Through the innovation of the education pattern, teaching paradigm, task processing, and collaborative communication, AI heralds the imminent era of AI lifelong learning, empowered by intelligent education (Rawas, 2024). The mode of task processing and knowledge creation, which involves the interactive and collaborative use of individual and intelligent technologies, has had a disruptive and challenging impact on the education industry (Xia et al., 2024). The ease of use, efficiency, and complexity of AI algorithms have led to significant and profound changes in predicting students’ learning states, recommending learning resources, and evaluating learning outcomes in higher education (Ouyang et al., 2022). Alqahtani et al. (2023) point out that AI technologies significantly drive innovative developments in the field of education, including AI technology to provide intelligent support, constructive feedback, automated assessment, customized courses, and personalized career guidance. Susnjak (2022) further posits that AI is trained on a large dataset of human conversations, confirming that AI can successfully solve complex problems across various domains, including education. Jaboob et al. (2025) proposed in a recent study on students’ AI behavior in Arab Higher Education that although the integration of AI technology with education is still in its first stage, students’ interaction with AI applications, perception of use, and enjoyment of pleasure all contributed to students’ high satisfaction. It advocates the establishment of intelligent education tools to help students collaborate efficiently.

However, in the current paradigm of intelligent education, the emphasis is predominantly on the convenience, support, and innovative technological contributions that AI brings to education, often overlooking academic integrity, technical ethics, educational equity, and issues related to controlling imbalances and declines in higher-order cognitive skills (Cotton et al., 2023). Therefore, technical regulation and risk control are not keeping pace with the innovative development and evolving needs of holistic smart education (Farrokhnia et al., 2024). Furthermore, the neglect of individual students’ internal motivation and their perception of external stimuli when using AI technology, the inaccurate assessment of students’ personal characteristics, and the lack of immediate oversight of their intelligent behavior and feedback on their willingness, all lead to the misuse of AI technology and the emergence of disorderly conduct (Liu, 2024). Soliman et al. (2025) pointed out that cognitive monitoring and behavioral assessment following the use of AI have become key indicators of whether students can develop healthy AI habits and become intelligent citizens. Currie (2023) points out that students without technological supervision are prone to learning dependency, which harms their professionalism, moral sense, academic integrity, and can even lead to technology addiction. Ultimately, this affects learning efficiency and task processing. Sun and Hoelscher (2024) argue that students’ overreliance on smart technologies can lead to a decline in higher-order knowledge skills, such as creativity, critical reasoning, and problem-solving abilities. This can result in a decrease in the use of smart technologies, and ultimately, learning burnout leads to a decline in the quality of academic achievement.

AI, as an intelligent computing program designed for thinking and simulating human behavior, is widely embedded in the field of education in various forms. The interactive dependency relationship between humans and intelligent systems is undergoing profound changes. Previous studies have shown that AI technology consistently enhances students’ perceived satisfaction and provides instant gratification through functions such as algorithmic recommendations, timely feedback, and automatic responses, ultimately fostering a reliance on AI (Xie et al., 2023). This overdependence behavior leads to a decline in creativity and situational understanding, the quality of knowledge storage, and the generation of “illusory truth” and the quantitative proliferation of “Fantasy Index” (Izak et al., 2025). Recent findings from Yankouskaya et al. (2025) point out that characteristics such as students’ personalized responses, affective validation, and sustained engagement may be attracted by GenAI’s claims of ability to increase productivity and efficiency, and suggest that students may be more likely to engage with a particular group of students, exacerbating the formation of their own AI dependence. This kind of AI dependence may seem like a “way to solve complex problems in the short term, but it is actually a hidden danger of functional dependence disorder,” a phenomenon known as erosion. This can lead to a practical disconnect between students and real-life learning and ultimately to learning burnout (Chakraborty et al., 2024).

Dependence on AI over the long term not only diminishes individuals’ intrinsic motivation to learn and their metacognitive abilities but also fosters emotional burnout, hedonic behavior, and lowers academic efficacy and expectations, leading to the development of AI addictive behaviors (Kooli et al., 2025). Dubey et al. (2024) proposed that the addictive behavior of AI can erode an individual’s basic abilities, such as independent reasoning, memory retention, and cognitive engagement, and can have a significant negative impact on their own productivity and values over the long term. Extensive empirical research has established behavioral and psychological patterns associated with intelligence addiction. Chou and Hsiao (2000) and Mitropoulou (2024) pointed out that addiction to AI increases individual loneliness, impairs learning functions, and weakens emotional perception and task satisfaction, leading to the generation of negative emotions such as avoidance. The proposal suggests that addiction to AI impedes the connection and sharing of knowledge, diminishes the likelihood of embodied communicative learning among groups, and even impacts knowledge sharing at the educational organizational level (Retkowsky et al., 2024). Ultimately, individuals are compelled to create fragmented knowledge islands. Hassan and Barber (2021) further point out that AI addiction leads to an “illusory truth effect” individual perceptions that, due to excessive addiction to technology, tend to ignore the authenticity of information and be deceived by repeated, meaningless assertions. It appears that at all levels, AI dependence and addiction have negative effects on personal information acquisition and knowledge quality (Retkowsky et al., 2024).

Learning burnout primarily refers to emotional exhaustion, depersonalization, and a diminished sense of accomplishment, which are caused by curriculum pressure, task complexity, or other psychological factors associated with learning. Learning burnout also encompasses persistent negative emotions that arise from individuals experiencing low educational achievement and well-being due to prolonged exposure to inappropriate learning behaviors (Zhang C. H. et al., 2021). Students with lower academic self-efficacy, career expectations, and learning motivation tend to show higher levels of learning burnout (Restubog et al., 2010). At the same time, students with learning burnout can not meet their needs through academic task processing, and their sense of academic achievement, self-efficacy, and mental health are lower (Lin and Huang, 2012). Therefore, individuals with higher levels of learning burnout tend to be more eager to use technology to obtain instant gratification. If not controlled, they may be more likely to use technology to obtain instant gratification, the “Compensation Hypothesis,” instead leads to increased learning burnout (Toth et al., 2021).

Burnout has become one of innovative education’s most significant challenges because of its corrosive effect on learning autonomy (Ahmad N. et al., 2023). At the internal psychological level, students lack self-control, emotion regulation, and reflection mechanisms after perceiving the practicality and ease of use of AI tools, and it is easy to fall into a vicious cycle of inertia, emotional exhaustion, and behavioral burnout (Ding, 2021). At the level of the external stimulation mechanism, learning burnout in intelligent education is a process of emotional exhaustion and psychological imbalance mediated by AI dependence and addictive behavior. Ding et al. (2025) found that AI dependence significantly enhances individual emotional exhaustion, which hinders innovative motivation and practical behavior, ultimately leading to individual industry burnout.

Based on pertinent research concerning AI dependence, AI addiction, and learning burnout, it has been established that individuals’ characteristics and external stimuli following the use of AI technology can exacerbate AI dependence and addictive behavior, thereby leading to the consequences of learning burnout. However, the majority of studies primarily concentrate on verifying the correlation between these influencing factors, lacking a comprehensive theoretical framework and model construction from a holistic perspective. Furthermore, the quantity of related studies is limited, and the research methodologies are relatively simplistic, typically centering on a singular linear genetic mechanism. Considering the limited number of studies with a comprehensive theoretical framework that encompasses students’ personal characteristics, technological engagement, reliance on AI, AI addiction, and learning burnout in higher education. The aim of this study is to examine the critical factors that contribute to the development of AI dependence and addictive behaviors following students’ use of AI technology. Additionally, the study seeks to understand how such dependence and addiction ultimately result in the manifestation of learning burnout among students. The primary research questions of this study are as follows:

Prior studies have indicated that the I-PACE framework requires ongoing refinement to remain consistent with new technological contexts and evolving patterns of use (Brand et al., 2019). Subsequent research has extended the model by offering a more delineated account of the processes underlying technology-related addictive behaviors. For instance, Dempsey et al. (2019) highlight that affective experiences of enjoyment and cognitive appraisals of technology are closely associated with addiction-related outcomes. In addition, susceptibility characteristics triggered by external technological stimuli have been shown to facilitate dependency-oriented behaviors, which may, in more extreme cases, develop into addiction-like patterns of use (Servidio, 2021). Evidence from personality research further suggests that inert thinking, as an individual disposition, is positively associated with technology dependence and addiction-related behaviors (Shiner, 2018). Studies focusing on perceived usefulness similarly indicate that it is linked not only to technology dependence but also to a subsequent shift toward more addiction-like use (Elhai et al., 2020). Nevertheless, empirical applications of the I-PACE model to the context of AI technologies remain limited. Recent studies in higher education have begun to address this gap by using the I-PACE framework to examine how students’ personal attributes and contextual factors shape AI dependence and AI addiction (Zhong et al., 2024). These findings provide initial support for the model’s relevance to AI-related dependency processes, while underscoring the need for further investigation.

On this basis, it is necessary to recognize that adoption-oriented and compensation-oriented models exhibit certain theoretical limitations in addressing these dimensions. The Compensatory Internet Use Theory and the Integrative Pathways Model primarily explain why individuals use technology to regulate emotions or fulfill psychological needs, whereas the Technology Acceptance Model and the Unified Theory of Acceptance and Use of Technology mainly account for the formation of usage intentions and continued use. Although these models emphasize motivational, attitudinal, and performance-related evaluations, they offer limited capacity to capture the progressive loss-of-control processes that may emerge when technology use becomes increasingly pathological, and they provide insufficient explanation of addiction-specific mechanisms such as executive-function impairment, compulsive escalation, and declining behavioral control. Consequently, they are more suitable for examining technology adoption, usage preferences, or sustained use than for explaining the developmental trajectory from normative use toward addiction and learning burnout (Ge et al., 2023). Against this background, and in view of the present study’s focus on how AI dependence may develop into AI addiction and how this process may be associated with learning burnout, the I-PACE model provides clearer addiction specificity, a stronger developmental orientation, and greater explanatory strength at the mechanistic level than CIUT, IPM, TAM, and UTAUT. Its application in describing and analyzing the antecedents and developmental processes of technology-related addictive behaviors is therefore theoretically appropriate and methodologically justified (Xiong et al., 2024).

In addition, beyond examining how key individual attributes in the I-PACE model—such as inert thinking, perceived enjoyment, and perceived usefulness—are associated with the development of AI dependence and AI addiction, the present study further considers how addiction-related processes may correspond to deeper behavioral outcomes, particularly learning burnout. Learning burnout is understood as a state in which chronic dependence and technology-related addictive tendencies are associated with fatigue, reduced academic fulfilment and achievement, and a gradual shift toward avoidance-oriented behaviors. Accordingly, this study extends the I-PACE framework by incorporating learning burnout as an additional outcome variable, thereby not only focusing on the affective–cognitive–executive pathway, but also examining the tangible consequences that may arise from AI dependence and AI addiction.

Utilizing a cross-sectional survey design, this study investigates the relationship between Chinese college students’ perceived usefulness of AI, inert thinking, perceived enjoyment of AI, AI dependence, AI addiction, and learning burnout. This study first determined the minimum sample size using the G * Power Tool (Faul et al., 2009), with the G * Power parameters set to: a moderate effect size of 0.15, an error type of 0.05, an effect strength of 0.8, and 5 predictors, with the G * Power parameter set to 0.8, the minimum sample size that makes the study credible is 92. The primary sampling method is stratified random sampling, and the stratification is mainly according to the subject’s specialty and age to ensure that students are represented in the research-related groups.

This study primarily employs the Questionnaire Star platform to conduct a structured questionnaire survey. The objective is to capture a multi-dimensional structure based on established theories, encompassing relevant variables such as the I-PACE model, AI dependence, and addiction. This study distributed the questionnaire via online media platforms, including Questionnaire Star, WeChat, and QQ, as well as through offline direct distribution. The data collection period spanned from June to September, 2025, ensuring accessibility across varying literacy levels and technical proficiencies. All participants who conducted the study provided informed consent upon completing the questionnaire. A total of 427 copies of the survey were distributed. Invalid responses were determined and subsequently excluded based on the established criteria: (1) key items were missing or not answered in the questionnaire, (2) the response time was less than the average frequency (i.e., less than 2 min). Ultimately, 15 invalid questionnaires were removed, and the response rate was 96.4%. Consequently, 412 valid questionnaires were obtained, and the total number of research samples was N = 412. The obtained data were imported into SPSS 26 software for demographic analysis. SmartPLS (V4.1) was utilized to evaluate the correlation between the models proposed in this study and the potential relationships of the individual variables. The analysis was conducted using SmartPLS software (see Supplementary material 1 for SmartPLS output), and the dataset was collected through a structured questionnaire (see Supplementary material 2 for the dataset).

The participants in this study are college students from higher education institutions in Fujian Province, China. They were recruited and screened using random sampling. The gender distribution of the sample is as follows: male students account for 21.8%, or 90 individuals, while female students account for 78%, or 322 individuals. In terms of educational distribution, undergraduate students make up the largest proportion at 93.7%, with a specific count of 386 individuals. The number of professionals in the distribution includes film and television, digital media, journalism and communication, with respective counts of 212, 83, and 116 individuals. Additional demographic details are presented in Table 1. This study was approved by the Ethics Committee of the School of Film and Communication, Xiamen University of Technology (Approval Number: XUT-SFC-2025-01, Date: January 9, 2025). All participants provided written informed consent in accordance with the Declaration of Helsinki.

First, a conceptual and semantic equivalence review of the original scales was conducted to ensure that the core constructs were contextually appropriate for Chinese university students. Two doctoral students with expertise in psychometrics and educational technology produced the initial translation, and another graduate student conducted a back-translation following Beaton et al. (2000) method to assess semantic consistency and potential conceptual deviation. Second, the translation and back-translation versions were reviewed by two associate professors in AI education, psychometrics, and behavioral addiction research to evaluate content validity, focusing on linguistic clarity, cultural appropriateness, and potential item bias. Third, a pilot test (N = 50) was conducted using the revised scales. Participant interviews and feedback were used to identify misunderstandings, ambiguous items, and semantic load issues, which informed minor revisions to improve comprehensibility and contextual relevance. Finally, during the formal survey, the reliability and structural validity of the scales were re-examined to confirm measurement stability and structural consistency following cross-cultural adaptation. Through this multi-stage procedure, the study preserved the conceptual integrity of the original scales while achieving contextual adaptation and measurement equivalence in the Chinese AI-mediated learning environment, thereby enhancing the methodological rigor and cross-cultural validity of the instruments.

This study mainly uses SmartPLS for convergent validity analysis to verify whether there are significant consistency problems within the scale constructed in this study to evaluate the model’s reliability. In measuring the model’s convergent and discriminant validity, we mainly draw on Cheng and Tsai (2020) and Habibi et al. (2020). The results of the multi-dimensional data analysis clearly show that the measurement model has good discriminant validity. Table 3 also shows Cronbach’s alpha, structural reliability, and extracted mean variance values for each measurement item in this study. The model is proven to have good convergent validity when the AVE value of all indicators is considered ideal when greater than 0.5 (Li, 2024). In this study, the lowest AVE value is 0.639 (0.5), so the measurement model proposed shows good convergent validity. In addition, internal consistency versus reliability was used to assess the consistency of all index results, where the values of CR (compliance reliability) and CA (Cronbach’s alpha) should be consistent with [0,1] and > 0.7 (Hair et al., 2021a), the CR and Ca values in this study were both higher than the recommended value of 0.7, thus the measurement model reliability was found to be somewhat persuasive.

This study also applies the Fornell-Larcker test in PLS-SEM (Fornell and Larcker, 1981) to further verify the discriminant validity of the hypothetical model. Table 4 shows that the AVE square root of each latent variable is greater than the correlation coefficient with other constructs, which further confirms the discriminant validity of the research model.

In addition, the value of variance inflation factor (VIF) needs to be evaluated before structural equation modeling analysis to further verify whether there is a potential collinearity problem between the individual measurement items. When VIF < 3, it can be further proved that there is no collinearity in the model (Hair et al., 2009). According to Table 5, the range of VIF values in this study was [1.307, 2.968], and the VIF was less than 3 for all variables in this study. It is demonstrated again that there is no significant collinearity between the variables of the structural equation model constructed in this study, and there is no potential problem such as distortion or inaccuracy of the model.

Based on the results of SEM-PLS operation, it is proved that the structural equation model used in this paper does not have the problem of deviation. In addition, common value numbers such as RMS, NFI, and normalized value residual (SRMR) were further adopted as PLS-SEM indicators to evaluate the fitting adequacy of the model. SRMR is usually in the numerical interval of [0,1], and when SRMR < 1.00, it can be demonstrated that the model fits well (Wijaya et al., 2022). The SRMR of the structural equation model was 0.072 < 1.00, which proved that the model fitting effect was sufficient. Moreover, Demler et al. (2015) noted that NFI reflects the degree of correspondence between the observed variables and the assumed model, with closer to 1.00 indicating better model fit. In this study, the NFI value is 0.821 > 0.8, which is within the acceptable range and further proves the adequacy of the model fitting.

Using SmartPLS (V4.1) software to test the direct effect, indirect effect, and overall effect of the structural equation model established in this study, Figure 2 shows this study’s validated structural equation model, and the correlation coefficient and significance level of each path.

Table 6 shows the variance, confidence interval, t-value, and p-value (significance) for each path calculated using the SmartPLS bootstrap method, with a total of 5,000 resamples. Through the structural equation model analysis, this paper clarifies how college students’ perceived usefulness, perceived enjoyment, and inert thinking lead to individual AI dependence and addiction, even though AI dependence and AI addiction as a mediating path ultimately lead to self-learning burnout behavior.

Table 6 records the variance, confidence interval, t-value, and p-value (significance) for each model path calculated using the PLS-SEM Bootstrap and repeated 5,000 times. Therefore, using structural equation modeling for empirical analysis, we can directly and significantly observe the relationship between college students’ perceived usefulness, perceived enjoyment, inert thinking, AI dependence, AI addiction, and learning burnout. Data analysis shows that the path coefficient between perceived usefulness and AI dependence is p = 0.001, p < 0.01 (SD = 0.06, T = 3.365), indicating that perceived usefulness positively impacts AI dependence. At the same time, the path coefficient between perceived usefulness and AI addiction is p = 0.00, p < 0.001 (SD = 0.05, T = 8.246), indicating that perceived usefulness also has a significant positive impact on AI addiction. The path coefficient between inert thinking and AI dependence was p = 0.0, p < 0.001 (SD = 0.051, T = 7.054), indicating that inert thinking significantly impacts AI dependence. However, the path coefficient between inert thinking and AI addiction is p = 0.751, p > 0.05 (SD = 0.049, T = 0.317), indicating no significant correlation between inert thinking and AI addiction. The path coefficient between perceived enjoyment and AI dependence was p = 0.001, p < 0.01 (SD = 0.068, T = 3.375), indicating that perceived enjoyment positively affects AI dependence. At the same time, the path coefficient between perceived enjoyment and AI addiction is p = 0.00, p < 0.001 (SD = 0.053, T = 7.157), indicating that perceived enjoyment also has a significant positive impact on AI addiction. The path coefficient between AI dependence and AI addiction was p = 0.00, p < 0.001 (SD = 0.044, T = 3.521), indicating that AI dependence significantly impacts AI addiction. AI dependence has a significant positive impact on AI addiction; it also shows that AI dependence has the possibility of further evolving into AI addiction. In addition, the path coefficient between AI dependence and learning burnout is p = 0.00, p < 0.001 (SD = 0.046, T = 8.724), indicating no significant correlation between AI dependence and learning burnout. The path coefficient between AI addiction and learning burnout is p = 0.00, p < 0.001 (SD = 0.049, T = 9.471), indicating no significant correlation between AI addiction and learning burnout. Therefore, the hypotheses proposed in this study, H1, H2, H3, H5, H6, H7, H8 8 and H10, are effectively supported, while the hypothesis H4 is not (see Table 7).

The path coefficient between perceived usefulness and learning burnout was p = 0.00, p < 0.001 (SD = 0.043, T = 6.666), indicating that students’ perceived usefulness of AI tools will also significantly impact learning burnout. The path coefficient between inert thinking and learning burnout is p = 0.00, p < 0.001 (SD = 0.037, T = 4.906), indicating that inert thinking cultivated by students using AI tools will also have a significant positive impact on learning burnout. The path coefficient between perceived enjoyment and learning burnout was p = 0.00, p < 0.001 (SD = 0.037, T = 4.906), indicating that students’ perceived enjoyment obtained by using AI tools will significantly impact learning burnout. At the same time, among the three paths that combine AI dependence as an intermediary variable: in the path of perceived usefulness and learning burnout, AI dependence as an intermediary path, path coefficient p = 0.003, p < 0.01 (SD = 0.027, T = 3.109), indicating that students’ perceived usefulness of AI tools ultimately leads to the formation of learning burnout through AI dependence. AI dependence as a mediating path in the path of inert thinking and learning burnout, with a path coefficient p = 0.00, p < 0.001 (SD = 0.028, T = 5.265), indicates that students’ inert thinking developed by using AI tools will eventually lead to the formation of learning burnout through the intermediary of AI dependence. AI dependence as a mediating path in the path of perceived enjoyment and learning burnout, the path coefficient p = 0.002, p < 0.01 (SD = 0.03, T = 3.111), shows that the perception and enjoyment of students using AI tools will eventually form the behavior of learning burnout through AI. In addition, attention was paid to three paths in which AI addiction exists as a mediating variable: in the path of perceived usefulness and learning burnout, AI addiction as a mediating path, path coefficient p = 0.00, p < 0.001 (SD = 0.034, T = 5.606), and learning burnout as a mediating variable, It shows that students’ perceived usefulness of AI tools will eventually lead to the formation of learning burnout through AI addiction. In the path of inert thinking and learning burnout, AI addiction acts as a mediating path, path coefficient p = 0.754, p > 0.05 (SD = 0.023, T = 0.314), which shows that AI addiction will not become an intermediate path between students’ inert thinking and learning burnout developed by using AI tools. AI addiction as a mediating path in the path of perceived enjoyment and learning burnout, the path coefficient p = 0.00, p < 0.001 (SD = 0.029, T = 6.097), shows that the perception and enjoyment of students using AI tools will eventually form learning burnout through AI addiction. Therefore, the hypothesis H9 proposed in this study is fully supported, and the hypothesis H11 is partially supported. AI addiction becomes an intermediate path between perceived usefulness, perceived enjoyment, and learning burnout; however, it has not become the intermediary path between inert thinking and learning burnout.

Further analysis shows that AI dependence and addiction can produce a significant serial mediating effect in the path of perceived usefulness and learning burnout, and the path coefficient is p = 0.00, p < 0.001 (SD = 0.006, T = 2.233). AI dependence and addiction can produce a significant serial mediating effect in the path of inert thinking and learning burnout, and the path coefficient is p = 0.00, p < 0.001 (SD = 0.037, T = 4.906). AI dependence and addiction can produce a significant serial mediating effect in the path of perceived enjoyment and learning burnout, and the path coefficient is p = 0.007, p < 0.01 (SD = 0.006, T = 2.685). Therefore, the hypotheses proposed in this study, H12, H13, and H14, are all effectively supported (see Table 8).

The results indicate that perceived usefulness is positively associated with AI dependence, providing empirical support for H1, which is also consistent with prior findings (Holtbrügge et al., 2025). The evidence suggests that when students perceive AI tools as effective in supporting personalized learning, reducing information barriers, and facilitating integrated access to resources, they are more likely to rely on such tools as a convenient and functional option for academic activities (Luckin and Holmes, 2016). Similarly, H2 is supported, showing that stronger perceptions of usefulness are also associated with higher levels of AI addiction. This result aligns with previous studies reporting that perceived functional benefits may reinforce reliance patterns that gradually shift from instrumental dependence to deeper psychological attachment (Acosta-Enriquez et al., 2025). Under such conditions, continued reliance on AI tools may coincide with craving-like tendencies, difficulty in disengagement, and persistent compulsive use behaviors that resemble technology-related addictive patterns (Tate et al., 2023).

The results provide support for H3, indicating a significant positive association between inert thinking and AI dependence. Students with higher levels of inert thinking are more likely to develop reliance-oriented patterns of AI use. This finding is consistent with the perspective of “cognitive offloading,” which suggests that when cognitive tasks are externalized to technological systems, reflective and analytic processing tends to weaken, thereby reinforcing functional dependence on technology (Risko and Gilbert, 2016). Similar evidence shows that when information acquisition is persistently delegated to technological tools, deeper cognitive processing and autonomous memory engagement may decline, increasing the likelihood of reliance-oriented usage tendencies (Sparrow et al., 2011). By contrast, H4 is not supported, suggesting that inert thinking alone does not directly predict ai addiction. This result is consistent with prior research indicating that addictive patterns are more often associated with emotional attachment, avoidance motivation, or compensatory use, whereas inert thinking primarily reflects reduced cognitive effort and perceived efficacy limitations rather than compulsive escalation mechanisms (Huang S. et al., 2024; Ahmad S. F. et al., 2023). In addition, the transition from dependence to addiction may be further shaped by contextual and institutional conditions, such as resource availability, technological support, and environmental or organizational characteristics (Hazzan-Bishara et al., 2025). Taken together, these findings suggest that inert thinking is more strongly associated with dependence-oriented use than with uncontrollable addictive behavior.

The results support H5, indicating a significant positive association between perceived enjoyment and ai dependence. Students who experience affective pleasure and emotional gratification during the use of AI tools are more likely to develop frequent and sustained reliance-oriented usage patterns. This finding is consistent with hedonic motivation theory, which suggests that affective gratification exerts a stronger influence on continuance behavior and usage persistence than purely instrumental utility (Lee et al., 2013). In academic contexts, AI tools may help relieve learning pressure and enhance perceived self-efficacy and achievement, thereby strengthening affective reward expectations and reinforcing dependence-oriented behavioral tendencies (Weger et al., 2022). H6 is likewise supported, showing that perceived enjoyment is not only associated with dependence-oriented use but may also be linked to higher levels of ai addiction. When pleasurable experiences gradually shift from short-term affective gratification to longer-term psychological attachment, students’ motivations for use may move from efficiency-oriented to avoidance- or compensation-oriented patterns, thereby increasing the risk of addictive tendencies (Yuan et al., 2024). Consistent with prior evidence, avoidance-driven motivations are more likely than purely instrumental motives to facilitate the escalation of technology use toward problematic or addictive trajectories (Meng et al., 2020). Furthermore, support for H7 reveals a progressive, stage-like relationship between AI dependence and AI addiction, suggesting that dependence-oriented use may constitute a precursor condition that, under the reinforcement of immediate feedback and immersive experience, gradually evolves into uncontrolled usage patterns (Maral et al., 2025). This result is consistent with behavioral addiction research highlighting a developmental pathway from reliance to habit formation and compulsive engagement, underscoring the central role of affective experience and habitual processing in the deepening of dependence (Jo and Baek, 2023).

The results provide support for H8, indicating a significant positive association between AI dependence and learning burnout. This finding suggests that when students increasingly externalize cognitive processing and task regulation to AI tools, cognitive engagement may gradually be replaced by “cognitive outsourcing,” accompanied by reduced self-efficacy and weakened performance expectations. These processes may, in turn, intensify emotional exhaustion and learning fatigue, thereby contributing to higher levels of learning burnout (Tarafdar et al., 2020). Furthermore, the support for H9 indicates that AI dependence plays a significant mediating role in the relationships among perceived usefulness, inert thinking, perceived enjoyment, and learning burnout. Specifically, positive evaluations of the functional and affective value of AI tools are associated with stronger reliance-oriented usage tendencies, while such dependence may be accompanied by diminished autonomous learning ability, reflective thinking, and cognitive investment, which together contribute to the gradual accumulation of burnout risk (Ma, 2024; Liu et al., 2023). AI dependence is not only a behavioral outcome of technology use, but also a key transmission mechanism linking individual technology perceptions to learning burnout. This provides empirical support for a developmental process in AI-mediated learning contexts, whereby reliance-oriented use is associated with psychological exhaustion and, ultimately, learning burnout (Huang C. et al., 2024).

Building on these findings, the results further indicate that AI dependence also plays a significant mediating role in the relationship between inert thinking and learning burnout. This suggests that students with higher levels of inert thinking are more likely to outsource information processing and decision-making to intelligent tools, thereby reducing deep thinking and active engagement. Over time, reliance-oriented use may gradually replace autonomous participation, weakening perceived control and competence and, in turn, contributing to the cumulative risk of burnout, which is consistent with evidence on the pathway from dependent use to reduced engagement and subsequent burnout formation (Maslach and Leiter, 2016). In addition, the findings show that AI dependence exerts a similar mediating effect in the relationship between perceived enjoyment and learning burnout. Pleasure-driven, high-frequency use may reinforce affective attachment and shift limited personal resources away from learning tasks toward immediate feedback and emotional reward, leading to sustained attentional and cognitive resource depletion and, ultimately, to burnout-related experiences (Hobfoll et al., 2018). Taken together, AI dependence functions as a key mediating mechanism linking individual psychological characteristics to learning burnout, revealing a unidirectional and progressive pathway of “resource depletion – reduced engagement – deepened burnout” in AI-mediated learning contexts (Fan et al., 2024).

The results provide support for H10, indicating that AI addiction is significantly and positively associated with learning burnout. This finding is consistent with prior evidence showing that addictive patterns of technology use are often accompanied by sustained attentional distraction and a preference for immediate gratification, which, through repeated cycles of use, may weaken perceived learning control and competence, heighten stress and feelings of loss of control, and ultimately contribute to avoidance tendencies and burnout responses (Yang et al., 2024). Related research also suggests that addictive use can trigger cognitive conflict and the accumulation of negative emotions, gradually leading individuals toward academic burnout through continued use, self-doubt, and functional depletion (Ma et al., 2025). The results further provide partial support for H11, showing that AI addiction plays a significant mediating role in the relationships among perceived usefulness, perceived enjoyment, and learning burnout, whereas no significant mediating effect is observed between inert thinking and learning burnout. In particular, perceived enjoyment influences learning burnout indirectly through AI addiction, indicating that when students become highly reliant on AI’s efficiency-enhancing and task-substitution functions, they are more likely to develop compulsive and persistent usage patterns that undermine learning autonomy and intrinsic motivation, consistent with evidence on the shift from tool-oriented dependence to loss-of-control behaviors (Yankouskaya et al., 2025). In addition, sustained addictive use may reinforce preferences for immediate feedback and promote passive coping tendencies in challenging learning situations, which can reduce motivational persistence and cognitive engagement, thereby increasing the risk of learning burnout (Marriott and Pitardi, 2024).

The results further indicate that perceived enjoyment influences learning burnout indirectly through AI addiction, confirming that AI addiction functions as a significant mediating mechanism between perceived enjoyment and learning burnout. This finding aligns with prior research showing that, within a self-determination perspective, perceived enjoyment operates as an intrinsic motivational driver that strengthens affective attachment and internalized regulation, thereby increasing the likelihood of compulsive and persistent technology use (Ganuthula et al., 2025; Li et al., 2025). In AI-supported learning contexts, higher levels of perceived interactivity and emotional gratification may reduce digital stress while reinforcing pleasure-oriented usage motives, which, over time, may foster addictive use patterns and elevate the risk of learning burnout (Ren and Wu, 2025). By contrast, no significant mediating effect of AI addiction is found in the relationship between inert thinking and learning burnout. This suggests that inert thinking reflects a relatively stable cognitive–motivational disposition rather than a proximal driver of compulsive, loss-of-control behavior, and its behavioral expression may depend on additional motivational, affective, or contextual conditions (Deemer et al., 2019). Moreover, inert thinking is more closely associated with reduced attention and task avoidance, whereas learning burnout typically results from cumulative mechanisms such as motivational decline, diminished self-efficacy, and emotional exhaustion arising from prolonged addictive use (Schaufeli et al., 2002).

The findings further show that the effects of perceived usefulness, perceived enjoyment, and inert thinking on learning burnout are primarily transmitted through a progressive chain-mediating pathway involving AI dependence and AI addiction. Specifically, perceived usefulness and perceived enjoyment first strengthen AI dependence, which subsequently facilitates the development of AI addiction and, in turn, contributes to higher levels of learning burnout. This sequential pattern is consistent with prior evidence indicating that reliance-oriented technology use may gradually escalate into uncontrolled addictive behavior and ultimately be associated with adverse learning outcomes (Han et al., 2023; Turel et al., 2011). In this study, therefore, AI dependence and AI addiction are conceptualized as two successive stages within this developmental process, whereas learning burnout functions as its downstream emotional and behavioral outcome.

The results further indicate that inert thinking does not directly predict AI addiction. Rather, its influence is transmitted indirectly through AI dependence and subsequently contributes to learning burnout. This pattern suggests that inert thinking primarily represents an underlying cognitive–motivational disposition that strengthens reliance-oriented use of AI, operating along a progressive pathway of “inert thinking-AI dependence-AI addiction - learning burnout.” Although previous studies have suggested possible reciprocal or reinforcing relationships between inert thinking and dependence-related behaviors (Fuglseth and Sørebø, 2014; Hong et al., 2019), the present study is based on cross-sectional data and a unidirectional structural model. Accordingly, the evidence supports only the chained mediation pathway identified here, and inferences regarding reverse or bidirectional effects should be treated with caution. Future research should therefore investigate the temporal dynamics of AI dependence and AI addiction, and examine whether, over extended timeframes, dependence-related behaviors may in turn reshape inert thinking, learning motivation, or executive-function regulation, thereby testing potential cyclical or stage-accumulative mechanisms.