Datasets collection
When selecting central state-owned enterprises as research subjects, several criteria must be considered to ensure the representativeness and credibility of the study. The following provides a detailed explanation of the criteria for selecting included in the study:
(1) Industry Representation
Central state-owned enterprises chosen should span diverse industries, including but not limited to energy, manufacturing, finance, and technology. This approach ensures the universality and representativeness of the research results, avoiding confinement to specific circumstances of particular industries.
(2) ESG Performance Level
Priority is given to selecting central state-owned enterprises with higher ESG performance as research subjects. The assessment of ESG performance can be based on publicly available data such as ESG ratings, reports, and sustainable development indices, ensuring that the selected research subjects have a certain level of ESG management and practical experience.
(3) Degree of AI Application
Consider the leadership and maturity of central state-owned enterprises in the application of AI. Select central state-owned enterprises with a certain strength and experience in the application of AI technology as research subjects to ensure comprehensive exploration of the impact of AI on ESG performance.
(4) Data Accessibility
Ensure that the selected central state-owned enterprises have publicly available ESG performance data and relevant corporate information for data collection and analysis. At the same time, central state-owned enterprises are willing to cooperate with the research and provide necessary data support and participation in the study.
To ensure the scientific rigor and consistency of the data, a sample pool was initially constructed by selecting all publicly listed central state-owned enterprises from 2016 to 2022 that met the criteria of publicly disclosed ESG reports and financial data. The sample pool was then stratified by industry category (such as energy, manufacturing, finance, technology, etc.) to ensure a balanced selection of representative enterprises from each industry. Within each industry stratum, the sample was further filtered based on AI usage, leading to the final selection of the study sample. The data for this study spans the period from 2016 to 2022 and includes relevant data from publicly listed central state-owned enterprises, primarily encompassing corporate ESG performance, AI application indicators, and control variable data. The Huazheng ESG Index serves as the primary measure for assessing corporate ESG performance. Compiled by a leading Chinese index provider, this index scores and ranks companies based on their performance across environmental, social, and governance dimensions, offering high authority and reference value. AI usage data is sourced from the annual reports of publicly listed central state-owned enterprises, with natural language processing techniques applied to analyze the text of these reports. The frequency of AI-related terms is calculated to construct a quantitative indicator of AI technology usage. Control variable data is sourced from the Wind Financial Terminal, the China Securities Regulatory Commission (CSRC) database, financial statements disclosed by the Shanghai and Shenzhen stock exchanges, and other relevant sources. Descriptive statistics for all variables are provided in Table 3.
As presented in Table 3, the average AI usage is 3.54, with a standard deviation of 1.09, indicating some variation in the level of AI technology adoption across different central state-owned enterprises. The maximum value is 5.29, and the minimum value is 1.69, reflecting a wide distribution of AI application among the sample enterprises. The mean of the SDI is 63.02, with a standard deviation of 4.87, suggesting some fluctuation in corporate sustainable development performance. ESG performance, with a mean of 68.26, indicates that the sample enterprises generally perform well in terms of environmental, social, and governance aspects. Additionally, control variables such as company size, leverage ratio, and asset turnover also exhibit variability. Notably, Tobin’s Q has a maximum value of 10.67, implying that some companies in the sample have significantly higher market values and more efficient asset allocations compared to others. Overall, the data distribution is reasonable and provides a robust foundation for subsequent regression analysis and empirical research.
A total of 213 questionnaires were distributed, resulting in the collection of 200 valid responses, yielding a response rate of 93.9%. Statistical analysis was performed utilizing SPSS 26.0 software.
Experimental environment and parameters setting
The study primarily employs support vector machines (SVMs) to analyze and model the collected data, investigating the impact mechanism and effects of AI on the ESG performance of central state-owned enterprises.
The specific steps of SVM modeling are outlined as follows:
(1) Data Preparation
ESG performance data and AI application indicator data are gathered from central state-owned enterprises, ensuring data integrity and consistency. The data is then split into training and testing sets, typically using a ratio of 70%-30% or 80%-20%.
(2) Feature Selection
Feature selection is conducted for both ESG performance data and AI application indicator data to identify relevant feature variables pertaining to the research objective. This process aims to reduce model complexity and computational burden.
(3) Data Preprocessing
Data preprocessing is carried out, encompassing tasks such as data cleaning, handling missing values, feature scaling, and data transformation. These steps ensure data quality and reliability.
(4) Model Training
The SVM model is trained using the designated training set. Throughout the training process, appropriate kernel functions (e.g., linear, polynomial, or radial basis function kernels), regularization parameters, and other hyperparameters are selected. For nonlinear problems, various kernel functions and parameter combinations are explored, and the optimal model is determined through methods such as cross-validation. Assume a training dataset as shown in Eq. (7):
$$\{({x}^{(1)},{y}^{(1)}),({x}^{(2)},{y}^{(2)}),…,({x}^{(m)},{y}^{(m)})\}$$
(7)
In Eq. (7), \({x}^{(i)}\) represents s the input sample, \({y}^{(i)}\) signifies the corresponding label, and \(i=\text{1,2},…,m\).
The kernel function \(\text{K}\left({x}^{(i)}\right),{x}^{(j)}\) is utilized to map input samples to a high-dimensional feature space. Commonly employed kernel functions include the linear kernel function, polynomial kernel function, and radial basis function (RBF) kernel function, among others.
The RBF kernel is identified as particularly effective for addressing the nonlinear characteristics present in the data, leading to its selection for the analysis. Concurrently, the regularization parameter C and the kernel function parameter γ are fine-tuned. The parameter C regulates the penalty imposed on the model for misclassifications, while γ influences the decision boundary created by individual training samples. Through a series of experiments, optimal values for C and γ are determined to achieve a balance between the model’s performance on the training set and its generalization capability on new data, thereby providing a robust predictive model for assessing the ESG performance of central state-owned enterprises.
The objective of SVMs is to determine an optimal hyperplane to separate samples from different classes. This task can be transformed into the following convex optimization problem, as depicted in Eq. (8):
$$\underset{\omega , b}{\text{min}}\frac{1}{2}{\Vert \omega \Vert }^{2}+C\sum_{i=1}^{m}max\left(\text{0,1}-{y}^{\left(i\right)}\left({\omega }^{T}{x}^{\left(i\right)}\right)+b\right)$$
(8)
In Eq. (8), \(\upomega \) stands for the normal vector (hyperplane parameter), b denotes the intercept (bias), C refers to the regularization parameter controlling the complexity of the model, \({x}^{\left(i\right)}\) indicates the input sample, \({y}^{\left(i\right)}\) is the corresponding label, and \(i=\text{1,2},\cdots ,m\).
Optimization algorithms such as gradient descent and coordinate descent can be employed to solve the aforementioned optimization problem, obtaining the optimal normal vector \(\omega \) and b. For a new input sample x, the class to which the sample belongs can be determined by calculating the value of \({\omega }^{T}x+b\), and based on the sign of this value. If the value is greater than 0, it is predicted to belong to the positive class; if it is less than 0, it is predicted to belong to the negative class.
To assess the performance of the model, a test set is commonly utilized, and various evaluation metrics are employed, including accuracy, precision, recall, and F1 score. The formulas for these metrics are as follows:
$$\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$$
(9)
$$\text{Precision}=\frac{TP}{TP+FP}$$
(10)
$$\text{Recall}=\frac{TP}{TP+FN}$$
(11)
$$\text{F}1\text{ Score}=\frac{2\times \text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$$
(12)
Here, TP (True Positive) represents the number of true positive instances, i.e., the number of positive class samples correctly predicted as positive; FN(False Negative) represents the number of false negative instances, i.e., the number of positive class samples incorrectly predicted as negative; FP (False Positive) represents the number of false positive instances, i.e., the number of negative class samples incorrectly predicted as positive; TN (True Negative) represents the number of true negative instances, i.e., the number of negative class samples correctly predicted as negative.
Performance evaluation
(1) An analysis of the confusion matrix for the SVM model constructed in this study reveals the following results in Table 4:
Subsequently, examination of the numerical outcomes of the evaluation metrics is presented in Fig. 1:
Numerical results of evaluation metrics for the SVM model.
From the graphical representation, the model’s F1 score stands at 0.878, representing the harmonic mean of precision and recall. The F1 score offers a comprehensive assessment by balancing precision and recall, making it particularly suitable for scenarios involving imbalanced data. In this instance, an F1 score of 0.878 indicates proficient performance of the model in predicting both positive and negative classes. Overall, the model in this example demonstrates robust performance in the prediction task, characterized by high accuracy, precision, recall, and F1 score, thereby showcasing accurate predictions for both positive and negative classes and indicating commendable classification proficiency.
(2) The results of the questionnaire survey are subjected to statistical analysis, with an emphasis on descriptive statistics of the sample and an analysis of group differences. The descriptive statistical findings are illustrated in Fig. 2A–C:
Descriptive statistical results of the study participants (A) gender, (B) age, (C) education.
From the depicted figures, it is evident that the proportion of males and females is equal, with each gender accounting for 50% of the sample, indicating a balanced gender distribution within the respondents. The majority of respondents fall within the age bracket of 20 to 30 years, comprising 40% of the total sample, reflecting the age composition of the survey population. Individuals aged 30 and above represent 45% of the total sample, indicating a significant proportion of middle-aged respondents. Respondents below the age of 20 constitute 15% of the total sample, possibly comprising students or young professionals who recently entered the workforce. Regarding educational attainment, the majority of respondents hold a bachelor’s degree, constituting 60% of the total sample, suggesting a prevalent presence of undergraduate students among the survey participants. Those with a master’s degree represent 25% of the total sample, indicating a proportion of graduate students. Respondents possessing a doctoral degree or higher account for 5% of the total sample, suggesting the inclusion of individuals from academia or senior professionals.
The distribution of gender, age, and education within the sample exhibits a relatively balanced representation, encompassing respondents from diverse age groups and educational backgrounds. Further statistical analysis of respondents’ backgrounds and roles within the company is presented in Table 5.
As illustrated in Table 5, a total of 200 valid questionnaires are collected, representing employees at various levels within central state-owned enterprises, including senior management, middle management, frontline employees, and technical staff. Specifically, 7 senior management personnel participate, accounting for 3.50% of the total sample. This group holds key decision-making and leadership roles within the company and possesses a profound understanding and influence over corporate policies and strategic direction. There are 26 middle management participants, representing 13.00% of the sample. As implementers and coordinators, they exert direct influence over daily operations and employee management. Frontline employees constitute the largest group, with 109 participants, or 54.50%, serving as the foundation of the company’s operations and possessing the most direct experience regarding the organization’s ESG practices. Additionally, 58 technical staff members participate, comprising 29.00% of the sample. They play a crucial role in the company’s technological innovation and application, particularly in the adoption and promotion of AI technologies. This sample distribution provides a comprehensive perspective, allowing for an analysis of AI’s application in ESG practices from the viewpoints of employees at different levels and ensuring the diversity and comprehensiveness of the research findings. Consequently, the sample possesses a certain level of representativeness, enhancing understanding of its characteristics and providing foundational support for subsequent data analysis and research conclusions.
Subsequently, an analysis of questionnaire scores is conducted, assuming a scoring scale of 1–5 for all questions. The scores pertaining to corporate governance (Q1-Q5), environmental protection (Q6–Q10), and social responsibility (Q11–Q15) are depicted in Figs. 3, 4, and 5, respectively:
Scores of corporate governance dimensions.
Scores of environmental protection dimension.
Scores of social responsibility dimensions.
Upon examination of the figures, the average scores for each dimension are as follows: corporate governance: 3.9, environmental protection: 3.7, and social responsibility: 4.2. Correspondingly, the standard deviations are: corporate governance: 0.8, environmental protection: 0.9, and social responsibility: 0.6. The mean score for corporate governance is 3.9 with a standard deviation of 0.8, indicating a generally positive evaluation of corporate governance among respondents, with scores exhibiting relative concentration. For environmental protection, the mean score is 3.7 with a standard deviation of 0.9, suggesting a notable level of concern for environmental issues among respondents, albeit with some divergence in opinions. Regarding social responsibility, the mean score is 4.2 with a standard deviation of 0.6, signifying a prevailing belief among respondents in companies’ effective fulfillment of social responsibilities, with scores demonstrating relative concentration.
Group difference analysis involves comparing the means and standard deviations of key indicators across different groups and conducting statistical tests to derive specific numerical outcomes. To illustrate, the scores of a pivotal question in the questionnaire are examined, with results displayed in Fig. 6:
Results of group difference analysis.
Upon inspection of the figure, it is apparent that the average score for males is 3.5, with a standard deviation of 1.2, whereas the average score for females is 4.4, with a standard deviation of 1.0. This discrepancy suggests that females exhibit a higher average score on Question 10 compared to males.
Subsequently, a t-test is conducted to scrutinize the disparity in scores on Question 10 between genders, as presented in Table 6:
Based on the outcomes of the t-test, the t-value is − 4.37, and the p-value is 0.00002, significantly smaller than 0.05. This indicates a statistically significant difference in scores on Question 10 between males and females. Hence, gender exerts a statistically significant influence on the scores of Question 10.
In the qualitative analysis, content analysis and thematic analysis methods are applied to systematically code and categorize the data, identifying key themes related to the application of AI. The findings of the analysis are summarized in Table 7.
The study reveals that many central state-owned enterprises have extensively applied AI technologies in the environmental dimension, particularly in resource management and pollution control. Through automated monitoring systems and intelligent optimization technologies, these enterprises have been able to more effectively reduce resource consumption, enhance energy efficiency, and decrease emissions. For example, some enterprises have explicitly reported in their annual reports that AI applications have contributed to reduced energy consumption and the optimization of wastewater treatment, thus improving their environmental performance. In the social dimension, AI applications primarily focus on employee welfare and social responsibility management. By utilizing intelligent employee management systems and data analytics platforms, companies have enhanced employee satisfaction and social responsibility awareness. Notably, some companies have employed AI technologies for health monitoring and safety management, especially in production environments. Predictive maintenance has played a key role in reducing workplace injuries and improving employee welfare. In the governance dimension, AI technologies are widely applied to optimize corporate governance structures, especially in risk management, internal audits, and decision support systems. Through AI-driven predictive models and data analysis tools, companies can more accurately identify potential risks and implement effective early warning and prevention measures. The use of these technologies has significantly improved decision-making transparency and efficiency, which has, in turn, bolstered trust among investors and stakeholders.
(3) Correlation Analysis.
A correlation analysis is performed for all variables, and the results are presented in Table 8.
Table 8 presents the correlation analysis results, revealing the linear relationships and significance levels between the variables. SDI is positively correlated with both AI and ESG performance, with correlation coefficients of 0.072 and 0.239, respectively, both significant at the 1% level. This suggests that improvements in AI technology adoption and ESG performance contribute to enhanced corporate sustainable development outcomes. Regarding the control variables, company size (Size) is significantly positively correlated with ESG performance (correlation coefficient = 0.201), indicating that larger companies may have more resources and capabilities to achieve stronger performance in environmental, social, and governance areas. The Lev ratio shows a significant negative correlation with SDI, AI, and ESG, implying that higher debt levels may hinder both corporate sustainable development and the adoption of AI technology. Additionally, total AT and RG rate are positively correlated with SDI, with correlation coefficients of 0.203 and 0.281, respectively. This suggests that efficient asset utilization and income growth foster better corporate sustainable development performance. Tobin’s Q is positively correlated with SDI and AI, but negatively correlated with ESG performance (correlation coefficient = -0.081), reflecting a complex relationship between market value and corporate governance performance. These correlations offer valuable insights for the subsequent regression analysis.
(4) Regression Analysis.
The baseline regression results are presented in Table 9. The first column reports the univariate regression results examining the impact of AI usage on SDI. The second column shows the regression results with industry and year fixed effects included, while the third column presents the regression results with control variables incorporated.
The results from Table 9 indicate that AI usage has a significant positive impact on SDI. In model (1), the regression coefficient for AI usage is 0.002, which is significant at the 1% level, suggesting that AI technology application positively influences sustainable development performance. In model (2), after adding two-way fixed effects for industry and year, the coefficient for AI remains at 0.002, and its significance remains unchanged, confirming the robustness of this result. In model (3), following the inclusion of control variables, the coefficient for AI slightly decreases to 0.001 but continues to be significant at the 1% level, demonstrating that the positive effect of AI on corporate sustainable development performance persists even after accounting for other factors. Among the control variables, Size, total AT, RG rate, and Tobin’s Q exhibit significant positive effects on SDI, while the Lev ratio shows a significant negative effect on SDI. This suggests that corporate resource allocation efficiency, growth capacity, and market performance are key drivers of sustainable development performance. The R2 value of the model increases progressively from 0.005 to 0.370, highlighting that the inclusion of control variables and fixed effects substantially enhances the explanatory power of the model. Overall, the regression results are robust and hold significant academic value.
In summary, listed companies that are larger in size, exhibit lower leverage, higher asset turnover, stronger growth, and better market performance tend to achieve superior sustainable development performance.
(5) ESG Mediating Effect Test.
The regression results for the mediating effect test are presented in Table 10.
Table 10 presents the regression results for the mediating effect test, which confirms the mediating role of ESG in the relationship between AI usage and corporate SDI. In Model Path_a, the regression coefficient for AI on ESG is 0.0007, significant at the 1% level, indicating that AI usage significantly enhances corporate ESG performance, thereby satisfying the first condition for establishing a mediating effect. In Model Path_b, the regression coefficient for AI on SDI is 0.204, also significant at the 1% level, suggesting a direct positive effect of AI on corporate sustainable development performance, which meets the second condition. In Model Path_c, which includes both AI and ESG in the regression, the coefficient for AI decreases to 0.0004 but remains significant at the 5% level. ESG shows a significantly positive coefficient, indicating that ESG partially mediates the relationship between AI and SDI. Control variables such as Size, Lev ratio, total AT, and Tobin’s Q also exhibit significant effects on ESG and SDI, further enhancing the explanatory power of the model.
In summary, the regression analysis supports the partial mediating effect of ESG, suggesting that AI not only directly enhances corporate sustainable development performance but also indirectly promotes performance improvement by enhancing ESG performance.