AI Product Factors and Pro-Environmental Behavior: An Integrated Model with Hybrid Analytical Approaches

Abstract

Based on three theories—the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and Responsible Environmental Behavior (REB)—the present study proposes a model of AI product factors and pro-environmental behavior. This study aims to investigate AI product factors that promote pro-environmental behavior by examining behavioral intentions to use AI technology. Unlike previous research, which predominantly focused on external variables such as social norms, cost, and inconvenience, or individual variables like demographic and psychological factors, this study emphasizes the underexplored role of technological factors. It integrates the Fuzzy Decision-Making Trial and Evaluation Laboratory (F-DEMATEL), Structural Equation Modeling (SEM), and Artificial Neural Network (ANN) approaches to assess the relationships among constructs. For the F-DEMATEL, opinions were collected from 20 experts in the environmental field, while SEM and ANN data were gathered from 1726 participants in Taiwan. F-DEMATEL results demonstrated causal relationships between external factors (perceived trust, self-efficacy, and perceived awareness) and the main variables of the TAM. Likewise, SEM results revealed that perceived trust (PT), self-efficacy (SE), and perceived awareness (PA) influence the main variables of TAM. However, the direct relationships between PT and behavioral intention (BI) and PA and BI were not significant. PT and PA indirectly influence BI through perceived usefulness (PU) and perceived ease of use (PEOU). The results also established that BI positively influences pro-environmental behavior. The author has also outlined how stakeholders aiming to encourage sustainable environmental behaviors can utilize the study’s findings to protect the environment.

1. Introduction

The growing global population and escalating energy consumption have intensified concerns about resource depletion and environmental issues such as climate change [1]. Addressing these challenges requires more than human effort; it necessitates leveraging advanced technological solutions. Specifically, the integration of machines and innovative technologies has emerged as a critical strategy for enhancing efficiency and effectiveness in mitigating environmental problems. Governments and organizations worldwide have acknowledged the urgency of combating climate change, as exemplified by the UN Climate Change Conference UK 2021 (COP26), which provided a collaborative platform for stakeholders to develop strategies to reduce emissions and preserve natural ecosystems [2]. Advancements in information technology have enabled many developed nations to adopt innovative tools to promote environmental sustainability and rural development [3]. Understanding the technological factors that drive protection motivation and pro-environmental behavior (PEB) is crucial for fostering actionable solutions to climate change and other environmental challenges. Among these technologies, artificial intelligence (AI) has gained prominence in the 21st century for its ability to perform tasks traditionally requiring manual effort and cognitive and non-routine processes [4]. AI offers transformative potential for addressing societal challenges and revolutionizing strategies for environmental conservation [5].

As AI continues to be integrated across various sectors, understanding how AI-related factors influence environmental conservation practices is critical for encouraging pro-environmental behavior and reducing harmful human activities. However, despite AI’s demonstrated effectiveness in addressing numerous challenges, research on environmental protection has largely overlooked the role of AI technological factors as determinants of PEB. Existing studies on PEB have primarily focused on external variables such as social norms, cost, and convenience, and individual variables like demographic and psychological factors [6]. Technological factors, particularly those related to AI, remain underexplored in this context.

This study addresses the gap by investigating the influence of AI product factors on pro-environmental behavior, mediated by behavioral intention to use AI technology. Building on the work of Kelly et al. [7], who emphasized the importance of understanding how acceptance, intention, and willingness to use AI technology translate into behavior, this research explores these dynamics in depth. To achieve this, the study integrates three theoretical frameworks: the Extended Technology Acceptance Model (TAM2), the Unified Theory of Acceptance and Use of Technology (UTAUT), and the Environmentally Responsible Behavior (ERB) model. These frameworks collectively offer a comprehensive perspective on the interplay between human behavior, technology, and environmental challenges. To examine these relationships, the study employs a novel hybrid methodology combining the Fuzzy Decision-Making Trial and Evaluation Laboratory (F-DEMATEL), Structural Equation Modeling (SEM), and Artificial Neural Network (ANN) approaches. This integrated approach enables the exploration of linear and nonlinear relationships among constructs and the identification of compensatory and non-compensatory models.

The study makes several significant contributions to the literature. First, it introduces a hybrid three-stage model that integrates F-DEMATEL, SEM, and ANN to analyze how AI product factors indirectly influence pro-environmental behavior. Second, it advances the application of expert systems and AI methodologies by demonstrating the effectiveness of ANN—a subset of machine learning (ML)—in identifying complex relationships. The objectives of this research are as follows: (1) to explore external factors influencing extended TAM technological factors, (2) to examine how technological factors affect individuals’ intentions to adopt and use AI products, and (3) to investigate the relationship between individuals’ intentions to use these tools and their pro-environmental behavior, particularly among users of environmental AI products.

The structure of the paper is as follows: Section 2 reviews the relevant literature and develops the research hypotheses. Section 3 details the research methodology, including the F-DEMATEL, SEM, and ANN approaches. Section 4 presents the results, while Section 5 discusses the findings. Finally, Section 6 outlines the theoretical and practical implications, as well as the study’s limitations and directions for future research.

2. Literature Review and Hypotheses Development

2.1. Theoretical Background

2.1.1. Extended Technology Acceptance Model (TAM 2)

Various models have been employed to investigate individuals’ intentions regarding technology adoption. Among these, the Technology Acceptance Model (TAM) is one of the most widely utilized frameworks for identifying the factors influencing technology acceptance across diverse contexts and domains [8]. TAM is an extension of the Theory of Reasoned Action (TRA), which posits that human behavior is guided by rational thought. TRA operates on the Principle of Compatibility, which asserts that attitudes and behaviors align only to the extent that they both refer to the same valued outcome or evaluative disposition [9]. According to TRA, attitudes can predict behavior, and intentions directly influence behavior. These attitudes are shaped by subjective norms and beliefs, while situational factors determine the relative importance of these variables [10].

In contrast, TAM focuses on behavioral intention as the primary determinant of an individual’s behavior. It introduces two key constructs: perceived usefulness (PU) and perceived ease of use (PEOU) [11]. According to TAM, external variables—such as media exposure and social influences—affect an individual’s PU and PEOU, influencing their intention to use technology and ultimately driving actual system usage. Despite its widespread adoption, TAM has faced criticism for not including behavior as its ultimate dependent variable. Critics argue that intention to use alone may not be a sufficient predictor of technology usage, as behavior represents the actual outcome and should be considered a means to an end [12]. To enhance its predictive accuracy, TAM has undergone extensions that incorporate additional variables, such as trust and knowledge. One such extension is TAM2, which integrates external and organizational factors to comprehensively explain PU and PEOU. These additional factors, absent in the original TAM, improve its ability to account for the complex dynamics of technology adoption [13].

2.1.2. Unified Theory of Acceptance and Use of Technology (UTAUT)

The Unified Theory of Acceptance and Use of Technology (UTAUT) was formulated by synthesizing eight theoretical acceptance models, including the Technology Acceptance Model (TAM) [14]. UTAUT posits that four key constructs—performance expectancy, social influence, effort expectancy, and facilitating conditions—predict behavioral intentions, subsequently influencing user behavior. These constructs have demonstrated robust predictive power, explaining over 60% of the variance in behavioral intentions across different cultural contexts [14]. Notably, performance expectancy and effort expectancy closely align with TAM’s perceived usefulness (PU) and perceived ease of use (PEOU) constructs, respectively. Given this overlap, the present study focuses on PU and PEOU as primary predictors of intention to use environmental AI products, alongside additional relevant factors. Consequently, performance expectancy and effort expectancy are excluded to avoid redundancy and maintain conceptual clarity.

2.1.3. Theory of Environmentally Responsible Behavior

Lee and Jan [15] and Steg and Vlek [16] described Environmentally Responsible Behavior (ERB) as actions undertaken by individuals or groups aimed at minimizing environmental problems as effectively as possible. Similarly, Akintunde [10] and Hines [17] emphasized that the intention to act is a critical factor influencing ERB. Their model highlights that variables such as intention to act, locus of control (an internalized sense of personal control over life events), attitudes, a sense of personal responsibility, and knowledge collectively determine whether individuals adopt such behaviors [10]. This study focuses specifically on the role of intention to act as a key determinant in fostering individuals’ environmentally responsible behavior, particularly through the use of environmental AI products.

2.1.4. Pro-Environmental Behavior (PEB)

Pro-environmental behavior (PEB) encompasses actions intentionally taken to reduce or mitigate the negative impact of human activities on the natural and built environment [18]. These behaviors include practices such as recycling, conserving water, saving electricity, using sustainable transportation, reducing emissions, consuming eco-friendly products, and engaging in sustainable farming. Given the significant human impact on land, air, and water through resource exploitation and environmental degradation, there have been increasing calls for behavioral changes to address these challenges [19]. As PEB involves actions that benefit the environment and the avoidance of actions that harm it, this study explores how AI products indirectly influence such behaviors through individuals’ intention to adopt these technologies. The investigation relies on self-reported assessments to examine the relationship between AI product usage and pro-environmental behavior.

2.1.5. Artificial Intelligence Products (AI)

Despite the rapid advancements in artificial intelligence (AI), particularly in replacing manual labor with automated systems, a universally accepted definition of artificial intelligence products (AI products) remains elusive [20]. AI is a technology that enables computers to perform intelligent tasks typically requiring human cognition. It represents a transformative capability that can revolutionize business practices, achieve objectives, and address critical societal challenges [5]. In contrast, AI products are tools, applications, or systems that leverage AI technologies to perform tasks traditionally associated with human intelligence, such as reasoning, learning, problem-solving, perception, or language comprehension [20]. Examples of AI products include smart lighting systems, robots, voice assistants, recommender systems, and smart home technologies. These products utilize AI methods such as machine learning, natural language processing, computer vision, and robotics to deliver innovative solutions, improve efficiency, enhance accuracy, and support informed decision-making across various domains.

2.2. Hypothesis Development

2.2.1. The Relationship Between Perceived Trust (PT), Perceived Usefulness (PU), Perceived Ease of Use (PEOU), and Behavioral Intention to Use Environmental AI Products

Trust is a subjective attitude that enables individuals to make decisions involving vulnerability [21,22]. Users who trust technology are more likely to believe that using a device will assist them in achieving their goals. Trust has been integrated into acceptance models to predict individual behavior and is a critical factor influencing attitudes toward technology use [7]. On the other hand, perceived usefulness (PU) serves as one of the most reliable indicators of an individual’s intention to adopt new technology, reflecting the belief that technology enhances daily life [8,11]. Perceived ease of use (PEOU) refers to the degree to which individuals believe that using technology requires minimal effort [11]. While PEOU is significant, it is often considered to have a smaller influence on technology acceptance than PU, as it primarily addresses the technical ease of device use [23].

Extensive research has explored the relationship between trust and the core components of the Technology Acceptance Model (TAM), namely PU and PEOU, as trust is a strong predictor of technology adoption [24]. Contemporary studies continue to support the role of trust within TAM. For example, Choung et al. [25] extended TAM and found that trust positively impacts perceived usefulness. Similarly, Miltgen et al. [26] identified trust as a key factor in adopting AI-based iris scanning technology, while Shin [27] demonstrated that trust in AI predicts both its perceived usefulness and ease of use, echoing earlier findings by Reid and Levy [28]. Building on prior TAM2-related studies that position trust as a predictor of PU and PEOU, the following hypotheses are proposed:

 Hypothesis 1a: 

Perceived trust in AI technology positively influences the perceived usefulness of environmental AI products.

 Hypothesis 1b: 

Perceived trust in AI technology positively influences the perceived ease of use of environmental AI products.

Furthermore, research has shown that trust significantly influences individuals’ behavioral intentions to use technology. For instance, studies by Miltgen et al. [26] and Wu and Cheng [29] revealed that trust directly impacts the intention to adopt biometric systems, highlighting its critical role in technology acceptance. Similarly, the present study posits that trust plays a pivotal role in encouraging the adoption of AI products. Trust directly shapes behavioral intentions and indirectly influences pro-environmental behavior (PEB). In this context, trust in AI products and the reliability of their technology suppliers are fundamental to fostering user confidence and promoting adoption. Based on this premise, the following hypothesis is proposed:

 Hypothesis 1c: 

Perceived trust in technology positively influences behavioral intentions to use environmental AI products.

Since trust influences both PU and PEOU, which in turn affects behavioral intentions (BI), it was anticipated that PU and PEOU would mediate the relationship between trust and BI. Therefore, the following hypotheses are proposed:

 Hypothesis 1d: 

Perceived usefulness mediates the relationship between perceived trust and behavioral intention to use environmental AI products.

 Hypothesis 1e: 

Perceived ease of use mediates the relationship between perceived trust and behavioral intention to use environmental AI products.

2.2.2. The Relationship Between Self-Efficacy (SE), Perceived Usefulness (PU), Perceived Ease of Use (PEOU), and Behavioral Intention to Use Environmental AI Products

The external component of the TAM referred to as self-efficacy (SE), reflects an individual’s confidence in their ability to perform a specific task. In the context of AI, self-efficacy pertains to a person’s perceived capability to use AI products effectively for both professional and everyday purposes. Limited exposure to technology can result in significant doubts about one’s ability to complete tasks, potentially deterring individuals from adopting these technologies [30].

Extensive research has explored the impact of self-efficacy on technology adoption, particularly its influence on perceived usefulness (PU), perceived ease of use (PEOU), and behavioral intention to use a given technology. Individuals with high self-efficacy are generally more confident in learning to use new technologies, enhancing their perception of the technology’s usability and usefulness. This confidence enables them to benefit more from environmental AI products or devices. Conversely, those with low self-efficacy may view the technology as ineffective, resulting in reluctance to adopt it. Research supports the idea that self-efficacy significantly affects behavioral intentions to use AI products and shapes perceptions of PU and PEOU. For example, Zhang et al. [31] identified a positive relationship between self-efficacy and the perceived usability of mobile health services. Similarly, Pang et al. [32] found that self-efficacy plays a critical role in shaping consumers’ intentions to purchase organic food. Based on these findings, the following hypotheses are proposed:

 Hypothesis 2a: 

Self-efficacy positively influences the perceived usefulness of environmental AI products.

 Hypothesis 2b: 

Self-efficacy positively influences the perceived ease of use of environmental AI products.

 Hypothesis 2c: 

Self-efficacy positively influences behavioral intentions to use environmental AI products.

Additionally, prior research has indicated that the relationship between self-efficacy and the intention to use technology is mediated by perceived usefulness (PU) and perceived ease of use (PEOU). For instance, it has been shown that students’ self-efficacy for self-regulated learning indirectly influences their intentions to use online learning platforms through their perceptions of usefulness and ease of use [33]. Similarly, behavioral intention to adopt computer-based evaluations has been found to be significantly influenced by self-efficacy through these mediating factors [34]. Drawing from this empirical evidence, it is anticipated that PU and PEOU will mediate the relationship between self-efficacy and the behavioral intention to use AI products. Based on this reasoning, the following hypotheses are proposed:

 Hypothesis 2d: 

Perceived usefulness mediates the relationship between self-efficacy and behavioral intention to use environmental AI products.

 Hypothesis 2e: 

Perceived ease of use mediates the relationship between self-efficacy and behavioral intention to use environmental AI products.

2.2.3. The Relationship Between Perceived Awareness (PA), Perceived Usefulness (PU), Perceived Ease of Use (PEOU) and Behavioral Intention to Use Environmental AI Products

Perceived awareness (PA) refers to the extent to which individuals are informed about new technological products [35]. It encompasses knowledge and mindfulness of a product’s existence derived from personal experience or available information [36,37]. In the context of AI technological devices, PA reflects the information an individual has about these devices, including their potential benefits and challenges in fostering pro-environmental behavior. Research has shown that individuals with prior experience using technology and AI devices are more likely to perceive these products as useful and easy to use [38]. For example, a study examining the role of awareness in auditors’ perceptions of big data adoption readiness in public sector auditing found that higher levels of awareness correlate with greater perceived usefulness (PU) and perceived ease of use (PEOU) [38]. Building on these findings, this study proposes that perceived awareness indirectly influences behavioral intentions to adopt AI technology through PU and PEOU. Consequently, the following hypotheses are proposed:

 Hypothesis 3a: 

Perceived awareness positively influences the perceived usefulness of environmental AI products.

 Hypothesis 3b: 

Perceived awareness positively influences perceived ease of use of environmental AI products.

 Hypothesis 3c: 

Perceived awareness positively influences behavioral intentions to use environmental AI products.

Additionally, since PU and PEOU are expected to mediate the relationship between perceived awareness and behavioral intention to use AI devices, the following hypotheses are proposed:

 Hypothesis 3d: 

Perceived usefulness mediates the relationship between perceived awareness and behavioral intention to use environmental AI products.

 Hypothesis 3e: 

Perceived ease of use mediates the relationship between perceived awareness and behavioral intention to use environmental AI products.

2.2.4. The Relationship Between Perceived Benefit (PB) and Behavioral Intention to Use Environmental AI Products

Perceived benefit (PB) refers to an individual’s assessment of the positive outcomes of a specific action [39]. PB has been identified as a significant predictor of technology acceptance, including fully automated driving, highlighting its critical role in adopting AI-related healthcare products [39]. Similarly, PB has been shown to exert the greatest influence on adopting autonomous vehicles [40]. In the healthcare sector, individuals are more inclined to adopt medical technologies when they perceive the benefits outweigh potential drawbacks, such as privacy concerns [41]. PB also plays a vital role in other fields, such as financial technology. For instance, a study investigating the factors influencing young Vietnamese individuals’ behavioral intentions to adopt financial technology during the COVID-19 pandemic found that PB was the strongest predictor of FinTech adoption [42]. These findings suggest that individuals are more likely to embrace new technologies if they believe the benefits significantly outweigh the costs. In the context of environmental AI technologies, emphasizing their potential advantages for environmental preservation could substantially enhance users’ behavioral intentions to adopt these innovations. Accordingly, the following hypothesis is proposed:

 Hypothesis 4: 

Perceived benefit positively influences behavioral intention to use environmental AI products.

2.2.5. The Relationship Between Perceived Usefulness (PU) and Behavioral Intention to Use Environmental AI Products

Perceived usefulness (PU) is one of the strongest predictors of an individual’s behavioral intention to adopt new technology [8]. PU is a crucial component of the TAM and has continuously impacted usage and intention. For instance, it was stated that purchasing intentions are significantly impacted by PU [43]. Similarly, another study discovered that PU positively predicted intentions to adopt mobile health (mHealth) applications in the healthcare industry [44]. Likewise, the present study expected that perceived utility would have a major impact on the intention to use AI products for environmental purposes, given the proven relationship between PU and behavioral intention in other contexts. Thus, the study proposed the following hypothesis:

 Hypothesis 5: 

Perceived usefulness positively influences behavioral intention to use environmental AI products.

2.2.6. The Relationship Between Perceived Ease of Use (PEOU) and Behavioral Intention to Use Environmental AI Products

Perceived ease of use (PEOU) refers to an individual’s perception of how effortlessly a technology can be used [11]. The TAM, which contends that consumers’ desire to embrace new technology is significantly impacted by their perception of its usability, includes PEOU as a fundamental element in addition to PU [45]. Both PU and PEOU influence behavioral intention, and a number of studies have found a significant relationship between PEOU and behavioral intention to adopt new technology in an array of disciplines. For example, it was found that digital literacy influences special education teachers’ intention to adopt AI technology by being associated with both perceived usefulness and ease of use [46]. PEOU is anticipated to serve similarly as a predictor of behavioral intention in the context of ambient AI products. Therefore, the following hypothesis is proposed:

 Hypothesis 6: 

Perceived ease of use positively influences behavioral intention to use environmental AI products.

2.2.7. The Relationship Between Social Influence (SI) and Behavioral Intention to Use Environmental AI Products

Social influence (SI) refers to the degree to which individuals perceive that important people believe they should adopt a new system [7,14]. It illustrates how people’s decisions to adopt AI products in their daily lives are influenced by their social surroundings [20]. People may use AI products as a result of social pressure from important others, which may promote more environmentally friendly behavior. Social influence shaped students’ behavioral intentions to use networking tools for learning, according to a study conducted by Alvi [47] on Indian college students. Similarly, it was demonstrated that social influence significantly affected students’ intention to use electronic library services [48]. Furthermore, another study also identified social influence as a significant factor driving the adoption of mobile payment services in India, such as m-wallets and m-banking [49]. These studies highlight the important role of social influence in shaping behavioral intentions toward technology adoption. When individuals use environmental AI products, their behavior can influence their peers’ intention to adopt similar technologies. Based on this reasoning, the following hypothesis is proposed:

 Hypothesis 7: 

Social influence positively influences behavioral intention to use environmental AI products.

2.2.8. The Relationship Between Facilitating Conditions (FCs) and Behavioral Intention to Use Environmental AI Products

Facilitating conditions, introduced through the Unified Theory of Acceptance and Use of Technology (UTAUT), are defined as the degree to which an individual believes that a supportive organizational and technical infrastructure is in place to facilitate system use [14]. Any tools or resources that help people use a system efficiently, such as organizational, technical, administrative, or knowledge support, are considered facilitating conditions [50]. Utilizing AI products frequently needs specialized knowledge, equipment, and technical setup. Access to sophisticated technical support is usually necessary for these products to function successfully [51]. Therefore, if people have access to sufficient technical know-how, training, and organizational support, they are more likely to adopt AI products. Providing the right resources and assistance encourages the adoption and use of AI technologies.

Previous studies have shown that facilitating conditions significantly impact behavioral intention to use technology. For example, studies found that facilitating conditions were a key determinant of behavioral intention to use mobile banking in Pakistan [52,53]. Similarly, another study highlighted that facilitating conditions had a substantial effect on behavioral intention to use mobile payment systems, which subsequently predicted actual use behavior [54]. In the context of environmental AI products, the availability of resources and support is expected to increase users’ willingness to adopt these technologies for pro-environmental behavior. Based on these findings, the following hypothesis is proposed:

 Hypothesis 8: 

Facilitating conditions positively influence behavioral intention to use environmental AI products.

2.2.9. The Relationship Between Behavioral Intention to Use Environmental AI Products (BI) and Pro-Environmental Behavior (PEB)

As Hines [17] argues in the theory of environmentally responsible behavior, the intention to act is a key determinant of environmentally responsible behavior. In this context, individuals more inclined to use environmental AI products are more likely to engage in pro-environmental actions. The findings suggest that behavioral intention is a significant predictor of pro-environmental behavior. Therefore, this study posits that the behavioral intention to use environmental AI products will positively correlate with individuals’ pro-environmental behavior. Based on this evidence, the following hypothesis is proposed:

 Hypothesis 9: 

Behavioral intention to use environmental AI products positively influences pro-environmental behavior.

Based on the hypotheses mentioned above, Figure 1 presents the research framework for SEM.

3. Methodology

This study adopts a unique methodology that integrates three advanced approaches: Structural Equation Modeling (SEM), Fuzzy Decision-Making Trial and Evaluation Laboratory (F-DEMATEL), and Artificial Neural Network (ANN). The F-DEMATEL method uses a dual approach to uncover relationships between factors, making it effective for analyzing complex systems and addressing interrelated issues [55]. F-DEMATEL is based on theoretical principles and expert opinions, in contrast to SEM, which can be prone to bias [56]. SEM collects data through surveys and nonprobability sampling, thoroughly investigating theoretical relationships among variables and allowing simultaneous estimation of all factors to confirm mutual causality [57]. ANN is a versatile method that uses advanced learning algorithms and input, output, and hidden layers [58]. Unlike SEM, ANN does not rely on multivariate assumptions or hypotheses, making it an ideal alternative [59]. Combining F-DEMATEL, SEM, and ANN provides a robust research approach by complementing each methodology. ANN enhances prediction accuracy by connecting input and output data through artificial neurons in hidden layers, while SEM effectively analyzes causal relationships. ANN also detects linear and nonlinear relationships and adapts using artificial intelligence techniques [60]. It identifies the input factors most strongly associated with output factors in the study’s model. At the same time, SEM examines the complex relationships between independent and dependent variables, including latent or measured constructs.

Two distinct studies were conducted: the first applied F-DEMATEL to explore the impact of external factors on the core TAM constructs (perceived usefulness and perceived ease of use), while the second employed SEM and ANN. SEM validated the theoretical framework and examined the relationships between variables, while ANN enhanced prediction accuracy by mapping input and output data through artificial neurons in hidden layers.

The following section provides a comprehensive overview of these three methods, detailing the survey participants, the cross-sectional research design, the sample size determination process, questionnaire development, measurement items, and the data analysis approach.

3.1. F-DEMATEL

The DEMATEL method provides a structured framework for analyzing the intricate relationships and dependencies between variables within a system [61]. It offers a systematic way to evaluate cause-and-effect relationships between different factors and components [57], making it valuable for understanding the framework of variables in complex decision-making scenarios [62]. DEMATEL has proven effective in various fields, including marketing strategies, social marketing, environmental issues, information systems, service quality, and business ecosystems. However, because the DEMATEL method is less suited for addressing the complexities of uncertain situations involving multiple factors, it has been expanded by integrating fuzzy theory. Fuzzy theory uses mathematical methods to address challenging decision-making problems where the relationships between variables are not clearly defined [55,63]. This approach is especially valuable when dealing with systems where variables cannot be precisely articulated, providing an effective means of addressing complex and uncertain situations.

The primary advantage of fuzzy theory is its ability to manage fuzzy conditions and adapt to uncertainties, making it a crucial tool for understanding human preferences [64]. The present study utilizes the F-DEMATEL method, which incorporates linguistic variables and expert input to analyze external factors which are the personal, systemic, or environmental factors outside the core factors of TAM that influence perceived usefulness (PU) and perceived ease of use (PEOU) [65]. These factors, in turn, influence individuals’ intentions to adopt environmental AI products. By establishing cause-and-effect links between criteria and reducing the uncertainty of expert assessments, the F-DEMATEL method enhances the reliability of the analysis [66]. The F-DEMATEL approach follows a series of structured steps, as outlined by [62,67], and these will be detailed in the subsequent section.

3.1.1. Step 1: Select and Collect the Viewpoint of the Experts on the Research Issue

The present research seeks to verify the causal relationships between external factors that influence the TAM main factors (PU and PEOU), using expected responses collected from 20 experts in the environment sector through a survey.

3.1.2. Step 2: Design and Define the Fuzzy Linguistic Scale

This step introduces a fuzzy linguistic scale to handle the inherent ambiguity in human assessments. This scale transforms qualitative evaluations into quantitative measures using triangular fuzzy numbers, which help capture the uncertainty and vagueness of human judgments.

Table 1. The correspondence of linguistic terms and values.

Linguistic Terms	Linguistic Values
Very high influence	4 (0.75, 1.00, 1.00)
High influence	3 (0.50, 0.75, 1.00)
Low influence	2 (0.25, 0.50, 0.75)
Very low influence	1 (0.00, 0.25, 0.50)
No influence	0 (0.00, 0.00, 0.25)

below presents the fuzzy values corresponding to various linguistic terms associated with “influence”, referring to the degree to which one criterion impacts another in a decision-making system. This concept reflects the causal strength and interdependence between elements [64]. “Influence” describes the extent to which one criterion (or factor) affects another, emphasizing the causal relationship between elements. Changes in one element can lead to changes in another, highlighting their interdependence. The elements are not independent but connected, influencing each other to varying degrees. Therefore, “influence” represents the strength of the causal relationship and the mutual dependence between the elements in the decision-making process. Each category is associated with specific fuzzy values, allowing for more accurate and flexible analysis when assessing direct effects in the decision-making process. Table 1 outlines the fuzzy values linked to these linguistic terms, serving as a foundation for translating subjective human evaluations into a structured format that can be analyzed systematically.

3.1.3. Step 3: Compute the Initial Direct-Relation Fuzzy Matrix ${\mathit{Z}}^{\mathit{k}}$

$Z^K$ is generated by having experts create fuzzy pairwise influence connections between components in an n × n matrix, with k representing the number of evaluators. By assigning values i = 1, 2, 3, ..., n for n evaluation criteria, the p experts assess criteria in pairs to develop $Z^{\left(1\right)}$, $Z^{\left(2\right)}$, …, $Z^{\left(p\right)}$. The fuzzy matrix $Z^k$ is the initial direct-relation fuzzy matrix of expert k, followed by Equation (1), which is an n × n non-negative matrix; x_ij indicates the direct influence of component i on factor j; and, when i = j, the diagonal elements x_ij = 0, k = 1, 2, 3.……p.

3.1.4. Step 4: Calculate the Normalize Direct-Relation Fuzzy Matrix

$$r^k=\left({\sum }_{j=1}^n{u}_{ij}^k\right) k=1,2\dots ,p$$

(1)

The linear scale transformation is then applied to compare the criteria, followed by obtaining the normalized direct-relation fuzzy matrix as $X^{\left(K\right)}$.

$$X^{\left(K\right)}=\left[\begin{array}{ccc}\begin{array}{cc}{X}_{11}^{\left(k\right)}& X_{12}^{\left(k\right)}\\ X_{21}^{\left(k\right)}& X_{22}^{\left(k\right)}\end{array}& \cdots & \begin{array}{c}{X}_{1n}^{\left(k\right)}\\ X_{2n}^{\left(k\right)}\end{array}\\ ⋮& \ddots & ⋮\\ \begin{array}{cc}{X}_{n1}^{\left(k\right)}& X_{n2}^{\left(k\right)}\end{array}& \cdots & X_{nn}^{\left(k\right)}\end{array}\right] k=1,2\dots ,$$

(2)

where $X_{ij}^{\left(K\right)}$=$\left(L_{ij,}^{\left(k\right)}{M}_{ij,}^{\left(k\right)}{U}_{ij}^{\left(k\right)}\right)$=$\left({\displaystyle \frac{Z_{ij}^{\left(k\right)}}{r^{\left(k\right)}}}\right)$=$\left({\displaystyle \frac{l_{ij}^{\left(k\right)}}{r^{\left(k\right)}}},{\displaystyle \frac{m_{ij}^{\left(k\right)}}{r^{\left(k\right)}}},{\displaystyle \frac{u_{ij}^{\left(k\right)}}{r^{\left(k\right)}}}\right)$.

The following Equation (4) is used to find the average matrix of superscript for k = 1, 2, …, p

$$\begin{gathered} L=\left[\begin{array}{ccc}{L}_{11}& \cdots & L_{1n}\\ ⋮& \ddots & ⋮\\ L_{n1}& \cdots & L_{nn}\end{array}\right], M=\left[\begin{array}{ccc}{M}_{11}& \cdots & M_{1n}\\ ⋮& \ddots & ⋮\\ M_{n1}& \cdots & M_{nn}\end{array}\right], U=\left[\begin{array}{ccc}{U}_{11}& \cdots & U_{1n}\\ ⋮& \ddots & ⋮\\ U_{n1}& \cdots & U_{nn}\end{array}\right] \\ \mathrm{where} L_{\mathit{ij}}={\displaystyle \frac{1}{p}}\sum \begin{array}{c}p\\ k=1\end{array}{L}_{ij,}^{\left(k\right)}{M}_{\mathrm{ij}}={\displaystyle \frac{1}{p}}\sum \begin{array}{c}p\\ k=1\end{array}{M}_{ij,}^{\left(k\right)}{U}_{\mathrm{ij}}={\displaystyle \frac{1}{p}}\sum \begin{array}{c}p\\ k=1\end{array}{U}_{ij,}^{\left(k\right)} \end{gathered}$$

(3)

3.1.5. Step 5: Generate and Analyze the Structure Model

The total-relation fuzzy matrix T is obtained by normalizing the direct-relation fuzzy matrix. The element tij represents the indirect influence relationship between factors i and j. The influence matrix T indicates the overall relationship of impact among the elements. The total-relation fuzzy matrix (T) is computed using the equation below.

$$T=\underset{m\to \infty }{\lim }\left(X+X^2+X^3\dots {+X}^m \right)=X(1-{X)}^{-1}$$

(4)

$$\begin{gathered} \mathrm{Where} \mathit{TL}=\left[{TL}_{ij}\right] =\underset{m\to \infty }{\lim }\left(L+L^2+L^3\dots {+L}^c \right)=L(1-{L)}^{-1}\mathit{TM} \\ =\left[{TM}_{ij}\right] =\underset{m\to \infty }{\lim }\left(M+M^2+M^3\dots {+M}^c \right)=M(1-{M)}^{-1} \end{gathered}$$

(5)

$$\mathit{TU}=\left[{TU}_{ij}\right] =\underset{m\to \infty }{\lim }\left(U+U^2+U^3\dots {+U}^c \right)=U(1-{U)}^{-1}$$

(6)

$$T=\left[\begin{array}{ccc}\begin{array}{ccc}{TL}_{11},& {TM}_{11},& {TU}_{11}\\ {TL}_{21},& {TM}_{21},& {TU}_{21}\end{array}& \cdots & \begin{array}{ccc}{TL}_{1n},& {TM}_{1n},& {TU}_{1n}\\ {TL}_{2n},& {TM}_{2n},& {TU}_{2n}\end{array}\\ ⋮& \ddots & ⋮\\ \begin{array}{ccc}{TL}_{n1},& {TM}_{n2},& {TU}_{n3}\end{array}& \cdots & \begin{array}{ccc}{TL}_{nn},& {TM}_{nn},& {TU}_{nn}\end{array}\end{array}\right]$$

(7)

3.1.6. Step 6: Calculate the Influence Degree, Affected Degree, Center Degree, and Cause Degree of Each Factor

The degree of influence indicated as Di, measures the extent to which different factors have a cumulative effect on other factors. Thus,

$${D_{i=\sum _{j=1}^{n}t}}_{ij}$$

(8)

The affected degree Rj indicates the extent to which the other factors influence each factor. Thus,

$${R_{j=\sum _{i=1}^{n}t}}_{ij}$$

(9)

The center degree,$Rj-Di$, indicates the importance of factors.

$$\mathrm{Therefore}, \mathrm{the} \mathrm{center} \mathrm{degree}=\left\{\left(Rj+Di\right)\left|i=j\right.\right\}.$$

(10)

For the cause degree, when $Rj+Di$ is positive, the factor belongs to the cause group, and when $Rj-Di$ is negative, the factor belongs to the effect group.

$$\mathrm{The} \mathrm{cause} \mathrm{degree}=\left\{\left(Rj-Di\right)\left|i=j\right.\right\}$$

(11)

3.1.7. Step 7: Establish and Analyze the F-DEMATEL Diagram

In the overall relation matrix $T$, the sums of the rows and columns are represented separately by the vectors Rj and Di. By plotting the dataset of $(Rj+Di, Rj-Di$), it is possible to create a cause-and-effect graph based on the values of $Rj+Di$ and $Rj-Di.$ The horizontal axis vector ($Rj+Di$), referred to as “Prominence”, is formed by combining $Rj$ and $Di$, signifying the importance of the criterion. In the same way, the vertical axis ($Rj-Di)$, designated as “Relation”, is formed by subtracting $Di$ from $Rj$, allowing for the classification of criteria into a cause category.

3.2. Structural Equation Modeling (SEM)

SEM is a data analysis technique employing regression analysis to forecast results using several variables. It integrates factor and path analysis to validate empirical studies and explore causal relationships among variables [68]. A theoretical framework is created to construct Structural Equation Modeling (SEM), with parameters being recognized and measured through path diagrams. The model’s goodness of fit is evaluated, and modifications are implemented [69].

3.2.1. Research Participants and Procedure

The study participants were the general public in Taiwan involved in environmental protection and had access to AI technological products. A quantitative approach was employed, and data were collected through a questionnaire survey utilizing a convenience sampling technique. This method is popular due to its cost-effectiveness, simplicity, and relatively quick execution [70]. Convenience sampling involves gathering data from readily available participants. Still, it also has limitations, such as potential selection bias and a lack of representativeness of the larger population due to its reliance on accessibility and motivation [71]. The questionnaires were created using Google Forms, an online survey tool, and distributed to participants through electronic channels such as social networking sites, email, Instagram, and Messenger. A back-translation method was used to ensure that the questionnaire items were culturally appropriate for the local context. Respondents participated voluntarily after being informed about the research purpose.

The valid dataset for analysis comprised 1726 respondents. Previous studies have stated that the number of samples should be more than 5 times of the questionnaire items [68,72]. In our study, we identified a total of 40 items in the SEM questionnaire; thus, at least 200 samples were required. Therefore, a total of 1726 samples indicate that the analysis results already had high reference values. Among the participants, 551 were male (31.9%) and 1175 were female (68.1%). Participants were asked to specify their age range, with the 18–25 age group being the most frequently selected; this group included 1125 respondents (65.2%), indicating a strong engagement among young people with new AI technology. Conversely, the 26–40 age group had the fewest participants, totaling only 144 individuals (8.3%). Participants were categorized based on their highest level of education attained. The largest group comprised individuals with bachelor’s degrees, accounting for 65.2% of the total (1125 individuals). The smallest group included participants with postgraduate degrees, totaling 601 individuals. Additionally, most participants were students, numbering 1125 (65.2%), while the smallest group consisted of employed individuals, totaling 221 (12.8%).

3.2.2. Measurement Items

The study utilized a structured multiple-item measurement approach to evaluate various constructs, including perceived trust, self-efficacy, perceived awareness, perceived benefit, perceived usefulness, perceived ease of use, social influence, facilitating conditions, behavioral intention to use, and pro-environmental behavior. All constructs were measured using 7-point rating scales, ranging from 1 (strongly disagree) to 7 (strongly agree). The wording of specific items was adapted to fit the research context.

Perceived trust: This construct was assessed using four items, such as “AI devices have features necessary to complete key tasks” and “AI devices are dependable [25]”. Self-efficacy: Three items were used, including “It would be easy for me to use AI devices by myself [20]”. Perceived benefit: Four items were employed, including “Using the AI device would provide me with a lot of enjoyment [68]”. Perceived usefulness: This was assessed using four items, such as “I find AI devices useful in my daily life [73]”. Social influence: Four items were used to measure this construct, including “Peers who are important to me think that I should use AI devices [14]”. Perceived ease of use: Items were adopted from [13,73], including “Learning how to use AI devices is easy for me”. Facilitating conditions: This construct was measured with three items such as “We have the resources necessary to use AI devices [47]”. Behavioral intention to use: Four items were used to measure behavioral intention to use, including “Given that I have access to the AI devices, I predict that I would use them [28,47]”. Pro-environmental behavior: Since it is challenging to measure personal PEB, most previous studies have used self-reporting to measure personal ERB [16,74]. Therefore, in the present study, PEB was assessed with six items previously used by Lucarell et al. [6], such as “I have changed/I will change to a more fuel-efficient car”.

3.2.3. Data Analytical Tools

The present study employed Variance-Based Structural Equation Modeling (VBSEM) to examine the structural relationships between nine variables and pro-environmental behavior. Specifically, Partial Least Squares-Structural Equation Modeling (PLS-SEM) and the Statistical Package for the Social Sciences (SPSS) software were used to obtain precise estimates, with sample sizes exceeding 50 [75]. PLS-SEM is particularly suited for complex structural models involving multiple constructs, indicators, and relationships, making it an ideal choice for this research, which includes 10 constructs and 40 items with various model relationships. Additionally, PLS-SEM is advantageous when dealing with non-normally distributed data, as it does not require specific distribution assumptions. This flexibility makes it a robust tool for testing intricate models and allows for better handling of data characteristics that may not conform to normal distribution patterns.

Furthermore, PLS-SEM offers an alternative when Covariance-Based Structural Equation Modeling (CB-SEM) produces inadequate results [76]. It is more effective at detecting significant relationships within a population, thus providing greater statistical power compared to other techniques [77]. To ensure the reliability of the measurement model, a reliability test was first conducted using Cronbach’s alpha. Following this, the measurement’s validity was assessed through both convergent and discriminant validity analyses. Confirmatory Factor Analysis (CFA) was subsequently employed to evaluate the model fit indices of the research framework, ensuring the appropriateness of the model. Finally, path analysis was conducted to test the proposed hypotheses and evaluate the path coefficients among the variables integrated into the structural model. This comprehensive approach allowed for a thorough investigation of the relationships between the variables and their impact on pro-environmental behavior.

3.3. Artificial Neural Network (ANN) Approach

ANN can serve as a complementary method to Structural Equation Modeling (SEM) [60]. ANN is an analytical approach that utilizes deep learning algorithms, structured around input, output, and hidden layers, without allowing for compensation [58]. This technique is highly adaptable and does not rely on the multivariate assumptions or hypotheses typically required by other methods [59]. The present research suggests that integrating F-DEMATEL, SEM, and ANN provides a more effective research approach. ANN enhances prediction accuracy by using artificial neurons in its hidden layers to link input and output data. While SEM is effective in analyzing causal relationships, ANN excels in identifying both linear and nonlinear relationships through artificial intelligence techniques [60]. SEM, on the other hand, can analyze complex relationships between independent and dependent variables, incorporating both factors and measured variables. It serves as a confirmatory method to evaluate and test hypotheses, refine existing models, or explore interrelated models [78].

4. Results

4.1. F-DEMATEL Results

In the first stage, opinions were gathered from a group of twenty experts in the environmental field. Following this, a fuzzy linguistic scale was developed and established in the second phase. The third step entails calculating the initial direct-relation fuzzy matrix. This matrix is formed by having evaluators establish fuzzy pair-wise influence relationships among the components in an n x n format, with k representing the number of experts, as illustrated in

Table 2. Direct-relation fuzzy matrix.

	T = L(I−L)1								T = M(I−M)1								T = U(I−U)1
	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC
PT	0.000	0.438	0.463	0.525	0.575	0.563	0.475	0.463	0.000	0.675	0.700	0.763	0.825	0.813	0.725	0.713	0.000	0.900	0.925	0.925	1.000	0.988	0.925	0.925
SE	0.375	0.000	0.400	0.488	0.538	0.538	0.488	0.488	0.613	0.000	0.638	0.725	0.788	0.788	0.738	0.738	0.850	0.000	0.875	0.913	0.975	0.975	0.950	0.950
PA	0.388	0.388	0.000	0.525	0.538	0.538	0.450	0.463	0.625	0.625	0.000	0.775	0.788	0.788	0.700	0.713	0.875	0.875	0.000	0.963	0.988	1.000	0.938	0.947
PB	0.313	0.325	0.325	0.000	0.363	0.375	0.375	0.388	0.563	0.575	0.575	0.000	0.613	0.625	0.625	0.638	0.813	0.825	0.825	0.000	0.863	0.875	0.875	0.875
PU	0.238	0.250	0.288	0.313	0.000	0.350	0.350	0.363	0.475	0.500	0.538	0.563	0.000	0.600	0.600	0.613	0.725	0.750	0.788	0.813	0.000	0.850	0.850	0.842
PEOU	0.275	0.263	0.275	0.313	0.338	0.000	0.375	0.375	0.500	0.513	0.525	0.563	0.588	0.000	0.625	0.625	0.750	0.763	0.775	0.813	0.838	0.000	0.875	0.875
SI	0.338	0.388	0.388	0.413	0.425	0.450	0.000	0.461	0.613	0.638	0.638	0.663	0.675	0.700	0.000	0.711	0.863	0.875	0.875	0.875	0.888	0.913	0.000	0.934
FC	0.375	0.388	0.388	0.463	0.488	0.475	0.475	0.000	0.625	0.638	0.638	0.713	0.738	0.725	0.725	0.000	0.888	0.888	0.888	0.938	0.950	0.938	0.938	0.000

Source: own calculations. After calculating the average direct-relation fuzzy matrix, the normalized direct-relation fuzzy matrix was determined by using a linear scale transformation for criteria comparison, resulting in the normalized direct-relation fuzzy matrix known as cap X. This is calculated using Equations (2)–(4) outlined in the methodology section. The normalized direct-relation fuzzy matrix is presented in Table 3 below.

below.

After calculating the average direct-relation fuzzy matrix, the normalized direct-relation fuzzy matrix was determined by using a linear scale transformation for criteria comparison, resulting in the normalized direct-relation fuzzy matrix known as cap X. This is calculated using Equations (2)–(4) outlined in the methodology section. The normalized direct-relation fuzzy matrix is presented in

Table 3. Normalized initial direct matrix (D).

	T = L(I−L)1								T = M(I−M)1								T = U(I−U)1
	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC
PT	0.000	0.125	0.132	0.150	0.164	0.161	0.136	0.132	0.000	0.129	0.134	0.146	0.158	0.156	0.139	0.137	0.000	0.137	0.140	0.140	0.152	0.150	0.140	0.140
SE	0.107	0.000	0.114	0.139	0.154	0.154	0.139	0.139	0.118	0.000	0.122	0.139	0.151	0.151	0.141	0.141	0.129	0.000	0.133	0.139	0.148	0.148	0.144	0.144
PA	0.111	0.111	0.000	0.150	0.154	0.154	0.129	0.132	0.120	0.120	0.000	0.149	0.151	0.151	0.134	0.137	0.133	0.133	0.000	0.146	0.150	0.152	0.142	0.144
PB	0.089	0.093	0.093	0.000	0.104	0.107	0.107	0.111	0.108	0.110	0.110	0.000	0.118	0.120	0.120	0.122	0.123	0.125	0.125	0.000	0.131	0.133	0.133	0.133
PU	0.068	0.071	0.082	0.089	0.000	0.100	0.100	0.104	0.091	0.096	0.103	0.108	0.000	0.115	0.115	0.118	0.110	0.114	0.120	0.123	0.000	0.129	0.129	0.128
PEOU	0.079	0.075	0.079	0.089	0.096	0.000	0.107	0.107	0.096	0.098	0.101	0.108	0.113	0.000	0.120	0.120	0.114	0.116	0.118	0.123	0.127	0.000	0.133	0.133
SI	0.096	0.111	0.111	0.118	0.121	0.129	0.000	0.132	0.118	0.122	0.122	0.127	0.129	0.134	0.000	0.136	0.131	0.133	0.133	0.133	0.135	0.139	0.000	0.142
FC	0.107	0.111	0.111	0.132	0.139	0.136	0.136	0.000	0.120	0.122	0.122	0.137	0.141	0.139	0.139	0.000	0.135	0.135	0.135	0.142	0.144	0.142	0.142	0.000

Source: own calculations. The structural model was subsequently created and examined by obtaining the total-relation fuzzy matrix $T$, following the normalization of the direct-relation fuzzy matrix. In this matrix, the element $tij$ represents the indirect influence between factors $i$ and $j$. The influence matrix T represents the comprehensive impact relationship among the elements. The total-relation fuzzy matrix ($T$) was determined using the formula T =$\underset{m\to \infty }{\lim }\left(X+X^2+X^3\dots {+X}^m \right)$ = X* (1$-{X)}^{-1}$, which incorporates an identity matrix. The findings from the total-relation fuzzy matrix ($T$) are displayed in Table 4.

below.

The structural model was subsequently created and examined by obtaining the total-relation fuzzy matrix $T$, following the normalization of the direct-relation fuzzy matrix. In this matrix, the element $tij$ represents the indirect influence between factors $i$ and $j$. The influence matrix T represented the comprehensive impact relationship among the elements. The total-relation fuzzy matrix ($T$) was determined using the formula T =$\underset{m\to \infty }{\lim }\left(X+X^2+X^3\dots {+X}^m \right)$ = X* (1$-{X)}^{-1}$, which incorporates an identity matrix. The findings from the total-relation fuzzy matrix ($T$) are presented in

Table 4. The total relation fuzzy matrix.

	T = L(I−L)1								T = M(I−M)1								T = U(I−U)1
	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC	PT	SE	PA	PB	PU	PEOU	SI	FC
PT	0.418	0.550	0.570	0.664	0.712	0.714	0.649	0.648	0.776	0.916	0.935	1.029	1.081	1.084	1.022	1.022	1.794	1.945	1.968	2.047	2.126	2.135	2.079	2.078
SE	0.494	0.417	0.533	0.629	0.675	0.679	0.626	0.628	0.856	0.775	0.899	0.994	1.044	1.049	0.995	0.997	1.884	1.800	1.937	2.019	2.096	2.107	2.055	2.054
PA	0.493	0.513	0.427	0.633	0.670	0.675	0.613	0.618	0.855	0.880	0.787	0.999	1.041	1.046	0.986	0.990	1.910	1.940	1.844	2.049	2.123	2.135	2.078	2.079
PB	0.389	0.408	0.418	0.392	0.512	0.519	0.486	0.490	0.742	0.765	0.777	0.749	0.889	0.895	0.854	0.858	1.751	1.780	1.799	1.760	1.940	1.951	1.906	1.905
PU	0.334	0.351	0.369	0.426	0.367	0.462	0.433	0.437	0.682	0.706	0.723	0.793	0.727	0.834	0.797	0.801	1.658	1.687	1.710	1.781	1.732	1.856	1.813	1.811
PEOU	0.351	0.363	0.375	0.437	0.467	0.382	0.449	0.450	0.694	0.715	0.729	0.801	0.838	0.740	0.810	0.811	1.679	1.707	1.727	1.800	1.865	1.762	1.835	1.835
SI	0.439	0.468	0.480	0.553	0.586	0.596	0.445	0.563	0.807	0.833	0.847	0.928	0.968	0.976	0.814	0.935	1.824	1.855	1.874	1.949	2.018	2.031	1.862	1.985
FC	0.466	0.488	0.501	0.589	0.626	0.628	0.588	0.470	0.829	0.854	0.868	0.959	1.002	1.004	0.959	0.839	1.874	1.904	1.924	2.007	2.078	2.086	2.038	1.913

Source: own calculations. The subsequent step consisted of calculating the influence degree, affected degree, center degree, and cause degree for each factor using Equations (9)–(12) outlined in the methodology section. The sums of the rows and columns were calculated to determine the total significance and overall impact. Table 5 and Table 6 presented below illustrate the sums of the rows ($Rj)$ and columns ($Di$) of the overall fuzzy relation matrix ($T$), along with the prominence, net effects, and categorization of the criteria into causes and effects.

.

The subsequent step consisted of calculating the influence degree, affected degree, center degree, and cause degree for each factor using Equations (9)–(12) outlined in the methodology section. The sums of the rows and columns were calculated to determine the total significance and overall impact.

Table 5. Sums of rows and columns.

	$\mathit{R}\mathit{i}$	$\mathit{D}\mathit{i}$	$\mathit{R}\mathit{i}$	$\mathit{D}\mathit{i}$	$\mathit{R}\mathit{i}$	$\mathit{D}\mathit{i}$
PT	4.925	3.385	7.864	6.240	16.17	14.37
SE	4.681	3.558	7.608	6.444	15.95	14.62
PA	4.642	3.673	7.585	6.565	16.16	14.78
PB	3.614	4.324	6.526	7.252	14.79	15.41
PU	3.178	4.615	6.063	7.590	14.05	15.98
PEOU	3.273	4.654	6.139	7.628	14.21	16.06
SI	4.130	4.288	7.108	7.237	15.40	15.67
FC	4.356	4.304	7.314	7.253	15.82	15.66

Source: own calculations.

and

Table 6. The prominence, net effects, and classification of the criteria into causes and effects.

	$\mathit{R}\mathit{j}+\mathit{D}\mathit{i}$	$\mathit{R}\mathit{j}-\mathit{D}\mathit{i}$	Identity
PT	17.6530	1.6542	Cause
SE	17.6201	1.2068	Cause
PA	17.8021	1.1214	Cause
PB	17.3074	−0.6842	Effect
PU	17.1573	−1.6311	Effect
PEOU	17.3221	−1.5743	Effect
SI	17.9418	−0.1851	Effect
FC	18.2378	0.0925	Cause

Source: own calculations.

presented below illustrate the sums of the rows ($Rj)$ and columns ($Di$) of the overall fuzzy relation matrix ($T$), along with the prominence, net effects, and categorization of the criteria into causes and effects.

The values of $Rj-Di$ in the F-DEMATEL analysis reflect the overall influence among criteria, classifying factors as causal or effect. A negative $Rj-Di$ i value implies that a factor is an effect factor, indicating its susceptibility to influence by other factors. Conversely, a positive $Rj-Di$ value designates the factor as a causal one, exerting influence over other factors. In this study, the F-DEMATEL results identify perceived trust (PT), self-efficacy (SE), perceived awareness (PA), and facilitating condition (FC) as causal factors, meaning they impact the system significantly. Meanwhile, perceived benefit (PB), perceived usefulness (PU), perceived ease of use (PEOU), and social influence (SI) emerge as effect factors, with negative $Rj-Di$ values, indicating they are more influenced by other variables.

, following the normalization of the direct-relation fuzzy matrix. In this matrix, the element $tij$ represents the indirect influence between factors $i$ and $j$. The influence matrix T represents the comprehensive impact relationship among the elements. The total-relation fuzzy matrix ($T$) was determined using the formula T =$\underset{m\to \infty }{\lim }\left(X+X^2+X^3\dots {+X}^m \right)$ = X* (1$-{X)}^{-1}$, which incorporates an identity matrix. The findings from the total-relation fuzzy matrix ($T$) are displayed in Table 4.

The sum of $Rj$ and $Di$ provides insight into the overall prominence of each factor in the cause-and-effect chain. Table 6 shows that FC, with the $Rj+Di$ value of 18.2378, has the highest prominence among causal factors, underscoring its essential role in shaping pro-environmental behavior (PEB) through intentions to use environmental AI products. FC, PA, PT, and SE have prominence values of 17.8021, 17.6530, and 17.6201, respectively, signifying their considerable influence on behavior. The results affirm that external factors such as PT, SE, PA, and FC exert notable effects on the primary variables within the TAM and UTAUT frameworks, influencing individuals’ intention to use environmental AI products and ultimately fostering pro-environmental behavior.

In the final analytical stage, a cause-and-effect diagram (Figure 2) is created, displaying the relationships between factors. The horizontal axis, labeled “Prominence”, represents $Rj+Di$, the sum of each factor’s total influence, indicating the factor’s importance. The vertical axis, labeled “Relation”, represents $Rj-Di$, differentiating between causal and effect factors by showing the net influence a factor has on or receives from others. This visual representation in Figure 2 clarifies the roles of each criterion within the system, highlighting causal pathways and relative impacts among the factors impacting PEB.

4.2. SEM Results

4.2.1. Multicollinearity Assessment

Tolerance and Variance Inflation Factor (VIF) were conducted to check for multicollinearity issues. Multicollinearity is a concern if a VIF value exceeds 10 [77]. The results of the multicollinearity analysis in the current study reveal that all study constructs have VIF values of less than 10. Thus, the study suffered from no multi-collinearity issues as the highest VIF value is 2.741.

4.2.2. Common Method Bias (CMB)

Harman’s single-factor test was performed to assess the presence of common method bias. If the variance extracted for the initial factor is less than 50%, it can be inferred that common method bias does not exist in the study [79]. In the current research, 40 factors were identified, with the first factor accounting for 24.353% of the explained variance. This suggests that there were not any significant issues with CMB, as the explained variance of the first factor in the dataset of the present study was under 50%.

4.2.3. Measurement Model Assessment

Composite Reliability, Discriminant, and Convergent Validity

To achieve satisfactory reliability, Cronbach’s alpha should meet or exceed a minimum threshold of 0.7, as recommended [77]. A Cronbach’s alpha value at or above this level suggests that the constructs are internally consistent and reliable for analysis. Each of the ten variables in this study meets this criterion, indicating acceptable reliability and consistency across the measurement items. Specifically, the Cronbach’s alpha values for each construct are as follows: perceived ease of use (PEOU) = 0.807, facilitating condition (FC) = 0.732, intention to use (IU) = 0.714, perceived awareness (PA) = 0.747, perceived benefit (PB) = 0.761, pro-environmental behavior (PEB) = 0.727, perceived trust (PT) = 0.712, perceived usefulness (PU) = 0.700, self-efficacy (SE) = 0.710, and social influence (SI) = 0.707. These values suggest that the items used to measure each variable are coherent and contribute to the overall reliability of the model, providing a robust foundation for further analysis. Figure 3 illustrates the PLS-SEM path analysis for the inner model, estimating the relationships between latent constructs. This path analysis helps interpret the structural relationships among the variables, is supported by the established reliability, and further enhances the model’s validity for understanding the interactions within the constructs.

Additionally, constructs are guaranteed to be reliable if their composite reliability (CR) value is higher than 0.7 [77]. The CRs vary from 0.804 to 0.873, as shown in

Table 7. Means, standard deviations, and correlations.

Variables	Mean	SD	CR	AVE	1	2	3	4	5	6	7	8	9	10
1. PT	4.482	0.876	0.837	0.633	0.795
2. SE	4.704	0.954	0.838	0.635	0.523 **	0.970
3. PA	4.854	0.758	0.839	0.572	0.435 **	0.625 **	0.757
4. PB	4.920	0.905	0.840	0.640	0.187 **	0.108 **	−0.149 **	0.800
5. PU	5.632	0.621	0.816	0.528	0.347 **	0.520 **	0.432 **	−0.128 **	0.726
6. SI	5.710	0.585	0.819	0.531	0.186 **	0.278 **	0.387 **	−0.074 **	0.143 **	0.728
7. PEOU	4.284	0.875	0.873	0.633	0.436 **	0.606 **	0.692 **	−0.218 **	0.425 **	0.500 **	0.796
8. FC	5.051	0.842	0.846	0.649	0.399 **	0.521 **	0.484 **	−0.237 **	0.632 **	0.127 **	0.478 **	0.806
9. IU	5.715	0.571	0.823	0.538	0.355 **	0.581 **	0.443 **	−0.052 *	0.591 **	0.180 **	0.473 **	0.526 **	0.734
10. PEB	5.681	0.503	0.804	0.412	0.333 **	0.337 **	0.419 **	0.011	0.233 **	0.214 **	0.517 **	0.186 **	0.223 **	0.642

Source: own calculations. Note: N = 1726; * p < 0.05; ** p < 0.01. Reliability coefficients are reported along the diagonal. PT = perceived trust; SE = self-efficacy; PA = perceived awareness; FC = facilitating condition; PB= perceived benefit; PU = perceived usefulness; PEOU = perceived ease of use; IU= behavioral intentions to use; SI = social influence; PEB = pro-environmental behavior.

below, demonstrating an exceptionally high-quality standard. AVE measures the proportion of a construct’s variance that can be ascribed to the construct rather than measurement errors. It is also stated that when evaluating the model’s constructs for convergent validity, values greater than 0.5 are deemed acceptable [77]. On the other hand, AVE values greater than 0.4 can be considered appropriate in certain fields of research [80]. Test results are shown in Table 7, which shows that most AVE values are higher than the 0.5 cutoff. Conversely, the AVE for PEB shows a slight decrease, settling at 0.412 and dropping below 0.5. Sometimes, AVE may be a conservative estimation of the measurement model’s validity, implying that researchers may find the convergent validity of the construct satisfactory based solely on composite reliability [81], even if more than 50% of the variance is attributable to error. However, some AVEs fall below the recommended level of 0.5, ranging from 40% to 50%. The composite reliability of all constructs is significantly greater than the suggested level, even though the internal reliability of the measurement items is deemed sufficient. Composite reliability numbers, which satisfy the minimum threshold, validate the reliability.

Similarly, discriminant validity is established because the diagonal values exceed all other values in their respective rows and columns. Bivariate correlations reveal small to moderate relationships between the independent and dependent variables, with correlation coefficients ranging from 0.011 to 0.692. In a nutshell, the findings from the reliability and validity tests conducted using Smart-PLS 3 demonstrate that the measurement items for the constructs were both reliable and valid. However, some items exhibited lower factor loadings (PT1, PB4, PEB 1, and PEB3), leading to their removal for falling below the minimum threshold.

Table 8. Descriptive statistics and factor loadings of the items (acceptable threshold values in brackets).

Construct Items	Mean (SD)	Factor Loading (>0.60)	Cronbach’s a (>0.70)	CR (>0.70)	AVE (>0.50)
Perceived trust in technology (PT)	4.482 (0.876)		0.712	0.837	0.633
PT2_ AI products have features necessary to complete key tasks PT3-_AI products are reliable PT4 _AI products are dependable		0.737 0.795 0.850
Self-efficacy (SE)	4.704 (0.954)		0.710	0.838	0.635
SE1_It would be easy for me to use AI products by myself. SE2_I could adopt AI products even if there is no one around to tell me what to do as I go. SE3_I could adopt AI products if I could contact someone when I got stuck.		0.864 0.839 0.674
Perceived Awareness (PA)	4.854 (0.758)		0.747	0.839	0.572
PA1_In general, I am aware of AI products. PA2_I have received enough information about the benefits of using AI products. PA3_I have gone through training programs about the overall features of AI products. PA4_I have known through social media about the overall features AI products.		0.644 0.613 0.855 0.876
Perceived benefit (PB)	4.920 (0.905)		0.761	0.840	0.640
PB1_I would have fun interacting with the AI product. PB2_Using the AI product would provide me with a lot of enjoyment. PB3-I would enjoy using the AI product.		0.918 0.658 0.803
Perceived usefulness (PU)	5.632 (0.621)		0.700	0.816	0.528
PU1_I find AI product useful in my daily life. PU2_Using AI product increases my chances of achieving tasks that are important to me. PU3_Using AI product helps me accomplish my tasks more quickly. PU4_Using AI product increases my productivity.		0.766 0.785 0.646 0.700
Social influence (SI)	5.710 (0.585)		0.707	0.819	0.531
SI1_Peers who are important to me would think that I should use AI product. SI2_Peers who influence my behavior would think that I should use product. SI3_My superiors who influence my behavior would think that I should use AI product. SI4_My superiors to whom I report would think that I should use AI product.		0.738 0.726 0.761 0.687
Perceived ease of use (PEOU)	4.284 (0.875)		0.807	0.873	0.633
EU1_Learning how to use AI products is easy for me. EU2_My interaction with AI products is clear and understandable. EU3_I find AI products easy to use. EU4_It is easy for me to become skilful at using AI products.		0.774 0.809 0.764 0.834
Facilitating conditions (FC)	5.051 (0.842)		0.732	0.846	0.649
FC1_We have the resources necessary to use AI products. FC2_AI products are compatible with the other technologies that I use. FC3_We can get help from others when we have difficulties using AI products.		0.867 0.686 0.852
Intention to use AI products (IU)	5.715 (0.571)		0.714	0.823	0.538
IU1_Assuming I have access to the AI products, I intend to use them. IU2_Given that I have access to the AI products, I predict that I would use them. IU3_If I have access to the AI products, I use them as much as possible. IU4_It is my responsibility to encourage my colleagues to adopt ways to counter climate change.		0.760 0.748 0.700 0.725
Pro-environmental behavior (PEB)	5.681 (0.503)		0.727	0.804	0.412
PEB2_I use public transport. PEB4_I have reduced the amount of waste I used to produce. PEB5_I have installed low energy light bulbs. PEB6-I turn off lights/fans/electrical appliances when not in use.		0.732 0.670 0.690 0.694

presents the factor loadings of the study constructs.

Confirmatory Factor Analysis (CFA)

This research performed Confirmatory Factor Analysis (CFA) using AMOS 21.0 to evaluate the model’s goodness of fit. The ten-factor model showed a good fit based on the CFA results, indicated by a χ2 value of 2494.455 and 620 degrees of freedom, with a significance level below 0.001. In addition, the CFI of 0.918, IFI of 0.919, TLI (rho2) of 0.908, and RMSEA of 0.042 suggest a satisfactory and acceptable level of model fit. Therefore, in our model, an RMSEA value of 0.040 signifies an ideal match between the model and the data.

4.2.4. Hypotheses Testing

The research utilized Structural Equation Modeling to investigate the hypotheses regarding the relationships between constructs, employing Smart-PLS 3.5. The test results show the expected significance and importance of the proposed relationships between variable groups that either support or disprove the hypotheses. The path coefficients are assessed on a scale that spans from negative to positive values. Values nearing +1 suggest a stronger positive correlation, whereas values approaching −1 indicate a stronger negative correlation. The results in

Table 9. Results of examination of direct research hypotheses.

Variables	Path Coefficients	p Values	Remarks
EU→IU	0.125	***	Supported
FC→IU	0.133	***	Supported
IU→PEB	0.270	***	Supported
PA→PEOU	0.503	***	Supported
PA→IU	−0.010	0.690	Not Supported
PA→PU	0.170	***	Supported
PB→IU	0.017	0.413	Not Supported
PT→PEOU	0.088	***	Supported
PT→IU	−0.009	0.685	Not Supported
PT→PU	0.078	***	Supported
PU→IU	0.316	***	Supported
SE→PEOU	0.244	***	Supported
SE→IU	0.282	***	Supported
SE→PU	0.372	***	Supported
SI→IU	−0.009	0.632	Not Supported

Source: own calculations. Note: *** p < 0.001. PT = perceived trust; SE = self-efficacy; PA = perceived awareness; FC = facilitating condition; PB = perceived benefit; PU = perceived usefulness; PEOU = perceived ease of use; IU = behavioral intentions to use; SI = social influence; PEB = pro-environmental behavior.

below demonstrate the results of path analysis.

Regarding the results presented above, PT, as an external factor, was found to have a positive relationship with both PU and PEOU (β = 0.078, p-value < 0.001 and β = 0.088, p-value < 0.001), respectively, thereby supporting hypotheses 1a and 1b, which proposed that PT influences the TAM main variables (PU and PEOU). However, the direct relationship between PT and IU was not supported (β = −0.009, p-value = 0.685). Therefore, hypothesis 1c which proposed a significant positive relationship between PT and IU was not supported. There was a significant positive relationship between SE and the two TAM main factors, PU and PEOU evidenced by the following results (β = 0.372, p-value < 0.001 and β = 0.244, p-value < 0.001), respectively. In this regard, hypotheses 2a and 2b were supported accordingly. Unlike hypothesis 1c, hypothesis 2c, which proposed the direct positive relationship between SE and IU, was supported (β = 0. 0.282, p-value < 0.001), implying that individuals with strong self-efficacy will develop greater intentions to use AI products as they will feel confident in attempting to learn the new technology. As another external factor that influenced the TAM, PA was positively related to both PU and PEOU (β = 0.170, p-value < 0.001 and β = 0.503, p-value < 0.001), respectively, supporting hypotheses 3a and 3b. On the other hand, hypothesis 3c, which proposed the direct relationship between PA and IU, has not been supported as the results demonstrated a non-significant impact (β = −0.010, p-value = 0.690). Similarly, hypothesis 4, which proposed a significant positive relationship between PB and IU, was not supported as the study results revealed a non-significant p-value of 0.413 and a coefficient of 0.017, implying that PB does not affect IU.

The study further established a positive relationship between PU and IU, as evidenced by a p-value < 0.001 and a coefficient of 0.316, thereby supporting hypothesis 5. Likewise, there is a significant positive relationship between PEOU and IU (β = 0.125, p-value ˂ 0.001) thus supporting hypothesis 6. As for hypothesis 7, which proposed a positive relationship between SI and IU, the study found contrary results as it revealed a non-significant p-value of 0.632 and a negative coefficient of -0.009, implying that SI does not influence IU in the present study. Hence, hypothesis 7 is not supported. As hypothesized, FC is significantly and positively related to IU (β = 0.133, p-value ˂ 0.001), thereby supporting hypothesis 8. Furthermore, the present study has demonstrated a significant and positive relationship between IU and PEB with a p-value of less than 0.001 and a coefficient of 0.270. The result supports hypothesis 9.

Mediation Analysis

A bootstrap analysis was conducted in SPSS using PROCESS macro 4.1 (Model 4) for the exploration of mediation effects of two main factors of TAM in the relationship between external factors and behavioral intention to use [82]. The exclusion of zero within the 95% confidence interval for estimates is indicative of the mediation effect’s significance [83]. As illustrated by the proceeding

Table 10. Bootstrapping results for indirect effects.

Rival Path	Estimated Effect	Bootstrapping 95% CI
		Bias-Corrected (LL, UL)
PT → PU → IU	0.1035	(0.0900, 0.1182)
PT → PEOU → IU	0.0677	(0.0537, 0.0829)
SE → PU → IU	0.1173	(0.1022, 0.1328)
SE → PEOU → IU	0.0455	(0.0289, 0.0631)
PA → PU → IU	0.1494	(0.1327, 0.1667)
PA → PEOU → IU	0.1087	(0.0793, 0.1373)

Source: own calculations. Note: PA = perceived trust; SE = self-efficacy; PA = perceived awareness; PU = perceived usefulness; PEOU = perceived ease of use; IU = intention to use.

, all rival paths encompass a 95% confidence interval without an estimate of zero. This outcome distinctly affirms the existence of significant mediation effects supporting hypotheses 1d, 1e, 2d, 2e, 3d, and 3e.

4.3. Artificial Neural Network (ANN) Results

In this study, three Artificial Neural Network (ANN) models were employed to assess the data, with SPSS 22 used for data analysis and interpretation. The models were designed to analyze various factors related to AI products that influence the intention to use and accept environmentally related AI products, as well as their impact on pro-environmental behavior. Model A evaluates perceived usefulness, incorporating three external factors: perceived trust, self-efficacy, and perceived awareness. Model B assesses perceived ease of use, using the same three external factors as inputs. Model C evaluates behavioral intention to use AI products as output, considering four significant factors: perceived usefulness, self-efficacy, perceived ease of use, and facilitating conditions as inputs.

The Root Mean Square Error (RMSE) measure, based on the methodology of Yadav and Pathak [84], was used to evaluate the accuracy of the various ANN models. Each model consists of a single hidden layer, with SPSS automatically determining the number of neurons in that layer [60]. For example, Figure 4 presents an ANN model for perceived usefulness (Model A). Using 70% of the data for network training and 30% for testing, the average RMSE values for the three ANN models are shown in

Table 11. RMSE values of Artificial Neural Networks (N = 1726).

	MODEL A		MODEL B		MODEL C
Network	RMSE (Training)	RMSE (Testing)	RMSE (Training)	RMSE (Testing)	RMSE (Training)	RMSE (Testing)
ANN1	0.624	0.678	0.391	0.391	0.452	0.451
ANN2	0.646	0.635	0.408	0.377	0.464	0.437
ANN3	0.642	0.628	0.387	0.392	0.454	0.458
ANN4	0.659	0.630	0.407	0.374	0.470	0.426
ANN5	0.634	0.653	0.402	0.343	0.466	0.451
ANN6	0.641	0.640	0.385	0.379	0.461	0.456
ANN7	0.640	0.631	0.398	0.365	0.453	0.457
ANN8	0.647	0.621	0.410	0.334	0.447	0.455
ANN9	0.635	0.661	0.412	0.402	0.473	0.409
ANN10	0.642	0.650	0.386	0.399	0.456	0.455

Source: own calculations.

below. These values range from 0.385 to 0.659 for the training sets and 0.334 to 0.678 for the testing sets. The findings indicate that the modeling of the relationships between independent and dependent variables demonstrates a moderate level of accuracy and reliability.

Furthermore, the outcomes derived from the Partial Least Squares Structural Equation Modeling (PLS-SEM) were compared to those from the ANN models to verify consistency and accuracy. As illustrated in

Table 12. Comparison between PLS-SEM and ANN analysis.

PLS-Path	Path Coefficient	Normalized Relative Importance	Ranking Based on PLS-SEM	Ranking Based on ANN	Remark
Model A
PT→PU	0.078 (<0.001)	30.2%	3	3	Match
SE→PU	0.372 (<0.001)	100.0%	1	1	Match
PA→PU	0.170 (<0.001)	59.1%	2	2	Match
Model B
PT→PEOU	0.088 (<0.001)	25.2%	3	3	Match
SE→PEOU	0.244 (<0.001)	31.8%	2	2	Match
PA→PEOU	0.503 (<0.001)	100.0%	1	1	Match
Model C
PU→IU	0.317 (<0.001)	100.0%	1	1	Match
SE→IU	0.282 (<0.001)	63.1%	2	2	Match
EU→IU	0.124 (<0.001)	34.0%	4	3	Mismatch
FC→IU	0.132 (<0.001)	24.6%	3	4	Mismatch

Source: own calculations. Note: (PT) = perceived trust; (SE) = self-efficacy; (PA) = perceived awareness; (PU) = perceived usefulness; (PEOU) = perceived ease of use; (FC) = facilitating condition; (IU) = behavioral intention to use.

below, Models A and B demonstrate identical factor rankings, suggesting a closer alignment in evaluating AI factors that influence IU and PEB. However, there seems to be a difference in the ranking of factors in Model C.

In Model A, both ANN and PLS-SEM approaches identified self-efficacy as the most significant determinant of perceived usefulness, closely followed by perceived awareness and perceived trust demonstrated significantly weaker influence in both approaches. Model B revealed that perceived awareness is the most critical predictor of perceived ease of use of AI products in ANN and PLS-SEM approaches. This implies that if the users have enough information concerning AI technology, they will find it easy to use AI products. The results indicate that perceived usefulness predominantly influences behavioral intentions to use Model C, followed by self-efficacy in ANN and PLS-SEM. However, perceived ease of use demonstrated a comparatively weaker impact in the PSL-SEM approach than in the ANN approach, while facilitating conditions is the weakest determinant of IU in the ANN approach.

5. Discussion

This study investigates the AI technological factors influencing pro-environmental behavior through behavioral intention to use environmental AI products, utilizing the Technology Acceptance Model (TAM), Unified Theory of Acceptance and Use of Technology (UTAUT), and the Environmentally Responsible Behavior (REB) model. The study addresses existing research gaps by recognizing the impact of AI product factors that prompt individuals to modify their behavior towards environmentally responsible practices. Specifically, the study integrates three approaches—F-DEMATEL, SEM, and ANN—to predict pro-environmental behavior (PEB).

The F-DEMATEL results reveal that external factors such as perceived trust (PT), self-efficacy (SE), and perceived awareness (PA) influence perceived usefulness (PU) and perceived ease of use (PEOU), suggesting that these external factors act as causal drivers, while the TAM variables are the effects dependent on these external factors. In line with the F-DEMATEL results, the Partial Least Squares Structural Equation Modeling (PLS-SEM) results also show positive relationships between these external factors and the TAM’s core variables (PU and PEOU), with PA emerging as the most influential factor. These findings align with previous studies [30,85], which found that internet self-efficacy positively impacts the perceived ease of use and perceived usefulness of internet and mobile health services, and with research by Choung et al. [25] and Shin [27], which confirmed that trust positively predicts perceived usefulness and ease of use of AI products. Additionally, individuals with prior experience with technology and AI (perceived awareness) are more likely to find AI products useful and easy to use [38]. The findings highlight the importance of awareness campaigns to educate the public on the use of environmental AI products.

Moreover, the study confirms the direct influence of self-efficacy, perceived usefulness, perceived ease of use, and facilitating conditions on behavioral intention to use environmental AI products, which in turn positively impact pro-environmental behavior. This is consistent with previous research [50,86]. The strong positive relationship between self-efficacy and the intention to use environmental AI products suggests that individuals with adequate technological skills are more likely to adopt these products. This underscores the need for environmental organizations to enhance public capacity by providing training on the effective use of AI products. The significant positive effects of PU and PEOU on behavioral intention to use environmental AI products correspond to the foundational tenets of the TAM [11], indicating that individuals are more likely to use AI products when they perceive them as useful and easy to use. These results also align with the UTAUT model, where facilitating conditions play a significant role in influencing behavioral intention [14].

However, the constructs of perceived trust, perceived awareness, perceived benefit, and social influence were found not to directly relate to behavioral intention to use environmental AI products, which contradicts some prior research [86], as well as the TAM and UTAUT models. The roles of perceived trust and perceived awareness may have been overshadowed by other TAM factors, such as perceived usefulness and perceived ease of use, which are more direct determinants of behavioral intention [11]. This is supported by the significant indirect positive effects of these factors through TAM variables, confirming their mediation role. While individuals may trust and be aware of the risks associated with environmental AI products, their intention to use them is likely influenced more by their perceived usefulness and ease of use. Furthermore, the limited availability of environmental AI products may prevent widespread access, leading individuals to seek information about their usefulness and ease of use before forming an intention to adopt them. The non-significant effect of social influence on behavioral intention to use environmental AI products aligns with studies by Almaiah et al. [86] and Nguyen et al. [87], which found no direct relationship between social influence and behavioral intention. This may be due to the relatively low public use of AI products, where social influence has minimal impact on the intention to use.

Regarding the REB theory, the PLS-SEM results indicate that behavioral intention to use environmental AI products has a significant positive effect on pro-environmental behavior, consistent with the REB model’s main hypothesis that the intention to act positively influences pro-environmental behavior [88]. This finding aligns with Hines et al.’s [17] argument that possessing the intention to act is a key factor in environmentally responsible behavior. Similar results were found in a study investigating the factors influencing the acceptance and use behavior of AI products in everyday environments [89].

The ANN models largely support the findings from F-DEMATEL and SEM. In Model A, the ANN results demonstrate that self-efficacy (SE) has a greater influence on perceived usefulness (PU), followed by perceived awareness (PA), mirroring the PLS-SEM results. This highlights the crucial roles of SE and PA in influencing PU, which in turn leads to the intention to use environmental AI products and pro-environmental behavior. In Model B, PA has the greatest impact on perceived ease of use (PEOU) among the three factors influencing PEOU, confirming the results from F-DEMATEL and PLS-SEM. Perceived trust (PT) has a weaker predictive impact on PEOU. This underscores the need for government agencies, environmental organizations, and NGOs to enhance awareness campaigns about the importance of using AI products to protect the environment. In Model C, the results differ slightly from those of F-DEMATEL and align more closely with PLS-SEM, where perceived usefulness (PU) has the greatest impact on behavioral intention to use, while facilitating condition (FCs) has less impact. In contrast, F-DEMATEL identified facilitating conditions as the most important predictor of behavioral intention to use.

5.1. Theoretical and Practical Implications

The study has made methodological, theoretical, and managerial contributions. In regard to the methodological approaches, the study integrated F-DEMATEL, SEM, and ANN approaches to analyze data, as they are thought to enhance each other where F-DEMATEL establishes cause-and-effect links between criteria, minimizes uncertainty in experts’ subjective assessments, and enhances the reliability of representation [55], ANN improves prediction accuracy by linking input and output data and identifying both linear and nonlinear relationships and can learn using artificial intelligence techniques while SEM has the capability of analyzing causal relationships.

5.1.1. Theoretical Implications

Theoretically, this study enhances the understanding of how AI product factors influence pro-environmental behavior through behavioral intentions to use AI products. By integrating the Technology Acceptance Model (TAM), the study confirms that external factors significantly affect the core variables of TAM—perceived usefulness (PU) and perceived ease of use (PEOU)—which, in turn, influence behavioral intentions to use environmental AI products. This supports the notion that perceptions of the product’s value and usability are central to individuals’ decisions to adopt AI technologies, particularly in the context of environmental sustainability. Additionally, this study expands the literature by introducing AI technological factors as key determinants of pro-environmental behavior through behavioral intentions. Previous environmental protection studies have predominantly focused on external factors, such as social norms, cost, and inconvenience, as well as individual factors, including demographic and psychological variables, as the primary drivers of pro-environmental behavior [6]. By incorporating AI-specific factors, such as perceived trust, self-efficacy, and perceived awareness, this study contributes a novel perspective on how technological innovations, specifically AI products, can shape individuals’ intentions to engage in environmentally responsible behaviors. This theoretical contribution underscores the growing importance of technology in influencing sustainable practices, offering a new lens through which to examine the intersection of technology adoption and environmental behavior.

5.1.2. Practical Implications

The present study has made important contributions to government agencies, environmental activists, and non-governmental organizations geared toward protecting the environment in Taiwan. Given the backdrop of global climate change, utilizing AI tools to safeguard and preserve the environment is vital. Nevertheless, the majority of individuals are reluctant to embrace AI technology products. In this regard, manufacturers of AI products should prioritize eco-friendly design in their creations. For example, smart home devices (smart lights, smart TVs, smart speakers) ought to be designed to encourage users to engage in more eco-friendly practices. Additionally, stakeholders should improve transparency regarding the functioning of AI and its environmental usefulness, as this can foster greater user trust and encourage the adoption of these products. Secondly, governments, environmental activists, and companies can offer incentives to encourage using AI technologies that support environmental sustainability. This might involve financial incentives for energy-efficient AI technologies or tax advantages for companies implementing AI in environmentally friendly initiatives. Governments and private entities, including academic institutions, should collaborate to develop innovative AI solutions to address environmental challenges. Furthermore, governments should allocate funding to support the development and implementation of environmental AI products. Public awareness campaigns are also essential to educate individuals about the benefits and applications of AI technologies in promoting environmental sustainability and to encourage active engagement with these innovations. As self-efficacy influences both perceived usefulness and perceived ease of use, which leads to behavioral intention to use, thereby affecting people’s pro-environmental behavior, government agencies, and environmental specialists should develop the public’s abilities to easily interact with recent AI technological tools by conducting training programs on how to use AI products effectively.

6. Limitations and Suggested Future Studies

Like any other research study, the present study has certain limitations that future studies may address. First, the current study is focused solely on the context of Taiwan, which raises concerns about the generalizability and applicability of its findings to other settings. Therefore, future research studies could take a cross-national approach to expand the scope of the current study. Second, the present study has not investigated the most commonly preferred and used AI products in environmental protection. Therefore, future studies can explore the most commonly used AI products to comprehend their impact on the environment. Third, regarding the nature of the external factors (perceived trust, self-efficacy, and perceived awareness) that influence the main variables of the TAM, they can vary and grow over time depending on various factors, which may affect the strength of their impact on TAM variables. Hence, with this study employing a cross-sectional design (collecting data at a single point in time), it may not have been able to demonstrate these changes over time; therefore, future studies may consider utilizing a longitudinal design to examine these changes. Fourth, the present study has focused on general AI products rather than one specific product. In this regard, the results might not represent a particular AI product as different products might produce different results; hence, future studies should consider focusing on particular AI products such as chatbots and robots. Fifth, the composition of young people, females, and well-educated individuals indicates that the results might have been biased towards the young and educated because they are more exposed to and comfortable with the technology, leading to quick adoption. Furthermore, the results might be different from those of older people, males, and uneducated people, who might face difficulties interacting with technology. Lastly, since our study used self-reported data, this might have provided room for social desirability bias in which respondents might have reported in their favor since being in favor of the environment is highly accepted [19], suggesting that future studies to consider investigating actual behavior through experimental study designs.

7. Conclusions

Based on the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and the Responsible Environmental Behavior (REB) model, the present study successfully achieved its primary aim by examining the factors related to AI products that influence behavioral intentions to use environmental AI products and, ultimately, pro-environmental behavior in Taiwan. This was accomplished using a novel three-staged approach that integrates Fuzzy DEMATEL (F-DEMATEL), Partial Least Squares Structural Equation Modeling (PLS-SEM), and Artificial Neural Network (ANN). The study identified eight key factors related to behavioral intention to use environmental AI products, which were analyzed using F-DEMATEL based on the opinions of 20 experts. These factors include perceived trust (PT), self-efficacy (SE), perceived awareness (PA), perceived behavior (PB), perceived usefulness (PU), perceived ease of use (PEOU), social influence (SI), and facilitating conditions (FCs).

The F-DEMATEL results revealed that PT, SE, and PA are external factors influencing the main TAM variables, PU and PEOU, which are considered effect factors. Additionally, FC was found to be a causal factor, while PB and SI were identified as effect factors in this study. These findings highlight the complex interrelationships between various external and internal factors that shape the behavioral intentions of individuals toward using environmental AI products.

In the PLS-SEM analysis, a sample of 1726 responses was collected and analyzed. The results indicated that the core TAM variables, PU and PEOU, are influenced by external factors such as PT, SE, and PA. Furthermore, the study found that only four factors—SE, PU, PEOU, and FC—are directly related to behavioral intentions to use environmental AI products. PT and PA, on the other hand, indirectly affect behavioral intentions through their influence on PU and PEOU. Importantly, the study also demonstrated a significant positive relationship between behavioral intention to use environmental AI products and pro-environmental behavior, supporting the idea that intentions to act play a critical role in fostering environmentally responsible actions.

To further validate the results and identify the most critical factors, the study employed ANN modeling. The ANN analysis confirmed that SE has a greater influence on PU, while PA has a stronger impact on PEOU. Additionally, PU emerged as the most influential factor in determining behavioral intention to use environmental AI products. These findings reinforce the importance of factors such as self-efficacy and awareness in shaping individuals’ perceptions of the usefulness and ease of use of AI products, which in turn influence their intentions to adopt these products.

This study makes significant methodological, theoretical, and practical contributions. Methodologically, it introduces a comprehensive approach combining F-DEMATEL, PLS-SEM, and ANN to examine complex relationships between factors influencing behavioral intention to use environmental AI products. Theoretically, the study extends existing models such as TAM, UTAUT, and REB by integrating AI-specific factors that impact pro-environmental behavior. Practically, the findings provide valuable insights for government agencies, environmental activists, and other relevant stakeholders in Taiwan, offering guidance for formulating policies and strategies that encourage the adoption of AI tools to support environmental preservation. By understanding the key factors that drive behavioral intentions to use environmental AI products, policymakers can better design interventions that promote sustainable practices and foster broader adoption of AI technologies in environmental protection efforts.