
	(−3.245)	(−3.091)	(2.170)	(2.257)	(1.978)	(1.975)
Age	−0.015 ***	−0.015 ***	−0.015 ***	−0.015 ***	−0.015 ***	−0.015 ***
	(−6.645)	(−6.628)	(−6.608)	(−6.618)	(−6.866)	(−6.880)
Growth	0.003 *	0.003 *	0.004 *	0.004 *	0.003	0.003
	(1.775)	(1.746)	(1.905)	(1.915)	(1.306)	(1.370)
Board	0.012 ***	0.012 ***	0.016 ***	0.016 ***	0.017 ***	0.017 ***
	(2.862)	(2.911)	(4.065)	(4.068)	(4.096)	(4.103)
Indep	−0.010	−0.008	−0.014	−0.014	−0.010	−0.010
	(−0.727)	(−0.564)	(−0.991)	(−0.982)	(−0.737)	(−0.735)
ATO	−0.003	−0.002	−0.001	−0.001	−0.001	−0.001
	(−1.406)	(−1.320)	(−0.560)	(−0.555)	(−0.511)	(−0.481)
INV	0.005	0.005	0.002	0.001	0.001	0.001
	(0.699)	(0.682)	(0.200)	(0.167)	(0.191)	(0.168)
Constant	−0.082 ***	−0.084 ***	−0.143 ***	−0.143 ***	−0.132 ***	−0.131 ***
	(−4.444)	(−4.528)	(−8.000)	(−8.007)	(−7.384)	(−7.320)
Observed Value	19,421	19,421	19,421	19,421	19,421	19,421
R2	0.072	0.073	0.066	0.066	0.067	0.068
Industry Fixed Effect	YES	YES	YES	YES	YES	YES
Year Fixed Effect	YES	YES	YES	YES	YES	YES

Note: The T value is in parentheses; *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

, this study introduced three moderators—institutional ownership (Inst), EDB, and media attention (Media)—to examine their moderating roles in the relationship between AI adoption and enterprise innovation efficiency. In Column 1, the coefficient of AI is 0.005 and significant at the 1% level. In Column 2, the coefficient of AI × Inst is 0.008, also significant at the 1% level, indicating that institutional ownership positively moderates the relationship between AI adoption and innovation efficiency. In other words, institutional investor participation strengthens enterprises’ innovation orientation and risk identification capabilities in AI adoption, thereby enhancing innovation efficiency. Thus, Hypothesis H2 is supported.

In Column 3, the coefficient of EDB is 0.006, significant at the 1% level. The interaction term AI × EDB in Column 4 is 0.002, significant at the 5% level, indicating that executives’ digital backgrounds enhance the effect of AI adoption on enterprise innovation efficiency. Executives with digital cognition and technical capabilities can better identify the potential value of AI-enabled innovation and improve innovation efficiency through resource integration and strategic guidance, confirming Hypothesis H3.

In Column 5, the coefficient of media attention (Media) is 0.022, significant at the 1% level. The interaction term AI×Media in Column 6 is 0.006, also significant at the 1% level, indicating that media attention positively moderates the relationship between AI adoption and innovation efficiency. External public opinion supervision and public exposure mechanisms improve transparency and social recognition of enterprises’ AI adoption, thereby promoting innovation efficiency, supporting Hypothesis H4.

In summary, institutional investor oversight, executives’ digital literacy, and the media’s public opinion environment significantly enhance the positive effect of AI adoption on enterprise innovation efficiency, demonstrating that external governance mechanisms and top management characteristics play important roles in promoting AI-enabled innovation.

Figure 2, Figure 3 and Figure 4 present marginal effect plots of the interaction between AI adoption and each moderator. Figure 2 shows that firms with higher institutional ownership exhibit a steeper positive slope, indicating that institutional investors enhance AI adoption’s effectiveness in improving innovation efficiency. Figure 3 shows that when EDB = 1, the marginal effect of AI on innovation is significantly stronger, suggesting that digital-savvy executives can better leverage AI to drive innovation outcomes. Figure 4 highlights that higher levels of media attention amplify the positive association between AI and innovation efficiency, consistent with the reputational pressure and signal-amplification mechanisms. These visualizations corroborate the regression results in Table 5 and provide intuitive evidence of the conditions under which AI adoption more effectively fosters innovation.

4.5. Robustness Test

To verify the robustness of the research results, this study examined the following aspects:

(1)

Replacement of the dependent variable. Initially, innovation efficiency was measured as the number of patent applications per unit of R&D investment. To address potential limitations of a single metric, the number of patent applications was replaced with a composite index, calculated as the total number of invention, utility, and design patent applications, weighted in a 3:2:1 ratio, plus the natural logarithm of 1. Innovation efficiency was then recalculated using this composite index.

(2)

Change in the regression model. The innovation efficiency variable contains many zeros and exhibits truncated data characteristics. Therefore, a Tobit model was employed to further examine the effect of enterprise AI adoption on innovation efficiency.

(3)

Exclusion of specific years. To account for the heterogeneous effects of the COVID-19 pandemic on enterprise production, operations, and innovation investment, samples from 2020 and 2021 were excluded, and the regression analysis was repeated.

(4)

Alternative measurement of AI adoption. A machine learning-based variable, implemented, was constructed to capture the number of AI-related sentences classified as “actually implemented” or “in practical use” in annual reports. The bert-base-chinese model was fine-tuned using 5000 manually annotated sentences to distinguish substantive AI applications from symbolic or promotional mentions. The binary classifier was trained with weighted cross-entropy loss, and hyperparameters were optimized on a validation set using grid search. The final model was selected based on the F1 score for implementation-type sentences. The constructed variable counts sentences with predicted label = 1, capturing the semantic depth of AI application. This approach provides a more nuanced proxy for AI adoption than keyword frequency and mitigates concerns of symbolic disclosure bias.

Table 6. Regression results of robustness test.

	(1)	(2)	(3)	(4)
Variables	Replace Dependent Variable	Change Regression Model	Exclude Special Years	AI Variable Substitution
AI implemented				0.002 ***
				(6.797)
AI	0.006 ***	0.022 ***	0.004 ***
	(8.039)	(7.290)	(5.629)
Size	0.002 **	0.031 ***	0.009 ***	0.009 ***
	(2.200)	(9.857)	(10.686)	(9.272)
Lev	−0.020 ***	−0.100 ***	−0.014 ***	−0.029 ***
	(−3.963)	(−4.463)	(−2.856)	(−4.583)
ROA	0.008	0.055	0.003	0.001
	(0.611)	(0.966)	(0.195)	(0.072)
Top1	0.003	0.048 **	0.007	0.007
	(0.701)	(2.276)	(1.491)	(1.017)
Age	−0.022 ***	−0.065 ***	−0.010 ***	−0.023 ***
	(−8.785)	(−6.540)	(−3.880)	(−7.600)
Growth	0.008 ***	0.022 **	0.002	0.001
	(3.864)	(2.343)	(0.668)	(0.326)
Board	0.019 ***	0.088 ***	0.014 ***	0.016 ***
	(4.465)	(4.476)	(3.045)	(2.865)
Indep	−0.010	−0.074	−0.023	−0.015
	(−0.668)	(−1.086)	(−1.447)	(−0.760)
ATO	−0.003 *	−0.019 *	−0.002	0.002
	(−1.671)	(−1.883)	(−1.114)	(0.744)
INV	−0.001	0.013	0.007	0.007
	(−0.078)	(0.330)	(0.766)	(0.648)
Constant	0.033 *	−0.942 ***	−0.145 ***	−0.113 ***
	(1.836)	(−9.878)	(−7.183)	(−4.466)
Observed Value	19,421	19,421	14,761	10,723
R2/Pseudo R2	0.052	0.094	0.062	0.076
Industry Fixed Effect	YES	YES	YES	YES
Year Fixed Effect	YES	YES	YES	YES

Note: The T value is in parentheses; *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

presents the robustness test results. Across all four tests, the regression coefficients for enterprise AI adoption were significantly positive, indicating that higher AI adoption improves innovation efficiency. These results are consistent with the baseline regression, supporting this study’s conclusions.

4.6. Endogeneity Test

As shown in

Table 7. Regression results of instrumental variable method.

	The First Stage	The Second Stage
Variables	InnoEff	InnoEff
L.AI	0.703 ***
	(58.121)
L2.AI	0.145 ***
	(11.940)
AI		0.002 **
		(2.326)
Size	0.016 ***	0.009 ***
	(2.862)	(10.361)
Lev	−0.047	−0.010 *
	(−1.209)	(−1.944)
ROA	0.110	0.016
	(1.070)	(1.202)
Top1	−0.045	0.015 ***
	(−1.191)	(2.660)
Age	−0.015	0.001
	(−0.729)	(0.258)
Growth	0.095 ***	−0.003
	(4.610)	(−1.368)
Board	0.029	0.013 ***
	(0.845)	(2.750)
Indep	0.026	−0.008
	(0.213)	(−0.471)
ATO	0.020	0.004 *
	(1.143)	(1.854)
INV	−0.166 **	−0.002
	(−2.480)	(−0.191)
Observed Value	12,317	12,317
R2	0.832	0.021
Industry Fixed Effect	YES	YES
Year Fixed Effect	YES	YES
Kleibergen–Paap LM Statistic Value		2648.543
		{0.000}
Kleibergen–Paap Wald F Statistic Value		9632.765
		{16.38}
Hansen Test		0.3705

Note: The T value is in parentheses; *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

, to address potential endogeneity, this study employed an instrumental variable approach using the first and second lagged periods of enterprise AI adoption as instruments. As lagged versions of current AI adoption, these variables are closely correlated with the present level of AI adoption and are exogenous, satisfying the key requirements for valid instruments. The “underidentification” test, based on the Kleibergen–Paap rk LM statistic, yielded a p-value less than 0.05, refuting the null hypothesis and indicating that the instruments were sufficiently identified. The weak identification test showed that the Kleibergen–Paap rk Wald F statistic (16.38) exceeded the 10% critical value from the Stock–Yogo test, allowing rejection of the null hypothesis of weak instruments and suggesting no evidence of weak identification. The Hansen J test yielded a p-value greater than 0.1, indicating no over-identification problem. These results confirm that the instrumental variables used in this study are valid.

The regression results of the second stage show that, after controlling for the endogeneity problem, the effect of enterprise AI adoption (the core explanatory variable) on innovation efficiency (the dependent variable) remains positive and significant at the 5% level. This indicates that the conclusions of the baseline empirical analysis are robust.

4.7. Heterogeneity Analysis

To examine whether the effect of enterprise AI adoption on innovation efficiency varies across different types of property rights, the sample was divided into state-owned enterprises and non-state-owned enterprises.

The regression results in

Table 8. Heterogeneity analysis of property rights.

	(1)	(2)
Variables	State-Owned Enterprise	Non-State-Owned Enterprise
AI	0.003 *	0.006 ***
	(1.677)	(7.630)
Size	0.006 ***	0.008 ***
	(3.915)	(9.281)
Lev	−0.003	−0.026 ***
	(−0.312)	(−5.207)
ROA	0.012	0.019
	(0.408)	(1.573)
Top1	0.012	0.000
	(1.162)	(0.046)
Age	−0.017 ***	−0.019 ***
	(−3.110)	(−7.796)
Growth	−0.004	0.006 ***
	(−0.765)	(2.970)
Board	0.003	0.012 **
	(0.306)	(2.512)
Indep	0.037	−0.031 **
	(1.178)	(−1.994)
ATO	−0.004	−0.002
	(−1.017)	(−1.086)
INV	−0.005	0.000
	(−0.310)	(0.048)
Constant	−0.056	−0.089 ***
	(−1.454)	(−4.020)
Observed Value	4550	14,862
R2	0.125	0.063
Industry Fixed Effect	YES	YES
Year Fixed Effect	YES	YES
Inter-group Difference Test (p-value)	0.51

Note: The T value is in parentheses; *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

indicate that, for state-owned enterprises, AI adoption remains positively correlated with innovation efficiency, but the effect is only significant at the 10% level, with a coefficient of 0.003. For non-state-owned enterprises, the coefficient is 0.006 and significant at the 1% level. Both the significance levels and coefficient magnitudes suggest heterogeneous effects of AI adoption across enterprise types. Fisher’s combined p-value test further confirmed the difference between the two groups, significant at the 10% level. These findings suggest that non-state-owned enterprises experience a stronger positive effect of enterprise AI adoption on innovation efficiency compared with state-owned enterprises.

5. Discussion and Conclusions

5.1. Discussion

Existing research generally finds that AI adoption positively affects enterprise innovation efficiency, although the mechanisms of action and boundary conditions remain controversial. Early research focused on the direct effects of AI on knowledge acquisition, process optimization, and product innovation [2,39], emphasizing that machine learning and data analysis can enhance innovation efficiency. However, subsequent studies show that this positive relationship is not universal, with some enterprises experiencing “high input and low output” from AI investments [5].

Empirical evidence further shows that the innovation impact of AI adoption is constrained by internal governance, the external environment, and leader characteristics [6,47]. In particular, when institutional frameworks are imperfect or managerial awareness is insufficient, AI often fails to translate into innovative outcomes [4].

Three main shortcomings remain in prior research. First, the perspective is narrow: most studies examine AI’s impact on innovation solely from technical or R&D perspectives, neglecting the role of governance factors such as equity structure, managerial characteristics, and external supervision [12,13]. Second, there is limited systematic discussion of the mechanisms through which AI affects innovation. Although existing studies verify that AI can improve innovation efficiency, they rarely analyze mediating and moderating mechanisms or the conditions under which AI exerts its influence [21]. Third, empirical samples and methodologies are limited. Most research relies on European, U.S., or high-tech industry samples [14], with limited focus on large samples of listed companies in emerging markets, and it often overlooks endogeneity and heterogeneity issues [48].

This study addressed these gaps in three ways. First, starting with the main line, “AI adoption—innovation efficiency,” it integrates enterprise governance (institutional ownership), the upper echelons (EDB), and external governance (media attention) into a three-dimensional moderation framework, forming a “dual internal and external mechanism”. This multilevel analysis goes beyond previous research conducted from a single technical or organizational perspective, revealing the complex governance logic of AI-driven innovation. Second, unlike prior studies that relied on cross-sectional data, this study employs a two-way fixed-effects model and an instrumental variable approach to address potential endogeneity, using panel data from A-share listed companies between 2012 and 2023 [11]. Robustness and heterogeneity tests further confirm the reliability of the results, addressing sample selection bias and estimation robustness identified in earlier literature [24]. Third, within the context of China’s digital transformation, this study reveals the unique mechanisms of AI-enabled innovation in emerging market institutions. Specifically, institutional investors enhance innovation returns from AI investments through governance effects, EDBs increase enterprise AI adoption [29], and media attention promotes innovation transformation via reputational pressure. These findings extend the applicability of research on AI and innovation efficiency and enrich interdisciplinary theories on digital governance and enterprise innovation performance.

5.2. Conclusions

Amid the accelerating wave of global digitization and intelligence, enterprises face a more complex innovation environment and greater technological transformation challenges. As a key driver of industrial upgrading and enterprise innovation, AI is becoming an important resource for gaining a competitive advantage and achieving high-quality development. AI adoption has reshaped enterprise operating models and profoundly influenced innovation productivity, resource allocation efficiency, and strategic decision-making. Based on a sample of Chinese A-share listed companies, this study systematically explored the impact of AI adoption on enterprise innovation efficiency, examining moderating mechanisms from the perspectives of institutional ownership, EDB, and media attention. The main conclusions are summarized below.

AI adoption significantly promotes enterprise innovation efficiency. Empirical results show that higher levels of AI adoption enhance innovation capability and technological transformation efficiency. AI improves R&D efficiency and innovation output through data analysis and predictive capabilities [49,50,51], highlighting its role in achieving technological breakthroughs and sustaining continuous innovation.

External governance mechanisms strengthen the effect of AI on innovation. Institutional ownership significantly and positively moderates the relationship between AI and innovation efficiency. By leveraging their information advantage and governance oversight, institutional investors promote the transparency and scientific rigor of AI project investment decisions and reinforce the normative and performance-oriented application of technology. This finding suggests that effective capital market supervision can amplify the economic and social benefits of AI-driven innovation.

Management characteristics and the external public opinion environment play significant roles in promoting the effectiveness of AI innovation. EDB significantly enhances the positive impact of AI on innovation efficiency, indicating that executives with technical capabilities enable AI to improve financial stability and innovation decision-making [2,9] and guide the integration of AI with innovation strategies. Similarly, media visibility in AI strategic decisions strengthens external supervision, indirectly promotes innovation investment, and increases enterprise transparency and social responsibility, further improving innovation performance.

In conclusion, AI adoption not only directly improves enterprise innovation efficiency but also generates synergistic effects through external governance, executive capabilities, and social supervision, enhancing the systematization and sustainability of AI-enabled innovation. To maximize these benefits, enterprises should strengthen governance structures, enhance the digital capabilities of management teams, and consider external public opinion supervision to achieve technology-driven, long-term innovation and growth in uncertain environments.

5.3. Revelations

This study has significant theoretical and practical limitations.

From a theoretical perspective, starting from the micro-level of enterprises, this study reveals the positive effect of AI adoption on innovation efficiency and systematically examines the moderating roles of institutional investors, EDB, and media attention, expanding the theoretical boundaries of existing research. The results show that AI functions not only as a technical tool but also as a strategic resource. Its impact on enterprise innovation capability is influenced by the interplay of governance structures and external supervision. This finding enriches interdisciplinary research at the intersection of artificial intelligence, enterprise governance, and innovation performance, providing a new analytical perspective and systematic interpretive framework for innovation theory in the digital economy.

From a practical perspective, enterprises should first define the strategic positioning of AI within their innovation systems, integrating AI into key areas such as R&D, production, management, and market decision-making, and establishing data-driven, intelligence-enabled innovation mechanisms. They should leverage institutional investors’ supervisory and incentive roles to promote rational AI investment allocation and scientific project evaluation through capital governance. Enterprises should also enhance the digital capabilities of executives by recruiting management talent with technical expertise and data analysis skills, thereby improving the top management team’s strategic understanding and execution of AI applications. Furthermore, enterprises should proactively engage with the media, increase transparency in information disclosure, and strengthen social responsibility. They should leverage external reputation mechanisms to boost social recognition and the long-term value of AI innovation.

At the policy level, the government should continue improving AI industry policies; promoting technological R&D, data openness, and collaborative industrial innovation; and creating an institutional environment conducive to AI adoption. Simultaneously, the capital market should be standardized, and institutional investors should be guided to fulfill long-term supervision responsibilities. AI-driven innovative enterprises must be supported in their growth. Senior executives’ digital capabilities should be developed through collaboration between government, industry, academia, and research institutions, with a multi-level digital education and vocational training system to strengthen enterprise digital governance capabilities. Finally, media supervision and information disclosure systems should be improved, and the active role of social supervision in promoting enterprise innovation transparency and ethical compliance should be encouraged, thereby building a high-quality innovation ecosystem based on the trinity of “technological innovation, governance optimization, and social consensus.”

5.4. Limitations and Future Research Directions

This study empirically analyzed a sample of Chinese A-share listed companies from 2012 to 2023. Although the sample is representative, it may not fully reflect conditions in other countries or regions. Future research could be expanded to include different economies, industries, and enterprise types, particularly small and medium-sized enterprises and emerging sectors, to test the universality and robustness of these findings. Broadening the sample and comparative dimensions will enhance the representativeness and international applicability of the results.

This study measures innovation efficiency using the ratio of ln(1 + patents) to ln(1 + R&D), a widely used and tractable indicator for large panel data [11,40]. However, unlike frontier methods such as DEA or SFA, this measure cannot accommodate multiple inputs and outputs or benchmark against best-practice firms. Given the focus on AI’s effect and the inclusion of moderator variables, DEA and SFA were not applied in the main analysis. Future research could employ these methods to further validate the findings.

Using lagged values of AI adoption as instruments is a widely accepted practice, but their validity depends on the strict exogeneity assumption—that is, lagged variables must be uncorrelated with the error term in the innovation efficiency equation. However, if unobserved factors are serially correlated (e.g., persistent strategic orientation, managerial preferences, or omitted variables), this exclusion restriction may be violated. We therefore acknowledge that this approach may not fully address all endogeneity concerns. Future research could explore alternative instruments—such as AI-related government subsidies, regional AI infrastructure rollout, or exogenous AI policy shocks—to strengthen causal identification.

Finally, this study does not examine potential external environmental influences on the relationship between AI adoption and enterprise innovation efficiency. For example, policy support, industry competition patterns, and regional digital infrastructure may moderate AI-enabled innovation. Future research should consider external variables such as the macro-policy environment, industry characteristics, and economic development stage to refine the research framework and reveal the multi-level mechanisms of AI-driven innovation.