AI vs. ESG? Uncovering a Bidirectional Struggle in China’s Sustainable Finance

Abstract

As global discourse increasingly centers on environmental, social, and governance considerations, ESG investment has become a major trend in financial markets. Artificial intelligence (AI), through its rapid evolution, has exerted a transformative influence that continues to reshape the fundamental structures of this domain. This study investigates the dynamic relationship between AI and ESG investment indices in China, aiming to reveal the bidirectional causal linkages and time-dependent interactions between these two critical areas. In methods, we used four different parameter stability tests to indicate that the Granger causality test based on the full-sample VAR model may produce biased results. Therefore, we employed a bootstrap rolling-window subsample Granger causality test using data from January 2013 to September 2024 in China. The results reveal a significant dynamic relationship between ESG investment and AI. In key findings, we find that AI exerts a negative impact on ESG investment. AI development attracts substantial capital inflows that favor technological advancement and commercialization over long-term ESG investments. Meanwhile, ESG investment shows both positive and negative effects on AI. The positive effect indicates that ESG investment promotes AI research and applications emphasizing energy efficiency, data privacy, and fairness, thereby supporting the sustainable development of AI technologies. However, driven by short-term economic returns, strict ESG standards and compliance requirements may, in the short term, constrain the development of certain energy-intensive or emerging AI technologies. In economic and political implications, our study provides policymakers with scientific evidence to improve the ESG investment environment and to design balanced policies that support both AI development and sustainable investment practices. It underscores the necessity of promoting coordinated development between AI and ESG investment to achieve global sustainability goals and recommends measures to align short-term economic interests with long-term ESG objectives. This study is expected to serve as a scientific basis for ESG goal-setting and contribute to the realization of China’s dual-carbon goals. In particular, it facilitates the convergence of artificial intelligence technologies with sustainable development initiatives and tells the importance of responsible technological progress for global sustainable development.

1. Introduction

As awareness of environmental, social, and governance (ESG) issues escalates in global society, ESG investment has emerged as a significant trend within financial markets. Investors are increasingly predilected toward investment strategies that underscore sustainability and social responsibility [1]. Departing from traditional investment paradigms, ESG investing emphasizes the assessment of both financial outcomes and the influence of environmental, social, and governance considerations on firm value and risk exposure [2]. It illuminates the attributes of positive externalities and sustainability, aligning closely with China’s current economic development objectives. Concurrently, the utilization of artificial intelligence (AI) technology in the financial sector is becoming pervasive, dramatically altering investment analysis and decision-making processes [3]. This paper examines how AI can enhance ESG investment, aiding investors in more effectively identifying ESG risks and opportunities, while also addressing potential challenges and impacts it may generate.

The 2008 global financial crisis underscored the significance of ESG investment principles for countries worldwide. In its aftermath, governments and regulatory bodies collaborated with international organizations to advance the development of ESG and encourage corporate ESG practices [4]. Irrespective of their scale, companies inevitably influence the ecological environment in which they operate. Concurrently, there has been a heightened focus on ESG-related challenges such as climate change and corporate social responsibility [5]. As environmental considerations gain prominence in today’s corporate landscape, embedding sustainability within strategic decision-making has become a critical priority for both business leaders and investors [6]. ESG mandates that companies not only prioritize financial metrics and profit optimization but also emphasize environmental stewardship, social accountability, and enhanced corporate governance. This approach is essential for fostering sustainable growth in China’s economy, achieving shared prosperity, and constructing an equitable society.

The rapid advancement of AI is profoundly impacting various industries, especially the investment sector. Its superior data processing and analytical capabilities are offering novel opportunities and challenges for ESG investment [7]. In recent years, China has experienced significant growth in AI development. As outlined in the “2024 AI Index Report”, China is the global leader in industrial robot installations and AI patent applications, and it holds second place worldwide in terms of generative AI development [8]. This surge in AI not only propels sustained economic growth in China but also provides technological support for enhancing corporate ESG performance. Since its inception, AI has been considered a pivotal force in driving economic transformation and upgrading, garnering the attention of global governments. The swift progress of AI equips ESG investment with innovative tools and methods. AI technologies, through their capacity for intelligent data interpretation and pattern recognition, streamline ESG-related assessments and contribute to more strategic investment judgments [9]. Although AI and ESG investment are both central themes in current financial and policy discourse, their relationship is not inherently harmonious. From the perspective of sustainable development theory, this interplay can be interpreted through two contrasting lenses: technological optimism and technological pessimism. The former asserts that technological advancements such as AI can play a pivotal role in achieving ESG objectives by enhancing energy efficiency, improving data governance, and promoting algorithmic fairness, and that AI technological development can be mutually reinforcing with ESG investments [10]. The latter warns that AI development and deployment could exacerbate environmental degradation, widen social inequality, and divert capital away from long-term sustainability goals, and that AI technology development and ESG investment could therefore be mutually impeding [11,12]. These conflicting views suggest that the relationship between AI and ESG investment may be more complex than assumed. Given the lack of dynamic and time-sensitive empirical assessments of AI and ESG investment in the existing literature, this study aims to provide timely insights into how AI and ESG investments interact and evolve in tandem. Against this background, this study empirically investigates the dynamic two-way relationship between AI technology and ESG investments in China. Rather than assuming a harmonious interaction between technological innovation and sustainable finance, this study explores if, when, and how these two forces strengthen or weaken each other over time.

The contributions of this study are as follows: First, the existing literature on the economic effects of artificial intelligence has primarily focused on the macroeconomic and industry levels, with insufficient attention given to AI as a pivotal technological driver within the ESG investment framework [13,14]. Moreover, past studies have mainly explored AI’s impact on capital markets in general, rather than focusing specifically on ESG investments, and have failed to examine the potential interrelationship between the two [15,16]. This paper innovatively and systematically investigates the relationship between AI development and ESG investment, which, to some extent, enriches the existing theoretical framework. Second, this research deepens insight into the technological dynamics underlying ESG investing, explicates AI’s contribution to this domain, and generates actionable knowledge for policy design aimed at advancing AI and fostering a more resilient ESG investment. Third, existing analyses often neglect the possibility that external contextual shifts could reshape the dynamic between AI and ESG. To address this, the study employs four different tests for parameter stability, demonstrating that the full-sample technique is unreliable. Therefore, this research uses subsample techniques with smaller estimation errors to identify the time-dependent transmission mechanism between AI and ESG investment. This study focuses on the dynamic interaction between AI and ESG investment indices in China. The primary research tasks are as follows: (1) examine whether there exists a bidirectional causal relationship between AI and ESG investment over time; (2) identify the temporal characteristics and variations in this relationship using rolling-window Granger causality tests; and (3) analyze the implications of these interactions for investment strategies and policy formulation. In this study, we do not pre-define a dependent or independent variable. Instead, we adopt a bidirectional perspective, investigating the mutual and dynamic causal interactions between AI and ESG investment. This approach reflects the theoretical possibility that AI development may influence ESG investment patterns, while ESG-oriented capital allocation may also shape the trajectory of AI research and applications.

The remainder of this paper is organized as follows. Section 2 reviews the existing literature. Section 3 introduces the empirical methodology of our models. Section 4 describes the data used in the study. Section 5 explains our empirical results systematically. Section 6 reaches our conclusions and implications. Section 7 is our research gaps and prospects for this study.

2. Literature Review

Although ESG investment strategies originated from socially responsible investing, tracing back to early ethically based investments with roots in religious beliefs, the term “ESG” was introduced as a distinct concept by The Global Compact in 2004 [17]. This was a move away from earlier views on socially responsible investing and corporate social responsibility, indicating the real start of ESG development. Since this formal introduction, investors and corporations adhering to ESG principles have gradually transitioned from a niche group to a mainstream one. ESG investment practices and corporate adherence to ESG principles have rapidly advanced and developed in a standardized manner [18]. In the digital economy era, the intelligent transformation of ESG implementation is set to accelerate, with the widespread application of intelligent technologies such as big data and artificial intelligence to all facets of ESG. This will create a new model for ESG development driven by intelligent technologies [19]. Over the past few years, there has been considerable academic focus on the integration of AI and ESG investment.

The literature on the impact of artificial intelligence (AI) on environmental, social, and governance (ESG) investing can be divided into two broad perspectives: technological optimism and technological pessimism. From the perspective of technological optimism, applications of AI in ESG investing are increasing. For example, AI is used to construct ESG scoring models and to help investors predict market risk. It can also extract data related to green bonds to evaluate the actual environmental impact of green bond projects [20,21]. The advent of AI allows for an improved collection and analysis of corporate ESG information, forming the basis for reliable ESG reporting and providing support and guidance for ESG investment initiatives [22,23]. In particular, AI addresses the earlier need for extensive manual analysis of corporate reports in making investment decisions. With its more accurate market analysis and decision-support capabilities, AI can assist investors in making more rational ESG investment decisions [24]. Through detailed analysis, AI technology can uncover companies’ potential strengths in environmental governance and social responsibility, thereby further enhancing their investment value [25]. From the perspective of technological pessimism, however, AI in ESG investing also faces certain challenges. As AI-based investment applications become more widespread, some investors who distrust AI refuse to use it, resulting in a reduction in investment and trading activities [26]. The transparency of AI algorithms in the ESG investment process has also come under scrutiny. A lack of interpretability is believed to undermine investors’ trust in AI, leading to calls for more responsible use of the technology [27]. Moreover, if AI is not properly applied, it may introduce potential risks to sustainable investment [28].

Another strand of the literature that we are focusing on is how ESG investments are affecting the development of digital technologies, particularly AI. As with AI’s impact on ESG investing, there exists a dynamic interplay between social value imperatives and technological innovation. Research has shown that strong ESG performance can foster digital innovation by boosting a company’s reputation and improving its ability to secure financing [29]. Conversely, inconsistent ESG ratings may lead investors to question a company’s true ESG credentials, thereby impeding its digital transformation efforts [30]. Although prior research has demonstrated that green finance and other forms of ESG investment can promote AI development, achieving a win–win scenario of ecological sustainability and technological advancement remains an open challenge in the face of rapid AI progress [31,32,33]. Moreover, AI’s formidable capabilities raise concerns regarding data security and privacy, potentially exacerbating social inequalities and deviating from core ESG objectives [34]. The value-driven goals of ESG investment may also constrain the expansion of energy-intensive digital infrastructures—such as data centers and computing facilities—thereby limiting AI-driven technological growth [35]. Consequently, AI research and development within an ESG framework must increasingly prioritize sustainability, and ESG investors must guide AI toward responsible innovation [36].

Overall, AI and ESG criteria present a symbiotic relationship. AI propels the standardized evolution of ESG investment, while the principles of ESG guide the sustainable progression of AI. However, most existing studies focus primarily on the one-way influence between these two areas. There is a noticeable absence of systematic investigation into their reciprocal relationship, particularly concerning how AI alters the efficacy of ESG implementation and how, conversely, ESG requirements shape the trajectory of AI applications [37]. To bridge this gap, our study employs subsample methodologies to discern the temporal reciprocal influence between AI development and ESG investment. Based on the previous analysis, we propose the following hypotheses, aiming to empirically analyze the potential bidirectional relationship between AI and ESG investment. Meanwhile, our study will follow the theoretical framework in Figure 1 and Figure 2.

A.

AI to ESG Investment

H1a. 

AI positively promotes ESG investment by applying to the ESG goal achievement process and enhancing the quality of ESG information.

H1b. 

AI negatively influences ESG investments through high returns on projects and potential environmental threats.

B.

ESG Investment to AI

H2a. 

ESG investments positively impact AI through green and responsible concepts.

H2b. 

ESG investments negatively impact AI by limiting the conduct of highly polluting projects and increasing data security and privacy costs.

3. Methodology

3.1. Full-Sample Technique

In traditional Vector Autoregression (VAR) models, it is essential to follow the standard asymptotic distribution for the Granger causality test statistic; deviations can precipitate analytical errors [38]. To circumvent this, our study leverages the bootstrap (RB) method, optimal for tests aligning with standard asymptotic distributions [39]. Notably, in scenarios with limited sample sizes, Monte Carlo simulations may not adequately enhance the Wald test. The likelihood ratio (LR) test provides a more accurate adjustment for both simulation power and size [40]. Consequently, our research employs the RB-augmented modified LR statistic to examine the interplay between AI and ESGI, ensuring a meticulous analysis of variables with non-normal distributions. We derive the bivariate VAR(p) system in the equation below to illustrate the RB-based modified-LR causality test.

X_t = β₀ + βX_t−1 + …… + β_pX_t−p + ε_t, t = 1, 2, ……, T,

(1)

We used the Schwarz Information Criterion (SIC) to identify the optimal lag order p The variable X is given by X_t = (AI_t + ESG_2t)′, from which Equation (2) can be derived.

$$\left[\begin{array}{c}A{I}_t\\ ESG{I}_{2t}\end{array}\right]=\left[\begin{array}{c}{\beta }_{10}\\ {\beta }_{20}\end{array}\right]+\left[\begin{array}{c}{\beta }_{11}(\mathrm{L}){\beta }_{12}(\mathrm{L})\\ {\beta }_{21}(\mathrm{L}){\beta }_{22}(\mathrm{L})\end{array}\right]\left[\begin{array}{c}A{I}_t\\ ESG{I}_{2t}\end{array}\right]+\left[\begin{array}{c}{\epsilon }_{1t}\\ {\epsilon }_{2t}\end{array}\right],$$

(2)

ε_t = (ε_1t, ε_2t) is a white-noise process; ${\beta }_{\mathrm{ij}}\left(\mathrm{L}\right)={\sum }_{\mathrm{k}=1}^{\mathrm{p}}{\beta }_{\mathrm{ij},\mathrm{k}}{\mathrm{L}}^{\mathrm{k}}, \mathrm{i},\mathrm{j}=1,2$. L are the lag operators; then, ${\mathrm{L}}^{\mathrm{k}}{\mathrm{X}}_{\mathrm{t}}={\mathrm{X}}_{\mathrm{t}-\mathrm{k}}$. We can establish the null hypothesis: there is no significant Granger causality between AI and ESGI, which is expressed as ${\beta }_{12,\mathrm{K}}=0, \mathrm{k}=1,2,\dots \dots ,\mathrm{p}$. Similarly, the reverse hypothesis is ${\beta }_{21,\mathrm{K}}=0, \mathrm{k}=1,2,\dots \dots ,\mathrm{p}$.

3.2. Stability Test of Parameters

Causality tests applied to the full sample generally assume that the Vector Autoregressive (VAR) model’s parameters are consistent throughout the entire period [41]. However, in reality, these parameters frequently exhibit instability, potentially leading to erroneous outcomes. The Sup-F, Ave-F, and Exp-F tests effectively mitigate this issue [42]. The Sup-F test helps identify structural changes in the model, while the other two tests examine the stability of the parameters. Furthermore, the Lc test enables the observation of whether the parameters adhere to a random walk process [43]. Time-varying parameters cause instability in the relationship between AI and ESGI, which can result in inaccurate outcomes when tests are performed on the full sample. To address this, the study adopts a subsample technique to examine and analyze the dynamic relationship between AI and ESGI.

3.3. Subsample Technique

The rolling-window test segments the entire time series into progressively smaller parts, spanning from the beginning to the end [44]. Nevertheless, a trade-off exists between parameter precision and subsample representativeness in this estimation method. The choice of an optimal window size is vital. A larger window size leads to better estimator accuracy but may lessen representativeness, whereas a smaller window size produces the opposite result. This challenge is addressed by stipulating a minimum window width in instances where parameters exhibit instability [45]. This study begins by assuming that the time series length is denoted as T and the rolling-window width as r, with each segment represented as r, r + 1,…, T, leading to T − r + 1 time series. Subsequently, the Granger causality of each subsample is determined using an RB-based revised-LR method, the LR statistics and p-values are organized chronologically based on the subsample methods. The average of all bootstrap estimations, ${\mathrm{N}}_{\mathrm{b}}^{-1}{\sum }_{\mathrm{k}=1}^{\mathrm{p}}{\widehat{\beta }}_{21,\mathrm{k}}^{*}$ and ${ \mathrm{N}}_{\mathrm{b}}^{-1}{\sum }_{\mathrm{k}=1}^{\mathrm{p}}{\widehat{\beta }}_{21,\mathrm{k}}^{*}$, represent the influence of AI on ESGI and the influence of ESGI on AI. $N^b$ represent the frequency of repeated bootstraps, while ${\widehat{\beta }}_{21,\mathrm{k}}^{*}$ and ${\widehat{\beta }}_{21,\mathrm{k}}^{*}$ are parameters of Equation (2). The confidence interval in this section is 90%, with the upper and lower bounds corresponding to the 95th and 5th percentiles.

This method was selected because it allows for statistically robust Granger causality testing without relying on large-sample assumptions or normally distributed errors, which is especially critical in small samples or when the underlying data exhibit structural instability. In our study, given the likelihood of structural breaks and time-varying dynamics between AI and ESGI, it is essential to explore their time-dependent relationship, rather than assuming parameter constancy over the full sample. As demonstrated in Zhao’s study, the full-sample Granger causality test may lead to estimation bias when structural breaks are present; the panel data regression approach does not account for time-varying causal relationships, limiting its applicability to our research focus; additionally, the nonlinear cointegrating autoregressive distributed lag model (NARDL) fails to capture structural shocks across different subsamples, potentially introducing bias into the analysis [46].

Therefore, in light of our research objectives, the bootstrap subsample rolling-window Granger causality test provides a well-balanced approach, combining theoretical rigor with empirical practicality. This approach is particularly suitable for capturing regime shifts, policy changes, or market transitions [47].

3.4. Interpretation Principles of the Empirical Results

To ensure a coherent understanding of the empirical findings, we follow the principles when interpreting the calculations below:

First, in both the full-sample and subsample technique, a p-value below 0.10 (at the 90% confidence level) is considered statistically significant evidence of Granger causality from one variable to another.

Second, the bootstrap-enhanced LR test used in both the full- and subsample settings improves inference robustness in small samples and under non-normal distributions. The mean values of the bootstrapped estimators reflect the average magnitude of causal influence, while the 90% confidence intervals provide a non-parametric measure of statistical reliability. If the confidence interval excludes zero, the causality is considered significant and robust.

Third, in the rolling-window approach, time-varying causality is assessed by observing the evolution of LR statistics and p-values across overlapping subperiods. Periods where the p-values fall below 0.10 indicate temporally significant causality, revealing dynamic interactions between AI and ESGI over time. Non-significant windows suggest either weak or absent influence during those intervals. Fourth, the results of the parameter stability tests (Sup-F, Ave-F, Exp-F, and Lc) serve as critical diagnostics for model validity. A significant Sup-F statistic signals the presence of structural breaks, while significant Ave-F and Exp-F statistics imply broad instability of VAR parameters across the sample. The Lc test, if significant, further indicates that parameters do not follow a random walk process. These findings provide theoretical justification for adopting a rolling-window strategy, as full-sample estimation would otherwise risk bias due to time-varying coefficients.

4. Data

This study uses monthly data from January 2013 to September 2024 to investigate the long-term link between AI and ESG investment in China. The choice of 2013 as the starting point is both conceptually and empirically grounded.

First, the year 2013 was a significant turning point in the development of AI. Through progress in neural network architectures and hardware capabilities, notable breakthroughs were achieved in deep learning, particularly in computer vision, natural language processing, and reinforcement learning [48]. The open-sourcing of deep learning frameworks, such as TensorFlow and PyTorch, on platforms like GitHub around this period further accelerated AI research and adoption [49]. Drawing on existing research for assessing AI development in China, this study employs the Wind Artificial Intelligence Concept Index (AI) (Code: 884201.WI) from the Wind database [50,51], as shown in Table 1. Also, according to the introduction of this index, the base date of the AI index is 31 December 2012, so our sample selection starts from January 2013 onwards.

Second, 2013 was also a turning point for ESG investment. It signaled the transition of ESG from a niche concern to a mainstream investment strategy. This shift is reflected in the significant growth of members in the United Nations-supported Principles for Responsible Investment (PRI), indicating a surge in global interest among investment institutions in ESG [52]. Morgan Stanley’s establishment of the Institute for Sustainable Investing in 2013 underscores Wall Street’s increasing focus on ESG investing. Advancements in policy regulations, data disclosure practices, and investment tools have enabled more investors to incorporate environmental and social factors into their decision-making, propelling ESG investing to become an essential pillar of global finance [53]. Therefore, draw on existing research, we selected the ESG300 (399378.SZ) index from the CNI INDEX of the Shenzhen Stock Exchange as a measure of ESG investment in China [54,55], as shown in Table 1. The index aims to provide reliable ESG investment services and guidance to Chinese domestic and foreign investors.

Therefore, the selected sample period not only captures the formative stages and structural shifts in both AI and ESG investment but also aligns with the availability and reliability of high-frequency data for both indices. This ensures the empirical validity of the dynamic interaction analysis conducted in this study. By analyzing the developmental trends of AI and ESGI, as depicted in the subsequent figure, we can further explore the interaction between AI and ESGI.

Table 1. The definition of variables.

Variables	Definition	Source	Reference
AI	Wind Artificial Intelligence Concept Index (Code: 884201.WI)	Wind database https://www.wind.com.cn	Qin et al. (2024) [50] Chen et al. (2024) [51]
ESGI	ESG300 Index (Code: 399378.SZ)	The Shenzhen Stock Exchange’s CNI INDEX http://www.cnindex.com.cn/	Dai (2022) [54] Wu et al. (2022) [55]

Table 1. The definition of variables.

Variables	Definition	Source	Reference
AI	Wind Artificial Intelligence Concept Index (Code: 884201.WI)	Wind database https://www.wind.com.cn	Qin et al. (2024) [50] Chen et al. (2024) [51]
ESGI	ESG300 Index (Code: 399378.SZ)	The Shenzhen Stock Exchange’s CNI INDEX http://www.cnindex.com.cn/	Dai (2022) [54] Wu et al. (2022) [55]

As shown in Figure 3, both AI and ESG witnessed a significant surge in development post-2013. From 2013 to mid-2015, the AI index surged from approximately 1000 to over 8000, representing an eightfold increase within two years, largely driven by breakthroughs such as AlphaGo. The DeepMind AlphaGo project marked a significant achievement by becoming the first computer Go program to defeat a professional player on a standard 19 × 19 board without handicaps. This remarkable feat was documented in the esteemed journal, Nature, in January 2016 [56]. The advancements demonstrated through this project significantly influenced the application of AI in intricate gaming and reinforcement learning domains, offering a foundational reference for subsequent AI endeavors in fields like gaming, decision-making, and optimization. In contrast, the ESGI index increased from about 1100 to nearly 2600 in the same period. The year 2015 marked a significant milestone for ESG, with the United Nations introducing its 17 Sustainable Development Goals (SDGs). These goals have since provided a guiding framework for worldwide ESG investments, offering investors a set of specific standards to gauge environmental and social impacts [57]. However, the subsequent stock market crash led to a sharp decline in indices, profoundly affecting both AI development and ESG investments. In the subsequent period, the AI sector experienced a volatile downward trajectory, with the AI index dropping from its peak of 8000 to approximately 4000. It was not until 2018, when Google released the BERT model—marking a breakthrough in natural language understanding—that AI regained public and academic attention [58].

In contrast, ESG investment surpassed 2500 in 2015 but fell to around 1800 in 2016 before entering a phase of significant recovery. This rebound can be attributed to a growing number of fund managers in Europe and North America establishing ESG-themed investment funds. As ESG investing gained momentum, a substantial influx of capital began to enter the Chinese market. Notably, in 2017, BlackRock—the world’s largest asset management firm—incorporated ESG principles into its investment framework, thereby propelling ESG into the mainstream investment discourse [59]. In 2018, MSCI, a leading index provider, launched its ESG rating system, equipping global investors with standardized tools to assess corporate ESG performance and contributing to the expansion of the global ESG investment landscape [60]. ESG investment continued its steady ascent, reaching a peak of 3425 during the sample period. Following the global outbreak of COVID-19, however, ESG investments experienced a persistent decline. In contrast, AI maintained an upward trend during the pandemic years (2020–2022), benefiting from its critical applications in healthcare, logistics, and public administration [61]. Although the Russia–Ukraine conflict in 2022 led to a temporary downturn, AI once again surged in 2023 with the launch of ChatGPT by OpenAI. Based on the GPT-3.5 model, ChatGPT introduced conversational natural language interaction, attracting widespread attention and reshaping the trajectory of AI development. The AI index approached 8000 once more, reflecting this renewed momentum [62]. Nevertheless, after 2023, a global economic slowdown severely impacted the Chinese market, causing downturns in both AI and ESG investments [63]. The AI index declined from approximately 8000 in early 2023 to around 5000, while the ESGI continued its downward trend, falling to nearly 2100. The observed data and trend curves suggest that the evolution of AI development and ESG investment does not exhibit a perfectly synchronized pattern, indicating the possibility of a complex Granger causality relationship between the two. Full-sample techniques may be inadequate for capturing such intricate dynamics. Therefore, this study adopts subsample-based methodologies to more effectively identify and characterize this evolving interrelationship.

Table 2. Descriptive statistics for AI and ESGI.

	AI	ESGI
Observations	141	141
Mean	5020.445	2154.663
Median	5420.312	2248.782
Maximum	8244.060	3425.190
Minimum	1048.480	1031.500
Standard Deviation	1672.712	578.759
Skewness	−0.881	−0.294
Kurtosis	3.168	2.595
Jarque–Bera	18.405 ***	2.998
Probability	0.000	0.223

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

presents the descriptive statistics of AI and ESGI. By observing the maximum, minimum, and average values, it can be seen that the selected variables exhibit substantial variability. AI exhibits a much wider range (approx. 7196) compared to ESGI (approx. 2394), indicating more substantial fluctuations in AI index performance over time. The standard deviation of the AI (1672.71) is approximately three times that of ESGI (578.76), further confirming the greater volatility in AI-related market activities. Both AI and ESGI display negative skewness, indicating that they follow a left-skewed distribution. The negative skewness of −0.881 for AI indicates a tendency for extreme low values, whereas ESGI’s skewness is closer to 0, implying a more balanced distribution. The kurtosis of AI and ESGI suggests that they adhere to a leptokurtic distribution, characterized by lower peaks and fat tails. The kurtosis value over 3 for AI (3.168) supports the presence of outliers or extreme market responses. The Jarque–Bera test shows that the null hypothesis of AI conforming to a standard normal distribution is rejected at the 1% level of statistical significance. Therefore, this study employs the RB-based revised-LR technique to address the issue of AI having a skewed distribution.

Over the entire observation period, the AI index exhibits more pronounced spikes and deeper troughs compared to ESGI, suggesting higher volatility and sensitivity to tech innovations and external shocks. Quantitatively, the AI index experienced a dramatic growth of over 700% from 2013 to 2015, followed by high volatility through 2023, while ESGI showed more gradual changes and peaked in early 2021. Statistical analysis confirms this contrast: AI’s standard deviation is nearly triple that of ESGI, with a more left-skewed and leptokurtic distribution (as shown in Table 2), highlighting the dynamic, event-driven nature of AI market sentiment—one that is more reactive to technological shocks compared to ESGI’s relatively steady trajectory driven by policy and regulatory frameworks.

5. Quantitative Analyses and Discussions

5.1. Data Test

This study applied the Dickey–Fuller test, Phillips–Perron test, and Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test to assess the stationarity of AI and ESGI after taking the first-order differences in both. According to

Table 3. The results of unit root tests.

Variables	ADF	PP	KPSS
AI	−9.787 (0) ***	−9.674 (8) ***	0.461 (6)
NUC	−9.188 (0) ***	−9.213 (2) ***	0.282 (1)

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

, the unit root tests reveal that both series are stationary. To test the Granger causality between the two variables, we utilize the VAR model on the full sample.

This study sets the optimal lag order as 3 based on the SIC, with 1000 bootstrap replications. The Granger causality results for the full sample between AI and ESGI are displayed in

Table 4. The outcomes of bootstrap full-sample method.

H0: AI Is Not the Granger Cause of ESG		H0: ESG Is Not the Granger Cause of AI
Statistic	p-Value	Statistic	p-Value
5.2768	0.321	14.160	0.016 **

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

. Based on the bootstrap p-values, we find that AI is not a Granger cause of ESG, but ESG is a Granger cause of AI.

However, the previous literature often assumes that there is no structural change in the time series and that a single Granger causality relationship holds throughout the entire sample period [44]. When structural changes occur, the parameters of the VAR model estimated for AI and ESGI using the full sample would vary over time. As a result, the Granger causality relationship between AI and ESG investing would become unstable. Given the assumption of constant parameters and a single Granger causality relationship throughout the sample period, the full-sample causality test is unreliable, and the resulting conclusions would be invalid.

In this study, we proceed by assessing parameter stability and determining if structural changes occur. To assess the time stability of the parameters in the VAR model with AI and ESGI, we apply the Sup-F, Mean-F, and Exp-F tests developed by Andrews and Ploberger [64]. Additionally, we employ the Lc test proposed by Nyblom and Hansen to test the stability of all parameters in the entire VAR system. The corresponding results are shown in

Table 5. Parameter stability tests.

Tests	AI		ESGI		VAR (s) Process
	Statistics	p-Values	Statistics	p-Values	Statistics	p-Values
Sup-F	51.589 ***	0.000	30.090 ***	0.003	49.562 ***	0.000
Ave-F	15.366 ***	0.006	13.456 **	0.018	15.147	0.331
Exp-F	21.822 ***	0.000	10.824 ***	0.004	20.288 ***	0.000
Lc					3.493 **	0.035

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

and

Table 6. Parameter stability tests in long-run relationship.

	Sup-F	Mean-F	Exp-F	Lc
AI = α + β × ESG	18.640 ***	9.639 ***	6.985 ***	1.050 **
Bootstrap p-value	0.002	0.002	0.001	0.011

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

[43,65].

The findings in Table 5 and Table 6 indicate that, based on the Sup-F, Exp-F, and Ave-F tests, both AI, ESGI, and the VAR system reject the null hypothesis, suggesting that the series and parameters have undergone significant fluctuations over time. This suggests that the chosen variables and coefficients in the VAR system undergo structural changes over time. Additionally, the Lc statistic supports the acceptance of the alternative hypothesis at the 1% significance level, confirming that the VAR(s) system does not follow a random walk. Through a series of tests, we demonstrate that there exists an evolving and dynamic interaction between AI and ESGI.

5.2. Rolling-Window Estimation

5.2.1. The Influence of AI to ESGI

To address structural changes, we use rolling-window estimation to analyze the Granger causality between AI and ESGI. Compared to the full-sample causality test, this approach better captures the causality between the two variables by accounting for time variation across different subsamples. In the rolling-window subsample causality test, we use the RB-based modified-LR causality test to examine the Granger causality between AI and ESGI. The rolling estimates for each subsample are plotted in Figure 4, Figure 5, Figure 6 and Figure 7.

Figure 4 shows that the null hypothesis “AI is not a Granger cause of ESG” is rejected at the 10% level of significance in two sub-periods: December 2016 to December 2017 and December 2020 to February 2021. When this finding is combined with that of Figure 5, it is found that AI significantly negatively affected ESG in these periods. In the first period, from December 2016 to December 2017, AI caused a negative effect on ESG, and the magnitude of this negative impact increased over time. One possible reason for these phenomena is as follows. Donald Trump’s election as President of the United States had a major influence on climate change and environmental policies [66]. During his election campaign in June 2017, Trump cast doubt on climate change and announced that the United States would pull out of the Paris Agreement. This led many global companies and investors to reduce their focus on ESG factors [67], and thus political and regulatory uncertainty may have contributed to the negative impact of AI on ESG investment in this period. Second, China’s AI industry experienced rapid growth in startups and capital influx, particularly with technology giants such as BAT (Baidu, Alibaba, Tencent) making investments and expanding in AI [68]. However, at that time, Chinese companies were relatively behind in ESG practices, and many AI companies did not integrate environmental, social responsibility, and governance into their core strategies. Instead, they focused heavily on technological innovation and market share, which likely resulted in a more noticeable negative impact of AI on ESG investments [69]. Third, during this period, China’s economic growth slowed, entering the “new normal”. In response to the economic slowdown, the Chinese government implemented a series of policy adjustments, focusing on innovation, technology-driven development, and structural reforms [70]. AI technology was widely applied in driving industrial upgrades, but these transformations and innovations tended to focus more on efficiency improvement and short-term profits, rather than long-term sustainable development [71]. This likely led to the neglect of ESG factors, with capital flowing into industries such as semiconductors and chips, which offered higher short-term returns, while ignoring the significant environmental impact and resource consumption associated with the large-scale deployment and application of AI [72].

From December 2020 to February 2021, AI also had a significant negative impact on ESG, but this negative effect began to alleviate in February 2021. The end of 2020 and the beginning of 2021 marked a critical period in the global pandemic. During this time, despite efforts by the Chinese government to control the pandemic, its impact on the national economy was unavoidable, and the market entered a recessionary state [73]. In late 2020 and early 2021, AI was likely applied more toward improving efficiency and stimulating traditional manufacturing industries, with less attention given to environmental and social responsibility factors in medium- and long-term ESG investments [74]. However, on 31 December 2021, with the Chinese State Council announcing that one of China’s domestically developed vaccines had been conditionally approved for market release, hopes for the control of the pandemic were raised globally. The expectation of economic recovery gradually emerged in early 2021, leading businesses to regain investment confidence in ESG and alleviating the previous negative impacts [75].

From the perspective of China’s economic policies, at the end of 2020 and the beginning of 2021, the government’s primary focus was on how to quickly recover the economy. This meant that short-term policy guidance was oriented toward using technologies like AI to help businesses resume production and accelerate investment in AI-related infrastructure [76], rather than using AI to address ESG-related issues, particularly neglecting environmental investments [77]. However, in January 2021, with the inauguration of the Biden administration, the U.S. began implementing more climate-friendly policies. The Biden administration’s policy direction re-emphasized the development of technology to address climate change and announced its intention to rejoin the Paris Agreement [78]. This policy shift signaled new expectations for the global market and encouraged companies worldwide to gradually prioritize ESG issues, particularly sustainable investments in the environmental sector [79]. This may have played a role in alleviating the negative impact in early 2021 and gradually restored confidence in ESG investments within the Chinese market.

In summary, the event-based analysis reveals that the impact of AI on ESG investments in China is highly contingent upon broader macroeconomic, political, and policy environments. In both periods where AI showed a significant negative Granger-causal effect on ESG, December 2016 to December 2017 and December 2020 to February 2021, short-term policy priorities, geopolitical uncertainties, and economic shocks contributed to a deprioritization of ESG concerns in favor of technological advancement and economic recovery. These findings suggest that the relationship between AI and ESG is not inherently contradictory, but rather shaped by the surrounding institutional and economic context. Consequently, aligning AI development with long-term sustainability goals will likely require targeted policy support, stronger ESG regulations, and a market environment that rewards responsible innovation.

5.2.2. The Influence of ESGI to AI

The examination of the impact of ESG on AI is illustrated in Figure 6 and Figure 7. During the periods from September 2015 to April 2016, April 2017 to June 2017, March 2018 to February 2019, February 2020, August 2021 to September 2021, November 2021 to January 2022, February 2023 to May 2023, and March 2024, the hypothesis that “ESG is not a Granger cause of AI” is rejected at the 10% statistical significance level. Specifically, ESG investments exerted a significant positive impact on AI during the periods of September 2015 to April 2016, April 2017 to June 2017, February 2020, February 2023 to May 2023, and March 2024. However, during the periods of August 2021 to September 2021 and November 2021 to January 2022, ESG investments exhibited a significantly negative impact on AI. Furthermore, from March 2018 to December 2019, the impact transitioned from negative to positive, with June 2018 identified as the critical turning point.

During the period from September 2015 to April 2016, global leaders formally reached the “Paris Agreement” at the Paris Climate Conference in December 2015. This agreement had a profound impact on China, prompting both the government and enterprises to increase their focus on sustainable technologies [80]. As an emerging technology, AI was rapidly applied to areas such as environmental monitoring, energy management, and carbon emission control, thereby driving investments in the ESG sector [81]. Moreover, 2016 marked the launch of China’s “13th Five-Year Plan”, which explicitly prioritized green development. This policy shift led to a heightened willingness among enterprises to invest in ESG, particularly in environmental protection and energy efficiency improvements [82]. The application of AI technology became a crucial means of achieving these objectives, with substantial ESG investments directed toward the construction of smart grids, the optimized management of energy-efficient equipment, and intelligent environmental monitoring, all of which contributed to the advancement of AI technology [83].

From April to June 2017, prior to the official release of the “Next Generation Artificial Intelligence Development Plan” by the Chinese government in July 2017, the government actively promoted the potential of AI across various sectors and signaled strong policy support [84]. In this period, ESG investors demonstrated significant interest in AI technologies, viewing AI as an essential instrument for boosting efficiency and promoting sustainable development [85]. This heightened interest spurred increased capital flows into projects integrating AI with sustainable technologies, significantly strengthening the impact of ESG investment on AI. However, as the commercial value of AI applications became increasingly apparent, capital markets began prioritizing the rapid commercialization of AI, particularly in consumer and financial sectors [86]. While ESG-related AI applications generated positive societal impacts, their economic returns were relatively slow in the short term. As a result, investment in ESG-driven AI applications declined [87].

In March 2018, the trade conflict between the U.S. and China intensified as the United States imposed new tariffs on Chinese products and restricted Chinese companies from investing in high-tech industries [88]. Within this environment, AI, as a core area of technological competition between the two countries, faced uncertainties related to supply chain disruptions and technological restrictions, leading to a more cautious attitude among enterprises and investors toward AI [89]. The increasing tensions in U.S.–China relations also led to stricter regulations on cross-border data transfers. Given ESG’s emphasis on data privacy and user rights, the regulatory restrictions could have placed further limitations on AI development, leading to adverse effects on AI [90]. In June 2018, China formalized the “Three-Year Action Plan for Winning the Blue Sky Defense Battle”, which was officially released in July. The plan aimed to significantly reduce air pollution and improve urban environmental quality [91]. It required key industries to undergo technological transformations by implementing intelligent technologies aimed at improving energy efficiency [92]. The integration of ESG investment objectives with AI applications began to generate positive value. Meanwhile, in response to U.S. technological restrictions, China increased its long-term commitment to AI research and development, with the government continuously strengthening policy support for AI [93]. The role of AI in social governance, energy efficiency, and emissions reduction expanded, reinforcing the ongoing positive impact of ESG on AI.

In February 2020, the early impact of the COVID-19 pandemic forced governments worldwide to seek new approaches to economic recovery [94]. China also began emphasizing a “green recovery”, leveraging intelligent technologies to minimize resource waste and enhance supply chain efficiency [95]. Starting in February, China extensively deployed AI to support pandemic control efforts, including disease monitoring, contactless temperature screening, and smart logistics. These applications represent typical areas where ESG investment drives AI development [96]. Consequently, ESG investment significantly and positively contributed to AI advancement.

From August to September 2021, China intensified its regulations on high-energy-consuming and high-pollution industries, leading to the implementation of power restriction policies in certain regions and resulting in electricity and production limitations across multiple areas [97]. Amid energy supply constraints, enterprises prioritized short-term survival and reduced investments in ESG-related AI applications, such as environmental monitoring and sustainable development projects [98]. Concerns over the pace of implementing the “dual carbon” policy and its potential economic impact led to a decline in investment enthusiasm for the integration of ESG and AI [99].

From November 2021 to January 2022, global energy supply tightened, natural gas prices surged, and coal supply–demand imbalances became increasingly evident. The energy crisis forced enterprises to prioritize production and energy security, leading to a continuous decline in their willingness to invest in AI-related ESG projects [100]. Moreover, the crisis prompted some companies to adopt a more cautious stance toward the short-term implementation of “dual carbon” goals, thereby weakening the positive impact of ESG investments on AI [101]. In November 2021, the U.S. Federal Reserve announced a gradual tapering of its quantitative easing policy, raising concerns about global liquidity tightening. Given the high research and development costs and long investment cycles associated with AI, the sector faced reduced capital inflows, particularly in ESG-related applications [102]. Under the combined pressures of policy constraints, the energy crisis, and uncertainties related to the pandemic, businesses and investors increasingly favored projects with immediate returns, further limiting support for the long-term application of AI in the ESG domain [103].

In early 2023, under the “dual carbon” goals of carbon peaking and carbon neutrality, the Chinese government further refined carbon reduction targets for various industries and introduced incentives for green technologies [104]. Between February and March 2023, the application of AI in environmental protection and corporate governance received direct policy support, leading to a significant redirection of ESG investments toward AI-related applications [105]. Additionally, the issuance of green bonds and the influence of capital markets rapidly expanded ESG investments into emerging technology sectors. As an essential tool for improving energy efficiency and curbing pollution, AI attracted considerable financial investment, while green finance provided strong backing for the long-term growth of AI applications [106]. However, due to the global economic slowdown and the persistent issue of local government debt in China [107], some local governments began reducing expenditures on long-term environmental and social governance initiatives in April 2023, allocating more resources to addressing financial difficulties. This led to a certain degree of constraint on ESG funding, directly impacting investments in AI technology [108]. In May 2023, China’s carbon trading market became increasingly active. In particular, as carbon emission quotas tightened, some enterprises sought more intelligent solutions to optimize energy consumption and emissions reduction [109]. The importance of AI in carbon emission monitoring, data analysis, and energy efficiency improvement once again surged [110], leading ESG investments to refocus on advancing AI applications.

In March 2024, ESG had a significant positive impact on AI. During this phase, the Chinese government introduced a range of policies aimed at advancing green technology innovation and promoting the fusion of high-tech industries with the green economy [111]. Notably, the publication of the “Guidance Catalogue for Green and Low-Carbon Transition Industries (2024 Edition)” in February 2024 and the emphasis on “strengthening ecological civilization and advancing green and low-carbon development” in the March 2024 government work report underscored the country’s commitment to sustainability. As a core innovation tool, AI received strong policy support and was widely applied in the green transformation of agriculture, energy, transportation, and manufacturing [112,113]. The implementation of these policies reignited market interest and investment in AI within the sustainable development sector.

In summary, the influence of ESG investments on AI development in China has exhibited significant temporal variation, closely tied to macroeconomic conditions, international relations, energy security, and domestic policy orientations. Periods of strong policy support for green development, such as during the implementation of the Paris Agreement, the early stages of the COVID-19 recovery, and the rollout of dual carbon targets, coincided with a pronounced positive impact of ESG on AI. In contrast, during times of geopolitical uncertainty, energy crises, and economic downturns, capital tended to shift toward short-term priorities, weakening ESG’s influence on AI innovation. These findings suggest that ESG can be a powerful driver of sustainable AI development when supported by coherent policy incentives and a stable investment environment. Sustained coordination between government initiatives, financial instruments such as green bonds, and market mechanisms are therefore essential to maintaining and amplifying the positive feedback loop between ESG priorities and AI advancement.

6. Conclusions and Implications

This study explores the relationship between Artificial intelligence (AI) and environmental, social, and governance (ESG) investment in China. First, using the full-sample Granger causality test, we found that ESG investment (ESGI) is a Granger cause of AI. However, based on parameter stability tests, we also discovered that the relationship between AI and ESGI is unstable. To ensure the accuracy of our results, we employed a rolling-window subsample Granger causality estimation, which further confirms the existence of a bidirectional causal relationship. The findings indicate that AI has a significant negative impact on ESGI, while ESGI exerts a bidirectional influence on AI. The positive effect is reflected in ESGIs recognizing the long-term application value of AI in environmental and social governance, while the negative effect arises from concerns related to energy consumption and data privacy issues associated with AI, leading to investor hesitancy. This study also provides important theoretical insights into the relationship between AI and ESG investments. First, it extends sustainable development theory by highlighting the tension between technological progress and long-term sustainability goals. While AI supports energy efficiency and fairness, its rapid commercialization may shift focus toward short-term economic gains, potentially undermining ESG objectives. Second, by integrating technological optimism and technological pessimism, this research provides a new framework for understanding AI’s dual impact on sustainable development. While AI is seen as a driver of progress, its fast-paced growth may also neglect environmental and social responsibilities. Lastly, the study contributes to financial innovation and technological innovation theories, showing that stringent ESG standards may hinder the development of energy-intensive AI technologies in the short term, thus emphasizing the need for balanced policies that align innovation with sustainability.

Based on these results, we offer the following policy recommendations: To address the fluctuations in the AI-ESG investment relationship, it is important to create stable and well-defined policy frameworks to reduce market uncertainty. By clearly outlining priority areas for AI applications in environmental protection and social governance, policymakers can enhance investor confidence in AI-driven ESG initiatives. Furthermore, pilot projects showcasing AI applications in environmental monitoring, carbon emission management, and social governance should be reinforced to provide ESG investors with clear use cases and return pathways. Government-guided funds and incentive mechanisms should be leveraged to encourage enterprises to explore innovative integrations of AI and ESG principles. Additionally, a more robust data governance framework should be established to ensure compliance in AI applications related to social governance and environmental monitoring. Strengthening data protection legislation and advancing privacy-enhancing technologies can further alleviate investor concerns regarding data risks, thereby fostering the synergistic development of AI and ESG investments. Finally, international cooperation in green technologies and AI should be actively promoted. China should take a proactive role in global sustainable development initiatives and climate governance programs by engaging in technology transfers, aligning standards, and collaborating on joint projects. Such efforts will enhance the nation’s capacity for innovation and competitiveness in the intersection of AI and ESG investments on the global stage.

7. Research Gaps and Prospects

The findings of this study offer valuable insights into the development of artificial intelligence (AI) and ESG investment within the Chinese context. Nevertheless, this research is subject to certain limitations. As the scope of the analysis is confined to China, the conclusions drawn may not be directly generalizable to other countries. Future studies could extend this line of inquiry to different geopolitical and economic contexts. For example, the United States, which is recognized for its leadership in AI technology, and Europe, which is known for its pioneering green policies, may exhibit distinct dynamics. With the rapid advancement of AI, scholars have increasingly highlighted its associated energy consumption and societal implications. In response, technological innovators are actively seeking solutions to mitigate these challenges. Looking ahead, greater attention should be paid to the broader application of AI in the ESG domain and its implications for the achievement of ESG objectives. This will also constitute an important direction for our future research.