Factors Influencing Generative AI Usage Intention in China: Extending the Acceptance–Avoidance Framework with Perceived AI Literacy

Abstract

In the digital era, understanding the intention to use generative AI is critical, as it enhances productivity, transforms workflows, and enables humans to focus on higher-value tasks. Drawing upon the unified theory of acceptance and use of technology (UTAUT) and the technology threat avoidance theory (TTAT), this research integrates perceived AI literacy into the AI acceptance–avoidance framework as a central variable. This study gathered 583 valid survey responses from China and validated its model using a dual-phase, combined method that integrates structural equation modeling and artificial neural networks. Research findings indicate that the model explains 51.6% of the variance in generative AI usage intention. Except for social influence, all variables within the extended framework significantly impact the usage intention, with perceived AI literacy being the strongest predictor (β = 0.33, p < 0.001). Additionally, perceived AI literacy mitigates the adverse effect of perceived threats on the intention to use AI. Practical implications suggest that enterprises adopt a tiered strategy, as follows: maximize perceived benefits by integrating AI skills into reward systems and providing task-automation training; minimize perceived costs through dedicated technical support and transparent risk mitigation plans; and cultivate AI literacy via progressive learning paths, advancing from data analysis to innovation.

1. Introduction

The digital transformation of industry is driving an evolution from automation-centric processes to more collaborative, human-centered approaches. While traditional industrial frameworks prioritized operational efficiency through technology-driven optimization, modern approaches increasingly emphasize the reintegration of human creativity into advanced technological systems [1,2]. This shift aims not merely to improve working efficiency, but, more importantly, to liberate employees from repetitive tasks through meaningful human–AI collaboration [3]. Generative AI has emerged as a pivotal driver of this transformation. Unlike traditional AI tools confined to narrow application scenarios, generative AI systems possess the capabilities for cross-scenario operation, interpretation of complex unstructured information, and support for natural human–computer interaction through multimodal interfaces [4]. These capabilities allow AI systems to augment human labor by automating routine tasks, thereby freeing cognitive resources for creative and higher-order thinking [1,5].

However, no matter how advanced the technology, its actual impact ultimately depends on human acceptance and meaningful utilization. This is particularly crucial in the context of Industry 5.0, which—like generative AI—is inherently human-centric. Specifically, Industry 5.0 relies on generative AI to automate intricate processes in smart manufacturing, healthcare, and logistics, while the effectiveness of generative AI in these scenarios directly hinges on human trust and proactive collaboration [2,5]. A vivid example is Siemens’ industrial copilots. Generative AI-based systems can respond to verbal instructions, dynamically optimize processes, and adapt to user preferences, intuitively demonstrating the potential of this technology to reshape workplace collaboration models.

Beyond industrial use cases, generative AI applications such as ChatGPT and DeepSeek have allowed users to efficiently generate coherent text, images, and code with minimal input, drastically reducing time and effort for various cognitive tasks [6,7]. In China, generative AI has exhibited substantial advancements. According to relevant data, by 2024, the number of enterprises involved in generative AI in China is expected to exceed 4500, with the core industry scale nearing CNY 600 billion [8]. An extensive industrial chain, encompassing fundamental research, technological development, and industrial application, has been established, thereby facilitating the widespread application and innovative advancement of generative AI across various sectors in China. These developments collectively highlight the increasingly crucial role of generative AI in shaping the future of work. However, the extent of technological implementation will ultimately depend on users’ actual acceptance and willingness to adopt it. Despite the rapid enhancement of generative AI’s technical capabilities, user engagement remains inconsistent. For instance, research indicates that while some users enthusiastically embrace these tools, others exhibit skepticism or resistance due to concerns about accuracy, privacy, job displacement, or ethical issues [9,10].

When it comes to the antecedents of generative AI usage intention, studies are mainly focused on UTAUT. While widely explored, the application of UTAUT to generative AI has yielded inconsistent results [11,12], and UTAUT primarily focuses on positive motivators for technology use, often neglecting users’ avoidance behaviors, particularly those triggered by emerging, unfamiliar, or potentially disruptive technologies, such as generative AI [13]. For example, while researchers have verified the significant influence of performance expectancy and facilitating conditions (both key factors of UTAUT) on mobile banking adoption [12], recent research found that these two constructs exerted a non-significant influence on generative AI adoption [11]. Moreover, existing UTAUT studies are largely concentrated in the educational sector, leaving sectors such as manufacturing, services, and other industries underexplored—sectors where human–AI collaboration and automation threat perceptions may significantly alter adoption dynamics [12,14]. To address this gap, TTAT provides a complementary perspective and posits that individuals evaluate technologies based on perceived threats, such as security vulnerabilities or loss of control, and that such perceptions shape avoidance motivations [15]. In the case of generative AI, users may be concerned about misinformation, AI-generated errors, or over-reliance, all of which can affect their decision to adopt generative AI [9,16]. Therefore, combining UTAUT and TTAT enables a more holistic understanding of both approach and avoidance motivations in the context of generative AI adoption.

However, the translation of user motivations into actual behavior hinges critically on their cognitive capacities to interact with AI systems, a dimension unaccounted for in either UTAUT or TTAT frameworks [17,18]. As generative AI becomes increasingly integrated into everyday work and life, the ability to understand, evaluate, and interact with AI technologies becomes crucial, not only for the adoption of AI but also for the responsible use of AI [19]. This introduces the concept of perceptual AI literacy (PAIL), which is defined as the ability to self-assess and meaningfully understand and interact with AI systems [20,21]. Perceptual AI literacy includes awareness of the capabilities and limitations of AI, understanding of ethical implications, and confidence in making informed decisions about the willingness to use AI [22]. A WEF report [23] indicates that 74.9% of business organizations plan to adopt AI technology by 2027. However, relatively few employees have received relevant skills training. This imbalance between technology penetration and capability readiness is widespread across countries. In this context, people with a higher awareness of AI literacy tend to better recognize the benefits of AI while dealing with the associated risks, thus more effectively balancing the acceptance and avoidance responses [24]. The contradiction between the speed of technological development and the level of public awareness makes perceived AI literacy a potentially key motivation and moderating factor influencing willingness to use generative AI [3]. Understanding this cognitive dimension can inform educational strategies, platform designs, and policy interventions aimed at achieving inclusive AI adoption.

Motivated by the theoretical and empirical gaps in China’s generative AI adoption landscape, this study constructs an integrated framework synthesizing the acceptance–avoidance paradigm with UTAUT, TTAT, and perceived AI literacy. Specifically, we explore how performance expectations, effort expectations, social impact, and facilitating conditions (from UTAUT); perceived threats and perceived avoidability (from TTAT); and cognitive self-efficacy (through PAIL) interact to influence behavioral willingness to use generative AI across various industries in China. This comprehensive, multi-dimensional theoretical model addresses existing research gaps by integrating cognitive factors into the framework of human–machine collaboration, considering proximity and avoidance motives as well as cognitive abilities. It elucidates the adoption of generative AI from multiple perspectives, acknowledging that willingness is influenced by both rational evaluation and emotional response, and is further modulated by an individual’s perceptual capacity. By clarifying the mechanisms through which various factors impact the willingness to engage with generative AI, this study enhances the theoretical understanding of generative AI adoption, achieves theoretical advancements, and contributes to the broader discourse on technology acceptance and AI applications.

The rest of the study is organized as follows: Section 2 provides an in-depth literature review and hypothesis development. Section 3 provides a detailed account of the research methods, and Section 4 presents the results of partial least squares (PLS) tests and artificial neural network analyses. Section 5 explores the theoretical and administrative implications of the findings. Finally, Section 6 points out the limitations of the study and presents directions for future research.

2. Theoretical Foundations and Hypothesis Development

2.1. UTAUT, TTAT, Perceived AI Literacy, and Generative AI Usage Intention

The swift advancement of generative AI, exemplified by large language models, image generators, and autonomous coding agents, has significantly transformed interactions between individuals, organizations, and intelligent systems, thereby enhancing innovation and economic performance in firms [25,26,27]. This shift is particularly pronounced in China, where the integration of generative AI is deeply rooted in a cultural emphasis on technological progress and a positive legal framework. At the cultural level, China has long regarded technological innovation as the cornerstone of national development, fostering social acceptance of AI tools through extensive digitalization and government-led “smart upgrade” campaigns across industries. As these technologies increasingly permeate areas such as education, healthcare, information technology, and manufacturing, understanding the factors that drive users to adopt generative AI has become a key research focus. The research regarding generative AI usage intention primarily follows the technology acceptance model [28], which identifies perceived usefulness and ease of use as key determinants of individuals’ and organizations’ technology adoption behavior. This model has served as a foundational framework for subsequent studies assessing user attitudes and intentions toward emerging digital technologies.

However, the unique opportunities and risks associated with generative AI necessitate an expanded analytical lens. To address this, researchers have developed integrated technology acceptance frameworks that extend beyond traditional models, including the UTAUT. As a meta-theory synthesizing eight prior models (e.g., TAM, TPB), UTAUT establishes performance expectancy, effort expectancy, social influence, and facilitating conditions as core predictors of IT adoption. Its application to generative AI tools reveals critical influences on user acceptance; students’ adoption intention is significantly shaped by perceived effort requirements (effort expectancy), output reliability concerns (performance expectancy), peer/organizational pressures (social influence), and technical/institutional support structures (facilitating conditions) [29].

Nevertheless, UTAUT predominantly emphasizes technology’s positive utility while inadequately addressing risk perceptions and avoidance tendencies toward emerging technologies. While generative AI offers notable advantages, users often experience psychological barriers stemming from perceived threats. For example, Pramod et al. [30] reported that concerns over misinformation generation and deepfake misuse significantly suppressed usage intentions of generative AI among social media users. Similarly, generative AI adoption in creative industries has been investigated [31], and the findings particularly emphasized security concerns in construction supply chains, where AI adoption is hindered by cybersecurity threats, cost–risk trade-offs, algorithmic opacity, and challenges in establishing trust and quantifiable benefits.

Therefore, some studies have applied TTAT to examine privacy concerns associated with AI-generated content, revealing that individuals with high perceived threat severity were less likely to adopt AI applications. TTAT, initially proposed by Liang and Xue [32], provides a threat appraisal framework that explains how individuals assess and respond to perceived IT-related threats. The theory outlines four key factors—perceived threat susceptibility, perceived threat severity, safeguard effectiveness, and self-efficacy—that jointly influence an individual’s motivation to avoid or reject a technology. Within the realm of generative AI, TTAT has been adopted to investigate users’ defensive reactions toward tools like ChatGPT, especially when potential harms are perceived to outweigh functional benefits.

Whether examining UTAUT’s motivational drivers or TTAT’s threat mitigation mechanisms, their operational efficacy remains fundamentally contingent on users’ cognitive engagement with AI technology. Literacy—originally denoting reading/writing proficiency—now encompasses domain-specific skill acquisition coupled with contextually adaptive abilities to understand, interpret, apply, and communicate such competencies [33]. Given artificial intelligence’s nature as a new technological science simulating and extending human intelligence [34], AI literacy emerges as critical for evaluating human–AI interaction capabilities. Specifically, it entails the capacity to ethically identify, utilize, and assess AI products [34], while enabling intuitive operational proficiency and effective tool management [35]. This construct formalizes core competencies, allowing users to critically evaluate AI systems, efficiently collaborate with algorithmic assistants, and confidently deploy solutions across diverse settings [20]. Empirical evidence confirms its impact; AI literacy accelerates educators’ technological–pedagogical knowledge, directly increasing generative AI adoption in higher education [18], while substantially strengthening students’ attitude–behavioral linkages [17].

Beyond productivity enhancements, generative AI simultaneously introduces ethical dilemmas spanning data privacy violations, algorithmic biases, and deepfake dissemination. Addressing these necessitates systematic literacy development, integrating three pillars—technical knowledge acquisition, applied skill refinement, and ethical critical thinking. Crucially, objective AI literacy diverges from perceived AI literacy—the latter more directly determines psychological acceptance and behavioral outcomes. Users with elevated perceived literacy demonstrate heightened ethical responsibility awareness and proactive risk mitigation, fostering sustainable AI implementation [36]. This perception significantly mediates relationships between perceived usefulness/ease of use and adoption intentions [22], underscoring the necessity of integrating AI literacy into user training, system design, and policy efforts to ensure equitable and confident usage of generative AI technologies.

Overall, existing research still exhibits three critical gaps in explaining willingness to use generative AI, necessitating an integrated theoretical framework. First, most empirical applications of UTAUT and TTAT are concentrated in the educational domain, particularly among students and educators. Research on generative AI adoption is limited in sectors such as manufacturing and customer-facing service industries, where contextual variables (e.g., automation threat or human–AI collaboration models) may lead to different adoption dynamics, and the existing results about UTAUT and TTAT on technology usage intention are still inconsistent. Second, while AI literacy is acknowledged as pivotal, existing research remains limited and fragmented, with few studies having integrated it into TTAT models to further explore its mechanisms on generative AI usage intention. This gap persists despite empirical evidence from Cho and Jeong [36], whose findings demonstrate AI literacy can play a critical moderating effect between privacy and perceived usefulness and ChatGPT usage. Third, there is a conspicuous absence of an integrated framework that combines AI literacy, opportunity-enhancing and threat-avoiding factors, and assesses the relative importance of competing drivers and inhibitors in forming user generative AI usage intention. Therefore, by integrating based on UTAUT, TTAT, and perceived AI literacy, this study endeavors to provide a comprehensive framework that elucidates the determinants and their significance for generative AI usage intention within the education, manufacturing, information technology (IT), and service sectors in China.

2.2. UTAUT and Generative AI Usage Intention

UTAUT points out that individuals’ and organizations’ behavioral intentions are determined by the following four core factors: performance expectancy (PE), effort expectancy (EE), social influence (SI), and facilitating conditions (FC) [29,37].

Performance expectancy refers to the extent to which individuals perceive that a technology will improve their productivity, efficiency, and overall outcomes [37]. Firstly, generative AI elevates learning outcomes by offering customized learning opportunities, generating study materials, and offering instant feedback [14]. For example, studies demonstrate that performance expectancy critically shapes students’ behavioral intentions toward using generative AI in educational contexts [38], as they perceive that these tools can improve their educational performance. Secondly, generative AI is capable of managing repetitive tasks, thereby enabling people to concentrate on intricate and creative work [14,39]. This perceived improvement in efficiency can drive users to adopt generative AI [40], and previous research also confirmed that performance expectancy critically influences individuals’ willingness to engage with generative AI [39]. Thirdly, generative AI can inspire new ideas and solutions by generating novel content and suggestions, enhancing innovation and creativity, which users might find beneficial and inspiring, particularly for new product development [13,41]. For example, Xia and Chen [13] pointed out that performance expectancy can significantly promote users’ generative AI usage intention in new product development, as generative AI not only improves innovation efficiency but also introduces new technologies and knowledge, facilitating the generation of new ideas. Therefore, when users perceive a higher level of performance expectancy, they are more inclined to use generative AI. Consequently, we have the following:

H1. 

Performance expectancy will positively affect users’ generative AI usage intention.

Effort expectancy is the level of ease in utilizing a particular technology [37], and its importance in generative AI stems primarily from its role in reflecting users’ opinions on the technology’s user-friendliness and accessibility [42]. Firstly, users are more inclined to adopt technologies that are easy to comprehend and require minimal effort to learn [43]. User-friendly designs reduce learning demands, increasing adoption intention when users perceive low operational effort. Thus, technology acceptance intensifies with system usability [44]. Additionally, the availability of generative AI tools across multiple devices (e.g., mobile phones, laptops, tablets) can enhance users’ intention to engage with the technology by reducing effort expectancy [45]. This accessibility enables users to interact with the technology more conveniently, thereby further augmenting their intention to use it. Consequently, we propose the following:

H2. 

Effort expectancy will positively affect users’ generative AI usage intention.

Social influence is the degree to which a person feels that significant others, such as colleagues, friends, or supervisors, think they should adopt a specific technology [37]. Social influence is a key factor in influencing users’ intentions to utilize a certain technology. Firstly, social influence can exert peer pressure and provide encouragement within a network, thereby increasing the likelihood that users will adopt generative AI if they perceive their peers and colleagues are utilizing it [45,46]. For example, behavioral intention shows significant dependence on social influences from academic and peer sources [47]. Moreover, social influence can also emanate from organizations, influenced by the attitudes and cultures of top management towards AI technologies, such as organizational AI training programs and positive reinforcement from supervisors [48]. When users perceive organizational support for generative AI, they are more inclined to adopt it to gain legitimacy within the organization [49]. For example, Yakubu et al. [14] demonstrated that social influence through peer pressure and environmental support can significantly enhance people’s learning and usage of content-generative AI tools. Nevertheless, the efficacy of social influence is subject to contextual contingencies, as its impact may be moderated by individual differences or situational factors [50,51]. Hence, we propose the following:

H3. 

Social influence will positively affect users’ generative AI usage intention.

Facilitating conditions reflect perceived availability of essential resources and support for utilizing a specific technology [37], and are significant as they embody user-perceived accessibility of technical infrastructure, organizational training, and resource support. Firstly, organizational support in terms of technical infrastructure—such as the provision of high-speed internet, well-integrated AI platforms, and access to digital resources—has a major effect on users’ willingness to engage with generative AI [48]. For instance, Jo and Bang [48] identified network quality, accessibility, and system responsiveness as critical predictors of ChatGPT adoption, usage intention, and user satisfaction. Furthermore, beyond infrastructure support, organizational training and education related to AI are vital for enhancing user satisfaction. Access to AI-related training programs and educational resources can bolster users’ confidence and competence within organizations, thereby mitigating and preventing the risk of AI-related issues [16,52]. Consequently, providing organizational technical infrastructure support along with AI-related training and education can enhance their willingness to adopt AI tools. Thus, we propose the following:

H4. 

Facilitating conditions will positively affect users’ generative AI usage intention.

2.3. TTAT and Generative AI Usage Intention

Malicious IT programs—including viruses, spam, spyware, and adware—have demonstrably compromised personal and corporate IT infrastructure, resulting in significant economic losses [53]. Developed through an integration of literature from healthcare, cybernetics, psychology, and information systems, TTAT elucidates the motivations and mechanisms underlying users’ avoidance of malicious IT threats [32,54]. The theory claims that people first evaluate the presence and magnitude of an IT threat before determining appropriate mitigation strategies [55]. Crucially, TTAT conceptualizes threat avoidance as a cybernetic process where individuals endeavor to expand the disparity between their current secure state and a potential insecure state [32,56]. The actions undertaken by users are contingent upon their assessment of the threat’s avoidability. When threats are perceived as avoidable, users typically engage in problem-focused protective measures; conversely, when threats are deemed unavoidable, users often resort to emotion-focused coping strategies [15]. Consequently, users may refrain from utilizing specific information technology systems, such as generative artificial intelligence, if they perceive substantial unmitigable risks, yet they may persist in their usage if they believe that threats can be effectively managed.

Central to the predictive efficacy of TTAT are two pivotal constructs, as follows: perceived threats and perceived avoidability. Perceived threats initiate the avoidance process by indicating potential harm, while perceived avoidability determines the feasibility and nature of the response [15,32]. Collectively, these constructs constitute the primary cognitive appraisal mechanisms through which TTAT explicates behavioral outcomes, making them indispensable for investigating the adoption of generative artificial intelligence. Their selection is also theoretically substantiated, as they embody the fundamental dimensions of threat evaluation (perceived threat) and coping appraisal (perceived avoidability) that collectively drive avoidance motivation and subsequent actions within the TTAT framework.

Perceived threats describe how a person assesses the danger or harm connected with a particular information technology event or system [15]. Information technology security threats are pervasive and indiscriminate, possessing the capacity to impact anyone, at any time, and in any place. Such threats often provoke negative emotional responses, including fear and anxiety, due to the potential for substantial losses. In the context of generative artificial intelligence, users who perceive threats express concerns about potential risks such as data breaches, misinformation generation, or data misuse, all of which could disrupt tasks or cause personal or professional harm. These perceived threats can negatively influence the willingness to adopt generative AI technologies. Supporting this notion, research demonstrates that perceived risks adversely affect attitudes toward AI-driven smart healthcare among non-clinicians [57] and shape stock investors’ intentions to use robo-advisors [58]. This evidence suggests that the decision to refrain from using such technologies is frequently regarded as the most effective strategy for risk mitigation. Therefore, we have the following:

H5. 

Perceived threats will negatively affect users’ generative AI usage intention.

Perceived avoidability reflects one’s assessment of the probability of successfully circumventing IT security threats related to a certain technology. This assessment is informed by the perceived effectiveness, feasibility, and cost-efficiency of available protective measures [32,59]. According to the technology threat avoidance theory (TTAT), perceived controllability influences coping strategies. Users tend to adopt problem-focused coping when they perceive a high degree of control, whereas emotion-focused coping is more common when perceived control is limited [59]. Therefore, perceived avoidability encapsulates the users’ perceived control over security threats related to generative AI. A heightened sense of control enhances feelings of safety and emotional stability during usage, whereas diminished perceived control may lead to insecurity and emotional disturbance. Consequently, when users perceive the level of avoidability is high, they are inclined to believe that effective and feasible protective measures are available, thereby reinforcing their sense of security. This perception of security and control over potential threats can positively impact their willingness to engage with the technology. Thus, the following hypothesis is proposed:

H6. 

Perceived avoidability will positively affect users’ generative AI usage intention.

2.4. Perceived AI Literacy and Generative AI Usage Intention

Perceived AI literacy refers to the level of understanding and competence individuals have in using and interacting with AI technologies [36]. Therefore, higher AI literacy can boost users’ confidence, making them more willing to adopt and integrate AI tools into their activities [18]. For example, some studies find that AI literacy can promote users’ perceptions and intentions to use AI tools in the education industry [22]. Moreover, Al-Abdullatif [18] points out that AI literacy can also promote the user’s perceived trust, which then increases the usage of generative AI. Secondly, increased literacy helps users comprehend the strengths and weaknesses of generative AI, resulting in more informed and positive perceptions of its potential benefits. According to Almatrafi et al. [60], AI literacy significantly enhances users’ beliefs concerning the potential advantages of AI tools. Similarly, Bozkurt and Sharma [61] highlighted that AI literacy can reduce the perceived complexity and effort required to use AI tools. That is, AI literacy empowers users to understand the impact of minor language variations on generative AI responses, leading to enhanced user-AI communication and interaction, which ultimately shapes their intentions to use this technology. Therefore, we propose the following:

H7. 

Perceived AI literacy will positively affect users’ generative AI usage intention.

Beyond its direct impact, perceived AI literacy plays a crucial role in moderating the relationship between perceived threat and users’ generative AI usage intention. Perceived threat refers to users’ subjective evaluation of systemic risks associated with generative AI applications, encompassing a range of multidimensional concerns, as follows: privacy and security vulnerabilities, such as data leakage and algorithmic bias; ethical challenges, including deepfakes and ambiguities in responsibility; and socioeconomic disruptions, such as employment restructuring and skill displacement [10]. Functioning as a multidimensional cognitive capability, perceived AI literacy encompasses technical understanding, ethical discernment, and risk mitigation proficiency. Enhanced literacy reduces threat perception through two mechanisms: cognitive restructuring allows users to comprehend AI’s foundational logic—such as machine learning workflows and data processing patterns—thereby transforming abstract risks into manageable challenges; while behavioral empowerment enables proactive countermeasures, including data anonymization techniques, smart contract audits, and critical output evaluation using established ethical frameworks [56,57]. In essence, users with advanced AI literacy can not only identify potential vulnerabilities but also develop a comprehensive protection system.

This competency fundamentally reshapes threat perception. When users possess demonstrated capacity to neutralize risks, perceived threat evolves from an overwhelming systemic hazard into a manageable technical challenge. Consequently, perceived AI literacy moderates threat effects by enhancing users’ risk control efficacy, ultimately attenuating its inhibitory influence on usage intention. Hence, the following research hypothesis is proposed:

H7a. 

Perceived AI literacy will positively moderate the relationship between perceived threat and users’ generative AI usage intentions.

Similarly, perceived AI literacy significantly strengthens the relationship between perceived avoidability and users’ generative AI usage intention. Perceived avoidability represents users’ subjective assessment of technology avoidance costs, integrating direct expenditures such as time and financial inputs with indirect costs including opportunity loss and social disconnection [15]. This concept includes direct costs (e.g., time and economic input) and indirect costs (e.g., missing potential benefits, social isolation), reflecting the risk trade-off mechanism in individuals’ technology decision-making process [62].

AI literacy serves a dual moderating function in this context. According to the cognitive value theory of technology adoption [63], enhanced AI literacy significantly elevates users’ cognitive understanding of the transformative potential of technology. This understanding extends beyond recognizing explicit benefits such as increased efficiency (e.g., automated content creation, accelerated data processing) and innovation empowerment (e.g., fostering creativity, integrating cross-domain knowledge) associated with generative AI. It also encompasses an appreciation of implicit values, such as technology-driven social advancement and industrial upgrading. When users develop a comprehensive understanding of AI’s multi-dimensional value, avoiding its use is no longer perceived as a risk-reduction strategy, but rather as an opportunity cost resulting from the deliberate forfeiture of potential value [63].

On the other hand, according to the risk perception and management theory [63], AI literacy equips users with a structured framework for risk response. Individuals possessing advanced literacy levels are adept at employing security protocols to convert perceived avoidability into a psychological safety boundary conducive to technology adoption. This transformative mechanism facilitates the evolution of users’ understanding of “avoidability” from a passive stance of “I can avoid harm” to an active approach of “I can safely harness technical value”. Through the combined effects of “value cognition enhancement” and “risk control empowerment”, AI literacy substantially amplifies the positive influence of perceived avoidability on the intention to use technology. Consequently, the following research hypothesis is proposed:

H7b. 

Perceived AI literacy will positively moderate the relationship between perceived avoidability and users’ generative AI usage intention.

As depicted in Figure 1, the research model is developed based on the assumptions mentioned above, as well as the control variables.

3. Methodology

3.1. Instrument Development

Eight constructs were measured on seven-point Likert scales. Existing scales were used when available; otherwise, items were derived from the most similar scales. Measures of performance expectancy, effort expectancy, social influence, and facilitating conditions were drawn from Venkatesh et al. [64] and Cao et al. [65]. Measures of perceived threat were adopted from Baabdullah et al. [66] and Liang et al. [15], and items of perceived avoidability were drawn from Liang et al. [56]. Perceived AI literacy is measured by the Perceived Artificial Intelligence Literacy Questionnaire (PAILQ-6) [67], and generative AI usage intention is measured by Venkatesh et al. [64] and Meng et al. [68]. Following previous studies concerning AI usage intention, gender, age, educational level, industry, and work experience are chosen as control variables [65,66,69]. Specific items for the scales are provided in

Table A1. Measurement items.

Constructs	Items	Sources
Performance Expectancy (PE)	PE1: I find Generative AI useful in my daily life.	Venkatesh et al. [64] and Cao et al. [65]
	PE2: Using Generative AI increases my chances of achieving tasks that are important to me.
	PE3: Using Generative AI helps me accomplish tasks more quickly.
	PE4: Using Generative AI increases my productivity.
Effort Expectancy (EE)	EE1: Learning how to use Generative AI is easy for me.	Venkatesh et al. [64] and Cao et al. [65]
	EE2: My interaction with Generative AI is clear and understandable.
	EE3: I find Generative AI easy to use.
	EE4: It is easy for me to become skillful at using Generative AI.
Social Influence (SI)	SI1: People who are important to me think I should use Generative AI.	Venkatesh et al. [64] and Cao et al. [65]
	SI2: People who influence my behavior think I should use Generative AI.
	SI3: People whose opinions I value prefer that I use Generative AI.
Facilitating Conditions (FC)	FC1: I have the resources necessary to use Generative AI.	Venkatesh et al. [64] and Cao et al. [65]
	FC2: I have the knowledge necessary to understand Generative AI.
	FC3: Generative AI is compatible with other technologies I use.
	FC4: I can get help from others when I have difficulties using Generative AI.
Perceived Threats (PT)	PT1: My fear of exposure to Generative AI’s risks is high.	Baabdullah et al. [56] and Liang et al. [32]
	PT2: The extent of my anxiety about potential loss due to Generative AI’s risks is high.
	PT3: The extent of my worry about Generative AI’s risks due to misuse is high.
Perceived Avoidability (PA)	PA1: Taking everything into consideration (effectiveness of countermeasures, costs, and my confidence in employing countermeasures), the threat of Generative AI could be prevented.	Liang et al. [32]
	PA2. Taking everything into consideration (effectiveness of countermeasures, costs, and my confidence in employing countermeasures), I could protect myself from the threat of Generative AI.
	PA3. Taking everything into consideration (effectiveness of countermeasures, costs, and my confidence in employing countermeasures), the threat of Generative AI was avoidable.
Perceived Artificial Intelligence Literacy (PAIL)	PAIL1: I understand the basic concepts of Generative artificial intelligence.	Grassini [67]
	PAIL2: I believe l can contribute to Generative AI projects. (dropped)
	PAIL3: I can judge the pros and cons of Generative AI.
	PAIL4: I keep up with the latest Generative AI trends.
	PAIL5: I’m comfortable talking about Generative AI with others.
	PAIL6: I can think of new ways to use existing Generative AI tools.
Generative AI Usage Intention (UI)	UI1: I intend to use Generative AI in the future.	Venkatesh et al. [64] and Meng et al. [68]
	UI2: I plan to use Generative AI in future.
	UI3: I predict I will use Generative AI in the future.

.

Following the procedures of the back-translation method [70], the English scales were translated into Chinese for the respondents. Initially, the authors translated the original items into Chinese, making minor wording adjustments to ensure alignment with the research context. Subsequently, a group of bilingual doctoral candidates independently performed back-translation to identify and reconcile conceptual discrepancies.

To establish content validity, a structured expert review was performed. Three domain experts (two academic researchers and one industry practitioner) independently evaluated each item’s theoretical relevance and contextual appropriateness. Their assessments underwent iterative consensus discussions until unanimous agreement was reached on a finalized item set comprehensively covering all target constructs.

A formal pilot analysis was then administered at a Chinese technology firm using a 30-employee sample. This phase employed cognitive interviews and iterative revisions to evaluate item clarity, logical flow, and administrative feasibility. Ambiguous expressions were refined through participant feedback cycles, with problematic items eliminated to produce the psychometrically optimized Chinese questionnaire.

In this study, the item PAIL2: “I believe I can contribute to Generative AI projects” was dropped from the perceived AI literacy (PAIL) scale due to both theoretical and empirical concerns. Firstly, conceptually, perceived AI literacy refers to individuals’ ability to understand, evaluate, and responsibly engage with AI technologies [21], rather than their capacity to contribute to AI projects. PAIL2 introduces a notion of self-assessed competence in contributing to AI development, which aligns more closely with self-efficacy or professional capability than with perceived literacy [20]. Including this item risks conflating cognitive understanding with project-based competence, thereby weakening the construct validity of the scale. Secondly, PAIL2 lacks dimensional alignment with the remaining items, which emphasize knowledge, critical judgment, communication, and creative application—dimensions more consistent with established definitions of perceived AI literacy [24]. Contributing to generative AI projects often requires advanced technical skills and contextual resources that go beyond literacy and may introduce irrelevant variance [71]. Thirdly, the results of reliability and validity of measurements also show that the factor loading of PAIL2 was below 0.50, which indicates that PAIL2 has a weak correlation with the underlying latent construct and suggests that the item does not contribute meaningfully to the measurement of perceived AI literacy. Therefore, in this study, the PAIL2 has been dropped.

3.2. Data Collection

This study employed purposive convenience sampling with sector stratification in a major western Chinese provincial capital—a national hub for education, technology, and manufacturing—to optimize accessibility while ensuring contextual representativeness. Industry stratification targeted five generative AI-intensive domains (IT, finance, manufacturing, education, services), identified through regional economic reports [72]. Companies were selected via academic–industry networks based on active AI implementation, with employee eligibility strictly limited to full-time staff (>6 months’ tenure), systematically excluding contractors and interns.

A total of 1000 questionnaires were administered through industry partnerships: senior managers and HR departments facilitated on-site distribution with Survey One (dependent variables/moderators/demographics), followed by Survey Two (independent variables) after a two-week interval. For employees who completed Wave 1 but were absent during Wave 2 on-site sessions, HR managers proactively arranged physical questionnaire completion upon their return to company premises. Completed paper surveys were then returned to the research team via express mail. Cross-wave matching utilized unique participant IDs (last 4 mobile digits) recorded in the top-right corner of both questionnaires, yielding 613 returns. After excluding submissions with missing data or response inconsistencies, 583 valid responses were retained (58.3% effective rate).

3.3. Data Analysis

Measurement adequacy and structural validity were examined via partial least squares (PLS), with SmartPLS 4.1.0 serving as the primary analytical tool. Given that PLS is primarily aimed at optimizing the prediction of target variables [73], it is well-suited for testing exploratory models and analyzing complex interactions [74], aligning with the objectives of our study. Moreover, the specific choice of SmartPLS 4.1.0 is justified by its advanced methodological capabilities essential for validating complex models and handling intricate interaction terms [75]. These enhancements collectively enable rigorous validation of complex frameworks while optimizing computational efficiency.

4. Results

4.1. Respondent Profile

As summarized in

Table 1. Respondent characteristics (n = 583).

Demographics		n	Percentage
Gender	Male	296	50.8
	Female	287	49.2
Age	20–29	194	33.3
	30–39	163	28.0
	40–49	172	29.5
	50–59	54	9.3
Educational level	High school	12	2.1
	College	31	5.3
	Undergraduate	388	66.6
	Graduate	148	25.4
	Doctoral	4	0.7
Industry	Service	89	15.3
	Financial	94	16.1
	Education	128	22.0
	IT	168	28.8
	Manufacturing	104	17.8
Work experience	<5 years	203	34.8
	5–10 years	220	37.7
	>10 years	160	27.4

, out of the 583 respondents, 50.8% were male. Participant demographics reveal balanced gender representation and high engagement of prime working-age cohorts (90.8% aged 20–49). The sample captures extensive industry experience (65.1% with >5 years’ tenure) and elevated education levels (92.7% bachelor’s or higher). Critically, industry distribution—IT (28.8%), manufacturing (17.8%), finance (16.1%), education (22.0%), and services (15.3%)—directly aligns with sectors identified in China’s AI Development Plan as primary adopters of generative AI [72]. This intentional inclusion of key generative AI-adopting sectors, combined with the city’s dual economic structure, ensures the findings capture critical variations in generative AI usage intention.

4.2. Non-Response Bias and Common-Method Bias

To rigorously evaluate potential non-response bias (NRB), we employed a dual-validation approach. First, temporal analysis comparing early and late respondents showed no statistically significant differences in demographic attributes and core research variables (e.g., generative AI adoption intention), indicating minimal NRB risk [76]. Second, structured callback interviews with five respondents and five non-respondents revealed comparable levels of adoption-related perceptions, further evidencing the absence of systematic response reluctance tied to study constructs [77]. Collectively, NRB does not compromise the validity of our findings.

Three methods were employed to control and test the extent of CMB. First, we implemented temporal separation by administering independent and dependent variables across distinct survey phases conducted at different time points. To enable accurate cross-phase participant matching, unique identifiers (last four digits of mobile numbers) were assigned during the initial phase and replicated in the second-phase questionnaire header. This temporal segregation mechanism served to disrupt respondents’ cognitive associations between variable sets, thereby mitigating the risk of artificially inflated correlations attributable to common method variance. Second, Harman’s single-factor test indicated acceptable common method bias with the largest factor accounting for 30.92% variance, less than the recommended value of 40% [78]. Third, as shown in

Table A2. Common method variance (CMV).

Construct	Indicators	Substantive Factor Loading (Ra)	Ra2	Method Factor Loading (Rb)	Rb2
PE	PE→PE1	0.757	0.573	0.040	0.002
	PE→PE2	0.704	0.496	0.113	0.013
	PE→PE3	0.731	0.534	−0.033	0.001
	PE→PE4	0.846	0.716	−0.131	0.017
EE	EE→EE1	0.805	0.648	−0.076	0.006
	EE→EE2	0.817	0.667	−0.095	0.009
	EE→EE3	0.802	0.643	0.009	0.000
	EE→EE4	0.661	0.437	0.162	0.026
SI	SI→SI1	0.850	0.723	0.026	0.001
	SI→SI2	0.847	0.717	0.004	0.000
	SI→SI3	0.898	0.806	−0.030	0.001
FC	FC→FC1	0.714	0.510	0.037	0.001
	FC→FC2	0.755	0.570	−0.040	0.002
	FC→FC3	0.750	0.563	−0.016	0.000
	FC→FC4	0.797	0.635	0.016	0.000
PT	PT→PT1	0.820	0.672	−0.013	0.000
	PT→PT2	0.830	0.689	0.000	0.000
	PT→PT3	0.799	0.638	0.012	0.000
PA	PA→PA1	0.815	0.664	0.014	0.000
	PA→PA2	0.760	0.578	0.061	0.004
	PA→PA3	0.909	0.826	−0.071	0.005
PAIL	PAIL→PAIL1	0.853	0.728	0.027	0.001
	PAIL→PAIL3	0.740	0.548	0.109	0.012
	PAIL→PAIL4	0.916	0.839	−0.031	0.001
	PAIL→PAIL5	0.916	0.839	−0.038	0.001
	PAIL→PAIL6	0.775	0.601	−0.071	0.005
UI	UI→UI1	0.776	0.602	0.071	0.005
	UI→UI2	0.774	0.599	0.074	0.005
	UI→UI3	0.930	0.865	−0.150	0.023
	Average		0.653		0.005

, the unmeasured latent method construct (ULMC) test shows that the substantive variance of each indicator significantly exceeds the method variance [79]. Consequently, CMB poses minimal concern.

4.3. Measurement Model Evaluation

The reliability and validity of measurements are evaluated following conventional approaches [79]. Construct reliability was tested using Cronbach’s alpha and composite reliability score.

Table 2. Loading, Cronbach’s alpha, CR, and AVE.

Constructs	Items	Loadings	Cronbach’s Alpha (α)	Composite Reliability (CR)	Average Variance Extracted (AVE)
Performance Expectancy (PE)	PE1	0.789	0.753	0.844	0.574
	PE2	0.784
	PE3	0.726
	PE4	0.730
Effort Expectancy (EE)	EE1	0.755	0.773	0.854	0.595
	EE2	0.750
	EE3	0.815
	EE4	0.763
Social Influence (SI)	SI1	0.875	0.832	0.899	0.748
	SI2	0.863
	SI3	0.856
Facilitating Conditions (FC)	FC1	0.777	0.746	0.839	0.566
	FC2	0.736
	FC3	0.704
	FC4	0.790
Perceived Threats (PT)	PT1	0.813	0.75	0.857	0.666
	PT2	0.829
	PT3	0.807
Perceived Avoidability (PA)	PA1	0.835	0.772	0.868	0.687
	PA2	0.810
	PA3	0.840
Perceived AI Literacy (PAIL)	PAIL1	0.882	0.895	0.923	0.708
	PAIL3	0.817
	PAIL4	0.897
	PAIL5	0.886
	PAIL6	0.709
Generative AI Usage Intention (UI)	UI1	0.838	0.766	0.865	0.681
	UI2	0.835
	UI3	0.801

demonstrates that all values for Cronbach’s alpha coefficient (α) and composite reliability (CR) surpass the 0.7 threshold, ensuring construct reliability.

The validity encompasses both convergent and discriminant aspects. Convergent validity of the construct was evaluated through item factor loading scores. Table 2 illustrates that each item’s loading on its respective construct is above 0.7, indicating adequate convergent validity. Discriminant validity was tested in four ways. First, all average variance extracted (AVE) values surpass the suggested threshold of 0.5, demonstrating sufficient convergent validity. Second,

Table 3. Discriminant validity.

	PE	EE	SI	FC	PT	PA	PAIL	UI
PE	0.758
EE	0.404	0.771
SI	0.559	0.207	0.865
FC	0.448	0.306	0.503	0.752
PT	0.172	0.255	0.291	0.313	0.816
PA	0.508	0.376	0.466	0.391	0.175	0.829
PAIL	0.384	0.285	0.471	0.434	0.222	0.502	0.841
UI	0.526	0.480	0.351	0.418	0.111	0.500	0.554	0.825

Note: Bolded diagonal values represent square roots of AVEs.

presents the Fornell–Larcker criterion results, which demonstrate that the square root values of AVE (bolded diagonal values) are greater than the correlation among the constructs (the relevant column and row values) and confirm the discriminant validity [80]. Third, all the correlations between constructs do not exceed 0.7, as shown in Table 3. Fourth, the condition for discriminant validity of these reflective constructs is satisfied, as indicated by the Heterotrait–Monotrait (HTMT) ratio values being below 0.85 [81], as shown in

Table 4. HTMT (Heterotrait–Monotrait criterion).

	PE	EE	SI	FC	PT	PA	PAIL	UI
PE
EE	0.518
SI	0.702	0.248
FC	0.595	0.391	0.647
PT	0.225	0.334	0.371	0.420
PA	0.665	0.481	0.580	0.516	0.226
PAIL	0.467	0.335	0.554	0.532	0.274	0.602
UI	0.682	0.622	0.429	0.539	0.148	0.641	0.657

.

4.4. Structural Model Analysis

We evaluated the structural model using a bootstrapping method with 10,000 resamples. Figure 2 displays the test results; 51.6% variance in the dependent variable is explained, reflecting an acceptable level of explanatory effectiveness. Since the SRMR value is 0.054, which is less than the 0.08 threshold [82], the model fit is satisfactory. Moreover, employing blindfolding with an omission distance of 7, the Stone–Geisser Q² value for the dependent variable generative AI usage intention was 0.337 (Q² > 0), confirming the model’s predictive relevance and strong out-of-sample predictive power [83]. As expected, performance expectations (H1: β = 0.245, p < 0.001), effort expectations (H2: β = 0.242, p < 0.001), and facilitating conditions (H4: β = 0.106, p < 0.05) positively affect users’ generative AI usage intention. However, the effect of social influence (H3: β = −0.085, p > 0.05) is not significant. Perceived threats (H5: β = −0.075, p < 0.05) are negatively associated with generative AI usage intention, and perceived avoidability (H6: β = 0.99, p < 0.05) positively affects users’ usage intention. Perceived AI literacy (H7: β = 0.330, p < 0.001) also positively influences the generative AI usage intention. Additionally, none of the control variables significantly affect the intention. Thus, H1-H2 and H4-H7 are supported, but H3 is not. The effect sizes of path coefficients within the research model were assessed using Cohen’s f² metric [84]. In accordance with established benchmarks, the high, medium, and small effects of paths are recognized by thresholds of 0.35, 0.15, and 0.02, respectively. Analysis of specific constructs revealed that PE (f² = 0.067), EE (f² = 0.086), and PAIL (f² = 0.134) occur with small to medium effects.

4.5. Moderation Test

To examine hypotheses regarding moderation effects, multiplicative terms were generated by cross-multiplying items related to perceived threats and perceived AI literacy, as well as perceived avoidability and perceived AI literacy. All variables were standardized prior to calculation to address multicollinearity. The findings indicate that the path coefficient of H7a (β = 0.081, p < 0.05) is significant, but H7b (β = −0.028, p > 0.05) is not. Thus, H1–H2, H4–H7, and H8b are supported, but H3 and H8a are not. Following Cohen et al. [84], simple slopes are used to graphically represent the moderating effect of perceived AI literacy. As shown in Figure 3, two regression lines are identified for the independent variable on the dependent variable, positioned one standard deviation above and below the mean of the moderating variable. When perceived AI literacy is low, the perceived threat has a negative impact on the intention to use generative AI, and this effect becomes weaker when the level of AI literacy is high, suggesting that perceived threat exerts a stronger effect on generative AI usage intention for users whose level of perceived AI literacy is low. Following Chin et al. [85], the effect size (f²) was calculated to confirm the overall moderating effects. The resultant f² value is 0.023, confirming a significant yet small-effect moderation in the proposed model [84].

4.6. ANN Analysis

In the empirical analysis’s second stage, we employed the artificial neural network (ANN) methodology, a computational model that emulates the structure of biological neural networks [86]. Figure 4 demonstrates the development of an ANN model using a multi-layer perceptron and the Sigmoid activation function. An input layer, which included six significant predictors (performance expectancy, effort expectancy, facilitating conditions, perceived threats, perceived avoidability, and perceived AI literacy) identified from the PLS-SEM model; a single hidden layer with 4 neurons; and an output layer representing the behavioral intention to use generative AI. The model was trained using the backpropagation algorithm [87], and its performance was evaluated using a 10-fold cross-validation strategy to mitigate overfitting. The data were randomly split into 90% for training and 10% for testing. As shown in

Table 5. RMSE value of ANN models (output: generative AI usage intention).

ANN	Training	Testing
1	0.067	0.055
2	0.071	0.058
3	0.066	0.054
4	0.065	0.066
5	0.066	0.062
6	0.068	0.062
7	0.068	0.045
8	0.066	0.058
9	0.066	0.071
10	0.070	0.060
Mean	0.067	0.059
Standard Deviation	0.002	0.007

, the average root mean square error (RMSE) values during the training and testing phases were 0.067 and 0.059, respectively, indicating that the ANN model demonstrated satisfactory predictive accuracy.

To ensure the robustness of our approach, ANN was selected as the optimal modeling approach. Among machine learning methods, SVM is best suited for high-dimensional classification, and decision trees are known for their interpretability in structured data. However, ANN excels at capturing complex, nonlinear relationships among variables, which are common in behavioral research [86,88,89]. This makes ANN highly complementary to PLS-SEM, which primarily models linear effects. Moreover, unlike traditional regression analysis techniques, ANNs are capable of modeling non-compensatory decision-making processes [90] and do not require assumptions of normality [91]. Given their causal nature and ability to detect nonlinear relationships, multi-layer perceptions (MLPs) have been recognized as an effective complement to PLS-SEM models [92], and widely used by previous studies [86,88,89].

Table 6. Sensitivity analysis.

Network	PE	EE	FC	PT	PA	PAIL
1	0.239	0.196	0.150	0.058	0.081	0.277
2	0.34	0.152	0.099	0.025	0.179	0.206
3	0.188	0.219	0.142	0.077	0.132	0.243
4	0.201	0.177	0.159	0.071	0.129	0.264
5	0.196	0.201	0.149	0.037	0.172	0.245
6	0.237	0.215	0.166	0.017	0.166	0.200
7	0.160	0.229	0.128	0.075	0.114	0.293
8	0.198	0.195	0.170	0.035	0.175	0.227
9	0.184	0.137	0.194	0.104	0.084	0.297
10	0.279	0.199	0.086	0.076	0.129	0.232
Average Importance	0.222	0.192	0.144	0.058	0.136	0.248
Normalized Importance (%)	89.45%	77.29%	58.09%	23.15%	54.79%	100%

illustrates the sensitivity analysis performed on the neural network, which evaluated how all independent variables predict the dependent variable. Notably, perceived AI literacy exhibited the highest relative importance (100%) in predicting usage intention, followed by performance expectancy and effort expectancy with relative importance of 89.45% and 77.29%, respectively. Ranking fourth and fifth are facilitating conditions and perceived avoidability, with relative importance of 58.09% and 54.79%, respectively. Perceived threats demonstrated the lowest relative importance at 23.15%. The predictive capability of input variables (independent variables) within the artificial neural network is quantified using normalized relative importance, an analogous metric to the path coefficient in PLS-SEM. As shown in

Table 7. Comparison between the results derived from PLS-SEM and ANN.

PLS Path	PLS-SEM: Path Coefficient	Normalized Importance	Ranking (PLS-SEM)	Ranking (ANN)	Remark
PE-UI	0.245	89.45	2	2	Match
EE-UI	0.242	77.29	3	3	Match
FC-UI	0.106	58.09	4	4	Match
PT-UI	−0.075	23.15	6	6	Match
PA-UI	0.099	54.79	5	5	Match
PAIL-UI	0.330	100.00	1	1	Match

, the importance ranking of independent variables is consistent in both the ANN model and PLS-SEM, thereby corroborating the robustness of the empirical findings from the initial stage.

4.7. Discussion of the Results

First, as for the technology acceptance constructs, performance expectancy, effort expectancy, and facilitating conditions have significant effects on generative AI usage intention, and performance expectancy is the strongest predictor among them according to the ANN analysis. This finding aligns with recent international studies [93,94], highlighting that anticipated gains in efficiency and effectiveness remain the central determinants of AI adoption across diverse settings. Although the UTAUT generally identifies social influence as a significant factor in technology adoption, our study did not find a significant effect of social influence on generative AI usage intention among Chinese enterprise employees. This finding contrasts with prior research in contexts such as educational technology adoption, where social norms and peer influence often play a decisive role [14,37,48]. One possible explanation is that generative AI remains a relatively novel and exploratory technology in many organizational contexts, so established social norms regarding its use may be weak or ambiguous [65,66,69]. In addition, the private and self-directed nature of many generative AI applications could diminish the influence of external expectations [64,95,96]. For example, in voluntary or highly self-directed contexts, social influence is not a significant predictor of behavioral intention to adopt AI-based technologies [93,95]; rather, perceived usefulness and self-efficacy are more influential. This suggests that, for innovative and personally controlled tools like generative AI, individual cognitive factors may outweigh traditional social pressures. Furthermore, cultural factors such as high institutional trust and strong performance-orientation in Chinese workplaces may lead employees to prioritize perceived usefulness and individual mastery over social endorsement [97,98]. For example, Kurniasari [50] and L. Wang & Xiao [51] have pointed out that when social influence conflicts with personal values or cognition, it may instead inhibit users’ willingness to use. Therefore, social influence cannot play a significant role in promoting employees’ generative AI usage intention among Chinese enterprises.

Second, generative AI usage intention is substantially driven by perceptions of both threat and avoidability, but they influence users in opposite directions. Perceived threats exert a negative impact on the usage intention, whereas perceived avoidability has a positive one. These results validate previous empirical evidence [56,93,94], indicating that the avoidance framework of TTAT is applicable to this context. In other words, understanding the interplay between perceived threat and perceived avoidability is essential for predicting and influencing user acceptance of AI. For example, Liang & Xue [15] found that both perceived threat severity and perceived threat susceptibility significantly reduced users’ willingness to use IT systems, while the belief in effective coping mechanisms (perceived avoidability) increased intention to use. More recently, Joo et al. [94] demonstrated that, in the context of AI-based smart services, users’ perceptions of technological threat suppressed adoption intention, but a higher sense of control or avoidability promoted it. By addressing users’ concerns and highlighting the benefits, developers and policymakers can foster greater willingness to use AI technologies.

Third, perceived AI literacy poses a notable positive impact on generative AI usage intention, which is in line with prior studies [17]. Our findings extend this work by showing that perceived AI literacy not only directly promotes usage intention but also moderates the negative effect of perceived threat. That is, when individuals feel more knowledgeable and competent about AI, they are less likely to be deterred by potential risks—a pattern echoed in recent studies that advocate for multidimensional AI literacy [93,95]. However, we did not observe a significant moderating effect of AI literacy on the link between perceived avoidability and usage intention. While perceived avoidability positively correlates with usage intention as hypothesized, AI literacy’s anticipated amplification effect was absent. An alternative explanation lies in the cognitive evaluation saturation. When users possess high baseline awareness of generative AI’s transformative benefits (e.g., efficiency gains, creative augmentation), further literacy increases may not enhance avoidability’s influence. For most adopters, perceiving avoidability sufficiently enables adoption decisions regardless of literacy level—akin to how internet users adopt search engines despite varied technical knowledge. This suggests perceived avoidability operates as a universal enabler in low-stakes contexts (e.g., routine content generation), while AI literacy’s moderating role might become critical only for high-risk or high-skill contexts.

5. Implications

5.1. Theoretical Implications

This research extends existing theory in three main ways. Firstly, this paper introduces an acceptance–avoidance framework aimed at understanding the users’ generative AI usage intention. This framework is built upon the UTAUT and the TTAT. Unlike prior work that considered acceptance and avoidance separately, our model provides a unified structure for analyzing how positive drivers and threat-based inhibitors interact to shape generative AI usage intention [15,64]. On the one hand, this study refines and expands the Integrated AI acceptance–avoidance model (IAAAM). The original IAAAM primarily focused on enterprise managers, examining how these high-level decision-makers perceive and adopt AI technologies [65]. By extending the application of this model to ordinary enterprise employees, this paper enlarges the research scope to incorporate a broader range of organizational stakeholders. On the other hand, this paper enhances the content related to technology threat avoidance. Based on the IAAAM model, this paper introduces perceived avoidance (PA) in addition to the original perceived threat (PT) to more comprehensively address the influencing variables related to users’ technology avoidance [56].

Secondly, this research makes a distinct theoretical advancement by integrating perceived AI literacy into established UTAUT and TTAT frameworks, offering a more holistic and cognitively grounded model of technology adoption. While most prior studies have focused primarily on motivational or threat-related factors, our approach recognizes cognitive readiness—specifically, perceived AI literacy—as essential for effective adoption and human–machine collaboration. This is especially relevant in the context of Industry 5.0, which prioritizes human-centric, synergistic integration between employees and intelligent technologies [3]. Furthermore, by empirically validating the model in China—a setting characterized by accelerated digital transformation, strong institutional trust, and performance orientation [95,98]—our work demonstrates how contextual and cultural factors shape the pathways to AI adoption. These contributions move beyond previous AI adoption models by addressing both cognitive and environmental prerequisites for responsible, widespread generative AI use.

Thirdly, this paper introduces perceived AI literacy as a moderating variable between perceived threat and generative AI usage intention. Unlike traditional TTAT constructs, perceived AI literacy reflects users’ cognitive capacity to understand and manage AI systems, which can buffer avoidance responses and promote adoption [36]. This integration contributes to a more cognitively grounded and context-sensitive model of generative AI acceptance under the human-centric vision of Industry 5.0, broadening the boundaries of studies in the generative AI usage intention field. Moreover, the current research further broadens the theoretical applicability of UTAUT and TTAT by shifting the focus from the well-studied educational context to underexplored sectors such as manufacturing and service industries. While prior studies, such as Xia and Chen [13], Yakubu et al. [14], and Zhou et al. [12] have predominantly investigated generative AI adoption among students and educators, this study tests these frameworks in a broad operational context, such as manufacturing, services, financial, and IT industries.

Finally, this study contributes to the cultural generalizability of technology adoption research. This study investigation covers major industries with intensive AI application—namely, information technology, manufacturing, services, finance, and education—based on survey data from a representative provincial capital in western China. While mainstream models such as UTAUT and TTAT have been widely validated in Western and educational contexts, our findings reveal that key mechanisms, such as the non-significant role of social influence, are shaped by distinctive features of the Chinese organizational environment, including high institutional trust and strong performance orientation [16,34]. This highlights the importance of organizational signals and individual cognitive resources over peer influence in technology adoption decisions. However, given the focus on a specific region and cultural setting, caution is required when applying our results to other contexts. Cultural, institutional, and regional differences may moderate the observed relationships, and the generalizability of our findings should be further tested in other national or cross-cultural settings [95]. Therefore, our study underscores the necessity of adapting global technology acceptance frameworks to local environments and encourages future research to systematically examine cultural and industry-specific boundary conditions.

5.2. Managerial Implications

This study offers practical management insights from a cost-benefit perspective. Specifically, the fundamental constructs of the UTAUT—namely, performance expectations, effort expectations, and facilitating conditions—are interpreted as perceived expected benefits by employees. These constructs positively influence employees’ willingness to adopt generative AI. Conversely, constructs derived from the TTAT—such as perceived threat and perceived avoidability—represent the psychological and operational costs associated with technology use, including security risks and cognitive load. Perceived AI literacy serves as a crucial moderating factor between these two sets of constructs, enhancing the conversion of perceived benefits while mitigating perceived costs. Consequently, in the context of promoting sustainable human–machine collaboration, management strategies should prioritize the expansion of perceived benefits, the reduction of perceived costs, and the achievement of equilibrium between the two through the systematic development of AI literacy. Based on the research findings, this paper proposes the following three recommendations:

First, with the aim of maximizing perceived benefits, organizations should prioritize cultivating employees’ belief in AI’s potential to aid in achieving their objectives and their confidence in mastering AI technology, as these are essential drivers of technology adoption. Practically, proficiency in AI applications should be directly integrated into the performance reward system. Additionally, targeted training programs should be developed to visually illustrate how AI can alleviate repetitive tasks, thereby allowing employees to concentrate on higher-value activities such as creative thinking.

Furthermore, although social influence has not demonstrated a significant direct impact, adequate facilitating conditions are crucial for a successful transformation. Organizations should move beyond formal commitments by making substantial investments, such as acquiring advanced AI infrastructure, establishing dedicated technical support teams, providing on-site guidance from experts, and allocating resources for employee innovation and research and development. These measures will help reduce adaptation challenges by ensuring a reliable supply of resources and promoting sustainable human–machine collaboration. The government can incentivize enterprises to increase relevant investments through tax incentives, while simultaneously coordinating the development of regional artificial intelligence public service platforms. These platforms would provide shared infrastructure and technical support for small and medium-sized enterprises, thereby lowering the barriers for industry-wide transformation and fostering a sustainable ecosystem of human–machine collaboration.

Second, with the aim of minimizing perceived costs, companies should focus on addressing employees’ concerns regarding the potential risks associated with AI. This can be achieved by enhancing employees’ confidence in AI risk management through transparent communication, practical safeguards, and efficient feedback mechanisms. Training programs should clearly articulate the data governance mechanisms integrated into AI systems and the risk mitigation solutions implemented by operators. Furthermore, establishing a rapid-response communication channel is crucial for incorporating employees’ concerns into the system optimization process, thereby helping employees effectively navigate technical uncertainties.

Third, understanding perceived AI literacy holds dual strategic significance; it directly facilitates technology adoption while simultaneously mitigating employees’ resistance to perceived threats. To optimize these benefits, organizations should develop a progressive learning trajectory that begins with foundational operational skills and advances towards innovative applications. Implementing a weekly expert consultation session can offer a readily accessible troubleshooting guide during working hours. Additionally, establishing incentive mechanisms to clearly reward the implementation of results and the enhancement of skills is crucial. As employees advance their competencies through practice, AI tools should consistently be framed as supportive aids that augment human decision-making, rather than as replacements.

6. Limitations and Future Research

This research recognizes several limitations that indicate potential directions. First, data sourced exclusively from China established cultural boundary conditions. As a result, caution is advised when extrapolating these conclusions to other countries or regions, as cultural, social, and technological variations might greatly affect AI usage willingness and behavior. Future investigations could encompass a range of international samples, aiming to boost the external validity of the findings and explore potential cross-cultural differences in AI adoption.

Second, this study focuses on measuring AI usage intention. While intent is an important predictor of behavior, it does not directly capture actual usage behavior. Future research should employ alternative methodologies, such as longitudinal designs, field experiments, or objective usage data collection (e.g., system logs), to capture how identified factors translate into sustained adoption and usage patterns. This approach would address the intent-behavior gap while generating actionable insights for practitioners.

Third, the measurement of AI literacy in this study uses a shorter version of the scale. While it provides some insight into the role of AI literacy, future studies could investigate different dimensions of AI literacy in more detail. By developing and validating more comprehensive measures of AI literacy, researchers can better understand how specific aspects of AI-related knowledge, skills, and attitudes affect users’ interactions with AI technologies. This may involve differentiating between technical and conceptual AI knowledge, as well as examining how different levels of AI literacy interact with other factors identified in this study.