Linking digital competence, self-efficacy, and digital stress to perceived interactivity in AI-supported learning contexts

Abstract

As artificial intelligence technologies become more integrated into educational contexts, understanding how learners perceive and interact with such systems remains an important area of inquiry. This study investigated associations between digital competence and learners’ perceived interactivity with artificial intelligence, considering the potential mediating roles of information retrieval self-efficacy and self-efficacy for human–robot interaction, as well as the potential moderating role of digital stress. Drawing on constructivist learning theory, the technology acceptance model, cognitive load theory, the identical elements theory, and the control–value theory of achievement emotions, a moderated serial mediation model was tested using data from 921 Chinese university students. The results indicated that digital competence was positively associated with perceived interactivity, both directly and indirectly through a sequential pathway involving the two forms of self-efficacy. Higher levels of digital stress were associated with a weaker indirect pathway, suggesting that elevated stress may constrain the extent to which self-efficacy is applied in artificial intelligence supported learning contexts. These findings suggest a multidimensional perspective on the cognitive and emotional factors associated with learner artificial intelligence interaction and may inform the design of digital learning environments that account for variation in learners’ competence and stress tolerance.

Introduction

Interactivity is widely regarded as a component of effective learning and has been associated with enhanced knowledge acquisition, increased learner engagement, and cognitive development¹. In recent years, the integration of artificial intelligence (AI) into education has introduced new modes of interaction, particularly between learners and intelligent systems². Advances in large language models have enabled AI tools to provide functions such as personalized instruction, real-time feedback, and flexible access, which may complement or modify established learning processes³. Systems such as ChatGPT and DeepSeek exhibit conversational capabilities that, in some contexts, approximate human interaction and may support more interactive learner–AI exchanges⁴. A central concept in this domain is perceived interactivity, which refers to learners’ subjective experience of meaningful, responsive, and controllable communication during interactive processes^5,6. As AI technologies develop, learners are increasingly engaged as active participants in exchanges with AI systems rather than serving solely as recipients of information⁷.

Digital competence is considered an important factor that facilitates productive interaction with AI tools. Learners with lower levels of digital competence may exhibit tendencies such as increased reliance on AI-generated responses without extensive critical evaluation, or reduced utilization of advanced functions due to limited technical skills⁸. These tendencies may be more evident in contexts requiring ongoing adaptation to complex digital content, where factors such as digital stress are associated with reduced capacity to apply skills effectively⁹.

In digital learning environments, individuals are often required to retrieve information using tools such as search engines, academic databases, and digital library systems. However, issues such as information overload, limited search proficiency, and challenges in evaluating information quality are associated with increased anxiety and reduced learning confidence among some learners¹⁰. In this context, information retrieval self-efficacy (IROWSE)¹⁰ has emerged as a relevant construct in information literacy and learning behavior research. At the same time, human–robot interaction self-efficacy (SE-HRI)¹¹ represents a related construct that focuses on individuals’ perceived competence and adaptability when interacting with AI-based systems. The development and application of these self-efficacy beliefs appear to be shaped by the intensity and complexity of the digital environment, where digital stress is related to differences in how learners perceive and sustain interactive engagement with AI-based tools¹².

Literature review

PILAI

Perceived interactivity of learner-AI interaction (PILAI) describes learners’ subjective perceptions of interaction with AI systems, emphasizing the sense of control, system responsiveness, and meaningful engagement^13,14. This construct differs from general measures of interactivity by focusing on the learner’s interpretation of being heard, understood, and actively engaged¹⁵. Technically advanced AI systems may fail to yield high perceived interactivity if such engagement is absent¹⁶. In educational contexts, higher PILAI has been associated with increased behavioral participation, sustained attention, and proactive involvement in AI-mediated learning^17,18. It has also been linked to motivational and cognitive benefits, including enjoyment, divergent thinking, and creativity, which have been related to academic performance^19–21. Moreover, higher PILAI may facilitate learner autonomy and self-regulation, enabling more independent navigation of AI platforms, adaptive responses to feedback, and greater control over the learning process^22–24.

Digital competence

The concept of digital competence has been described as evolving from basic operational skills to include higher-order abilities such as problem-solving, critical thinking, and knowledge construction through digital technologies^25,26. In AI-supported learning environments, higher levels of digital competence have been linked to greater confidence in evaluating AI-generated content, managing adaptive feedback, and engaging in self-directed learning strategies²⁷. Such competencies have also been associated with lower tendencies toward reliance on system outputs without critical appraisal or limited engagement with available platform functions²⁸. The application of digital competence appears to be shaped by contextual variables, including cognitive and emotional demands linked to digital stress²⁹. Conditions such as high task intensity or frequent system feedback have been reported in the literature to coincide with greater cognitive load, which may be associated with constraints on the effective deployment of digital skills³⁰.

IROWSE and SE-HRI

Self-efficacy refers to learners’ beliefs in learners’ capability to perform specific tasks³¹. Within AI-mediated learning contexts, two specific forms of self-efficacy are of particular relevance. IROWSE reflects learners’ confidence in their ability to locate and evaluate information, including their perceived control over search strategies and their engagement in inquiry-based learning and information assessment³². SE-HRI refers to confidence in interacting with AI agents or robotic systems, encompassing adaptability, operational fluency, and perceptions of system reliability. Previous studies have reported that higher SE-HRI has been associated with higher levels of trust in AI systems, adaptability, and willingness to use robotic technologies^33,34. While prior work has examined the roles of IROWSE and SE-HRI in technology-enhanced learning^14,35,36the ways in which these constructs operate jointly in AI-supported educational environments remain less well explored.

Digital stress

Digital stress has been described as the psychological strain associated with ongoing interaction with digital technologies, arising from factors such as frequent digital interruptions, sensitivity to social evaluation, and exposure to excessive or irrelevant information^37,38. These conditions reflect the challenges posed by persistent connectivity and high information density. In AI-supported learning environments, such stressors may be particularly salient. Immersive and adaptive systems may deliver frequent feedback, prompts, and recommendations³⁹which can place demands on learners to sustain cognitive processing, shift between tasks, and regulate emotional responses, particularly when content is personalized and dynamically updated. Such demands have been associated with higher cognitive load and may correspond with challenges in applying digital competence, which in turn could relate to variations in how learners perceive interactivity with AI systems^40,41.

Overall, while prior research has examined PILAI, digital competence, IROWSE, SE-HRI, and digital stress individually, limited work has integrated these constructs within a unified framework. This study addresses this gap by proposing a model to examine how technical literacy, self-efficacy, and psychological factors jointly shape engagement with AI systems in educational contexts.

Research hypotheses and model

The direct influence of digital competence on PILAI

While AI grows increasingly integrated into educational environments, learners’ experiences of interaction quality with these systems may be associated with their capacity to operate and adapt digital tools effectively⁴². From a Constructivist Learning Theory^43,44learners actively construct knowledge through interaction with their environment⁴⁵. Increased digital competence may correlate with interactions with AI systems that better align with learners’ objectives⁴². First, digital competence is related to the ability to formulate precise and context-specific prompts, which may increase the likelihood of obtaining outputs that address informational needs⁴⁶. Second, digital competence is associated with the capacity to evaluate AI-generated feedback for accuracy and relevance, to integrate pertinent information into ongoing tasks, and to disregard content that does not contribute to intended objectives^47,48. Third, learners with higher digital competence may be more likely to adjust interaction strategies, such as modifying prompt phrasing to elicit targeted explanations, selecting system functions appropriate to task requirements (e.g., summarization, problem-solving, language translation), and organizing these operations in sequences informed by system responses⁷. These capabilities have been linked in the literature to three components of perceived interactivity: (a) the ability to guide the interaction process, which corresponds with a sense of control; (b) the capacity to evaluate and act on system feedback, which is related to perceptions of responsiveness; and (c) the integration of AI-generated content into meaningful learning activities, which is associated with a sense of meaningfulness when the interaction is perceived as contributing to the achievement of learning objectives⁴⁹.

Empirical studies have shown that learners with higher digital competence tend to evaluate AI systems more positively, report higher satisfaction, and express stronger intentions to continue using them^50,51. These findings are consistent with conceptualizations of perceived interactivity, which emphasize autonomy, responsiveness, and relevance as central dimensions^52,53. Considering this both theoretical and empirical evidence, the following hypothesis is proposed:

H1 There is a significant effect of digital competence on PILAI.

The moderated mediation effect of digital stress

According to the Technology Acceptance Model (TAM) proposed by Davis⁵⁴perceived ease of use and perceived usefulness are regarded as key factors associated with individuals’ acceptance and use of technology⁵⁵. Higher levels of these perceptions have been linked in the literature to greater behavioral engagement and more frequent interaction with AI-based platforms⁵⁶. Learners with higher digital competence tend to exhibit greater familiarity with digital tools, higher efficiency in navigating interfaces, and increased confidence in managing interactive systems⁵⁷. These characteristics may be associated with stronger perceptions of ease of use and usefulness⁵⁸. When digital tools are perceived as easy to use, learners may develop greater confidence in their ability to engage with these tools, which could contribute to higher levels of interactive readiness and willingness for system engagement (IROWSE). IROWSE, in turn, may provide a cognitive and motivational basis for perceiving the system as useful, potentially aligning with higher levels of perceived interactivity. For example, when learners apply digital competence to process AI-generated feedback efficiently, respond to adaptive content, and engage in structured interactions, they may be more inclined to perceive the AI system as useful. Under such conditions, some studies have reported an association with greater cognitive engagement, including sustained attention, more systematic feedback interpretation, and goal-oriented learning behaviors⁵⁹. Within this framework, digital competence may be indirectly related to PILAI through its association with IROWSE. IROWSE may operate as a cognitive link that facilitates the application of learners’ skills in evaluating and acting upon information provided by AI, thereby supporting more consistent and confident interaction with AI systems.

However, the mediating role of IROWSE may vary across contexts depending on the level of digital stress experienced by learners. Digital stress can arise from multiple sources, including high volumes of incoming information, expectations for immediate responses, sustained connectivity, and concerns about evaluation in online environments^60,61.

According to Cognitive Load Theory (CLT), working memory has a limited capacity, and task performance may be constrained when task demands exceed available cognitive resources^62,63. Empirical studies have reported that digital stress can operate as an extraneous cognitive load, drawing on cognitive resources that could otherwise be allocated to learning processes^63,64. In AI-supported learning, this load can manifest when learners must interpret large volumes of adaptive feedback under time pressure, switch frequently between tasks, or troubleshoot unfamiliar system functions⁶⁵. These conditions may make it more difficult to apply digital competence or to activate self-efficacy beliefs, even for learners with strong capabilities. For example, research indicates that high information load and frequent interruptions can diminish perceived control and hinder goal-directed engagement with learning systems⁶⁶.

Conversely, when digital stress is lower, cognitive resources are less likely to be diverted to managing strain, leaving a greater proportion available for task-related processing. Evidence suggests that in low-stress digital environments, learners are more likely to sustain attention, process feedback effectively, and adapt interaction strategies to meet task requirements⁶⁷. In such situations, digital competence and IROWSE may be more fully mobilized, which could be associated with a stronger sense of control, greater confidence in navigating system features, and more consistent perceptions of PILAI. Based on this rationale, the following hypothesis is proposed:

H2 Digital stress moderates the indirect effect of digital competence on PILAI via IROWSE. The indirect effect is stronger when digital stress is lower.

Self-efficacy is a well-established psychological construct that has been associated with knowledge construction through its relation to motivation, persistence, and strategy use⁶⁸. SE-HRI may operate as a psychological mechanism linked to sustained and goal-directed engagement with AI tools. It may be associated with a greater willingness to explore system functions, adapt interaction strategies, and refine approaches in response to feedback, which could relate to more extensive knowledge construction processes⁶⁹. An empirical study in technology-enhanced learning contexts has found that higher self-efficacy in interacting with intelligent systems is associated with more frequent exploratory use, greater adaptability in problem-solving, and more effective integration of feedback into learning tasks⁷⁰.

From a constructivist learning perspective, learners acquire a sense of control and operational fluency through repeated interactions with external tools, gradually internalizing strategies and refining their responses⁷¹. Learners with higher digital competence tend to understand the operational logic, feedback structures, and affordances of AI systems more readil⁷²which can facilitate task comprehension and reduce uncertainty during interaction. This understanding may foster stronger motivation to experiment with system features and may reinforce SE-HRI over time⁷³. Empirical evidence supports this link: individuals with higher levels of digital competence have been shown to exhibit greater confidence and persistence in using AI-based learning tools^74,75. Higher levels of SE-HRI may be associated with an increased capacity for co-constructive engagement with AI, including the collaborative identification of problems, the interpretation of system-generated feedback, and the integration of new knowledge into existing frameworks. When such interactions are perceived as responsive and meaningful, they may be associated with higher levels of PILAI⁵³.

However, this pathway may be attenuated when digital stress is elevated. Research in online learning contexts indicates that high digital stress is associated with reduced attention control, lower persistence, and diminished self-efficacy beliefs⁷⁶. Within the framework of CLT, excessive cognitive load competes for the limited capacity of working memory, which may constrain the resources available for processing and integrating new information⁷⁷. The development of SE-HRI has been described in the literature as involving repeated cycles of performance, feedback interpretation, and strategy adjustment. Each of these cycles relies on the learner’s ability to attend to relevant cues, evaluate outcomes, and store effective strategies for future application. When digital stress introduces extraneous load (e.g., frequent notifications, rapid task switching, or the need to process large volumes of system-generated feedback), these cognitive operations may be disrupted. As a result, learners may have fewer opportunities to consolidate successful interaction patterns, limiting the reinforcement of self-efficacy beliefs⁶⁴. Over time, this constraint may be connected with a reduced likelihood of engaging in exploratory and adaptive behaviors related to greater SE-HRI, which in turn could be related to weaker perceptions of interactivity in AI-supported learning environments. Under these conditions, even learners with high digital competence may experience challenges in developing or maintaining SE-HRI, potentially limiting their capacity to perceive AI interactions as responsive and controllable⁷⁸. Conversely, when digital stress is relatively low, learners may be able to allocate more cognitive resources to exploring system features, applying feedback, and engaging in adaptive strategies, which could support the development of SE-HRI and be associated with higher perceptions of interactivity. Based on this reasoning, the following hypothesis is proposed:

H3 Digital stress moderates the indirect effect of digital competence on PILAI via SE-HRI. The indirect effect is stronger when digital stress is lower.

Knowledge or skills acquired in one context can be transferred to another when shared structural elements are activated⁷⁹. The Identical Elements Theory (IET) posits that transfer is most likely to occur when two tasks share common stimulus elements^80,81. IROWSE and SE-HRI appear to share several underlying cognitive structures, including goal-directed navigation, interpretation of system feedback, and adaptive regulation. Such shared cognitive processes can reduce the learning effort required to adapt from web-based information retrieval to interactive engagement with AI systems, thereby facilitating near transfer. Prior studies have found that skills in digital information search and evaluation can enhance confidence in related technological tasks, especially when operational procedures and feedback mechanisms are similar^82,83. Learners with higher levels of IROWSE may enter AI-supported environments with a greater readiness to recognize interface patterns, anticipate system logic, and apply adaptive strategies. This readiness may be associated with lower cognitive resistance and could support the development of SE-HRI. Over time, this confidence and fluency in managing digital systems may support more active and collaborative engagement with intelligent technologies⁸⁴. Conceptually, IROWSE and SE-HRI may represent sequentially related forms of self-efficacy, where competence in one domain supports the progressive development of efficacy in another, ultimately enhancing PILAI.

However, the efficiency of this transfer pathway may depend on digital stress levels. Elevated digital stress, which may arise from factors such as unfamiliar interfaces, operational uncertainty, or the pressure of constant connectivity, can be associated with the diversion of cognitive resources toward managing stress-related concerns rather than engaging in core learning activities⁸⁵. Within the framework of CLT, extraneous cognitive load from stressors such as cluttered interfaces, excessive notifications, or ambiguous operations competes with the working memory resources required for processing and integrating new knowledge⁶³. Studies on empirical data has found that technology-related anxiety and information overloading are associated with lower working memory efficiency, less frequent exploratory behaviors, and lower effectiveness of skill transfer in digital learning contexts⁶⁴. When digital stress is high, learners may have fewer opportunities to consolidate effective interaction strategies, which could limit the reinforcement of SE-HRI and be related to weaker perceptions of interactivity. Conversely, when digital stress is relatively low, cognitive resources may be more readily allocated to recognizing structural similarities between tasks, applying previously learned strategies, and building confidence in new interaction contexts, which could support the indirect pathway from IROWSE to PILAI via SE-HRI. Based on this reasoning, the following hypothesis is proposed:

H4 Digital stress moderates the indirect effect of IROWSE on PILAI via SE-HRI. The indirect effect is stronger when digital stress is lower.

According to the Control-Value Theory of Achievement Emotions (CVTAE)⁸⁶learners’ emotional experiences and engagement in learning tasks are shaped by their appraisals of control and value. Perceived control refers to the belief in one’s capacity to influence task processes and outcomes, encompassing constructs such as self-efficacy, mastery experiences, and expectations of success. Subjective value reflects the perceived importance or usefulness of a task, including interest value, utility value, and intrinsic value^87,88. Digital competence may function as a foundational cognitive resource for establishing perceived control, as it is associated with the technical and strategic capacity to manage task demands effectively⁸⁹. Empirical studies in technology-enhanced learning contexts have reported learners who have enhanced digital competence tend to show greater confidence in initiating and completing online information retrieval tasks, indicating that digital competence may be related to an initial and observable form of control in technology-based learning^90,91.

As task demands progress from basic information access to more complex human–AI interaction, the control experiences gained during information retrieval may transfer to broader contexts. Skills such as navigating complex interfaces, interpreting system feedback, and adapting search strategies can be applied to AI-based interactions, supporting the development of SE-HRI⁹². This transfer process aligns with findings in technology use that prior task mastery in related domains can reduce cognitive resistance and facilitate skill adaptation in new environments⁹³. In this way, digital competence may contribute to a layered construction of perceived control, progressing from IROWSE to SE-HRI. Within the CVTAE framework, higher levels of perceived control are associated with greater emotional stability and more positive anticipatory emotions during interaction^88,94which can encourage learners to explore AI functionalities more actively, potentially enhancing PILAI.

However, the serial mediation pathway from digital competence to PILAI, via IROWSE and SE-HRI, is unlikely to operate with equal strength across all learning contexts. One factor that may moderate this relationship is digital stress. Prior studies in online learning environments have shown that high levels of digital stress are associated with increased cognitive load, reduced attentional control, and diminished capacity for effective self-regulation^64,76.

As outlined in the CVTAE, negative achievement-related emotions such as anxiety and stress can erode learners’ perceived control, disrupt the regulation of achievement emotions, and limit the effective deployment of adaptive learning strategies⁸⁶. Stress from continuous connectivity and information overload reduced learners’ willingness to engage with new digital tools, even among those with high baseline digital competence. From a cognitive perspective, CLT posits that extraneous cognitive load competes for the limited capacity of working memory, thereby restricting the resources available for essential learning processes⁶³. Digital stress may function as an extraneous load by drawing cognitive resources toward managing interruptions, processing excessive feedback, or addressing operational uncertainty. These demands may be associated with interruptions to iterative processes such as interpreting feedback, refining strategies, and consolidating approaches, which are considered relevant for strengthening both IROWSE and SE-HRI. In such circumstances, even learners with comparatively high digital competence may experience difficulties in maintaining the confidence and adaptability related to perceiving AI interactions as controllable and responsive.

In contrast, low digital stress can help preserve cognitive resources for goal-directed engagement, enabling learners to apply prior skills in information retrieval to more complex AI-based interactions, thereby reinforcing self-efficacy and enhancing perceived interactivity. Accordingly, this study proposes the following hypothesis:

H5 Digital stress moderates the serial mediation effect of IROWSE and SE-HRI on the relationship between digital competence and PILAI. That is, the indirect effect is stronger under low levels of digital stress.

Hypothesis model

The hypothesized structural model of this study is presented in Fig. 1.

Fig. 1.

Hypothesized model of the study.

Method

Participants and procedure

This study involved undergraduate students from two universities in China, located in Jiangxi Province and Inner Mongolia Autonomous Region. These institutions were selected to capture regional diversity and to reflect variation in technological infrastructure and student digital exposure. Given the study’s focus on learner–AI interaction, participants were selected based on two inclusion criteria: (1) they reported using AI tools at least once per week; and (2) they were currently using AI applications to support academic learning or problem-solving tasks.

Data were collected via Questionnaire Star, a widely used online Survey platform in China. A snowball sampling method was adopted to increase recruitment efficiency and to reach participants who met the inclusion criteria. Initial respondents were invited to share the Survey link with peers who met the same conditions, facilitating broader access within relevant student communities. The study was conducted in accordance with the ethical standards of the institutional research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Ethical approval was obtained from the Ethics Committee of Jiangxi University of Science and Technology, which covered all procedures and participant groups, including those recruited from the collaborating institution in Inner Mongolia. All participants provided informed consent prior to data collection.

This study utilized G*Power 3.1⁹⁵ to determine the minimum required sample size. Based on Cohen’s⁹⁶ guidelines, the analysis specified a medium effect size (f² = 0.15), a significance level (α) of 0.05, and a statistical power (1 − β) of 0.95. The results indicated that a minimum of 146 participants was required to ensure adequate power (λ = 21.90, critical F = 2.16, actual power = 0.9508). A total of 1,000 responses were collected. After excluding questionnaires with excessive missing data (> 20%), completion times shorter than 120 s (indicating insufficient reading or response time), failed attention checks (e.g., incorrect responses to instructed items), or uniform response patterns (e.g., selecting the same option for ≥ 90% of items), 921 valid responses were retained for analysis, yielding a response rate of 92.1%. The decision to collect a larger sample was motivated by the complexity of the proposed structural model, which includes multiple mediating and moderating variables. A larger sample size was expected to enhance the stability and precision of parameter estimates in Structural Equation Modeling (SEM), reduce standard errors, and improve the robustness of the findings. Additionally, given the geographic and cultural diversity of the sample, drawn from two distinct regions in China, a broader dataset contributes to the generalizability and representativeness of the results across varying learner–AI interaction contexts.

Measures

Perceived interactivity of learner-AI interaction scale

PILAI was measured using the Perceived Interactivity of Learner-AI Interaction Scale, which was originally developed and validated by Wang et al.⁵³ to assess how learners perceive interactivity when engaging with artificial intelligence in educational contexts. The scale includes 17 items distributed across four dimensions: responsiveness (5 items), learner control (3 items), learning engagement (6 items), and personalization (3 items). Each item is rated on a six-point Likert scale ranging from 1 (strongly disagree) to 6 (strongly agree), with higher scores reflecting greater perceived interactivity. The original study demonstrated excellent internal consistency, reporting a Cronbach’s α of 0.948 for the overall scale, and values ranging from 0.820 to 0.915 for each subscale. Confirmatory factor analysis (CFA) also supported the scale’s structural validity (χ²/df = 2.148, CFI = 0.938, TLI = 0.926, RMSEA = 0.061, SRMR = 0.062), and convergent and discriminant validity were also well established. In the current study, the PILAI scale continued to show excellent psychometric properties. The internal consistency of the scale was confirmed with a Cronbach’s α coefficient of 0.93 for the full scale. The results of CFA indicated a good model fit (χ²/df = 2.232, CFI = 0.934, TLI = 0.932, RMSEA = 0.063, SRMR = 0.063), suggesting that the factor structure remained stable and valid across different samples. These findings confirm that the PILAI scale is a reliable and valid instrument for measuring university students PILAI in digitally supported learning environments.

Digital competence scale for university students

The digital competence was measured using the Digital Competence Scale for University Students (DC-US), which was originally developed and validated by Wang et al.⁹⁷ to Support digitally enhanced learning in higher education. The DC-US consists of 10 items grouped into two dimensions: technical literacy (4 items) and digital skills (6 items). Each item is rated on a five-point Likert scale ranging from 0 (strongly disagree) to 4 (strongly agree), with higher scores indicating stronger digital competence. The original study reported high internal consistency, with a Cronbach’s α coefficient of 0.94 for the overall scale. CFA also supported the structural validity of the instrument (χ²/df = 4.047, CFI = 0.961, TLI = 0.948, RMSEA = 0.079, SRMR = 0.036). In the present study, the DC-US continued to demonstrate excellent reliability, with a Cronbach’s α of 0.91 for the overall scale. The results of CFA indicated a good model fit (χ²/df = 2.913, CFI = 0.921, TLI = 0.913, RMSEA = 0.076, SRMR = 0.073), confirming the construct validity of the scale. These findings suggest that the DC-US is a reliable and valid instrument for assessing digital competence among university students in China.

Information retrieval on the web self-efficacy scale

Students’ IROWSE was measured using the Chinese version of the Information Retrieval on the Web Self-Efficacy Scale, Rodon and Chevalier¹⁰ adapted from the original scale developed by Rodon and Meyer⁹⁸. The adapted Chinese version consists of 8 items, each rated on a four-point Likert scale ranging from 1 (false) to 4 (true), with higher scores indicating stronger self-efficacy in web information retrieval tasks. In the original validation study of the Chinese version, Rodon and Chevalier reported psychometric properties including Cronbach’s α of 0.90 and CFA fit indices (GFI = 0.996, AGFI = 0.993, PGFI = 0.553, RMSR = 0.018). These metrics, while informative for the scale’s initial adaptation, were derived from the prior validation study rather than the present analysis. In the current study, we re-examined the measurement properties using Mplus and reported only the CFA results obtained from our own dataset to avoid methodological ambiguity. The scale demonstrated strong internal consistency (Cronbach’s α = 0.92), and CFA supported the one-factor structure with good model fit (χ²/df = 2.241, CFI = 0.956, TLI = 0.942, RMSEA = 0.052, SRMR = 0.058). These results confirm the construct validity of the Chinese version of the IROWSE in the present sample.

Self-efficacy in human–robot interaction scale

Students’ SE-HRI was assessed using the Chinese version of the Self-efficacy in Human–Robot Interaction Scale¹¹which was adapted from the original instrument developed by Pütten & Bock⁹⁹. The adapted Chinese version consists of 13 items covering two dimensions: robot control efficacy (6 items) and robot use efficacy (7 items). Each item is rated on a six-point Likert scale ranging from 1 (strongly disagree) to 6 (strongly agree), with higher scores indicating stronger perceived self-efficacy in interacting with robots. The SE-HRI scale demonstrated robust psychometric properties, with a Cronbach’s α coefficient of 0.91 and satisfactory structural validity indices (χ²/df = 3.49, CFI = 0.91, TLI = 0.89, RMSEA = 0.084) in prior validation studies. In the present study, the Chinese-adapted version of the SE-HRI scale exhibited excellent internal consistency, with a Cronbach’s α coefficient of 0.92. CFA indicated a good model fit (χ²/df = 2.965, CFI = 0.921, TLI = 0.937, RMSEA = 0.0742, SRMR = 0.0731), confirming the construct validity of the scale. These results demonstrate that the Chinese version of the SE-HRI scale is a reliable and valid instrument for assessing SE-HRI among university students in China.

Digital stress scale

Students’ digital stress was assessed using the Chinese version of the Digital Stress Scale (DSS-C), which was developed and validated by Zhang et al.³⁸. The DSS-C consists of 24 items across five dimensions: availability stress (3 items), approval anxiety (5 items), fear of missing out (4 items), connection overload (6 items), and online vigilance (4 items). Items are rated on a five-point Likert scale ranging from 1 (never) to 5 (always), with higher scores indicating greater levels of digital stress. The original DSS-C demonstrated strong psychometric properties, with Cronbach’s α values showed excellent internal consistency (α = 0.937) and confirmed the five-factor structure through CFA (χ²/df = 6.39, CFI = 0.938, TLI = 0.929, RMSEA = 0.063, SRMR = 0.044). In the present study, the DSS-C demonstrated similarly robust psychometric performance. The Cronbach’s α coefficient for the overall scale was 0.92, indicating high internal consistency. CFA indicated an acceptable model fit (χ²/df = 3.98, CFI = 0.911, TLI = 0.902, RMSEA = 0.065, SRMR = 0.061), supporting the construct validity of the scale. These findings confirm that the DSS-C is a reliable and valid instrument for assessing digital stress among Chinese university students.

Data analysis

A cross-sectional research design was employed in this study. Data analyses were conducted using SPSS version 27.0 and Mplus version 8.3. To assess potential common method bias, a CFA was first conducted in Mplus based on the single-factor test procedure. Following this, descriptive statistics and Pearson correlation analyses were performed in SPSS to calculate the means, standard deviations, and intercorrelations among the study variables. Gender, grade level, and academic discipline were included as control variables in all Subsequent analyses. To test the hypothesized moderated serial mediation model, SEM was conducted using Mplus. This analytic strategy allowed for the simultaneous examination of mediation and moderation effects within a unified model. A nonparametric bootstrapping procedure with 5,000 resamples was used to estimate the bias-corrected 95% confidence intervals for indirect and interaction effects. Effects were deemed statistically significant if the corresponding confidence intervals did not contain zero.

Additionally, to assess the potential influence of participant fatigue and sample homogeneity, robustness checks were conducted using multi-group SEM by response speed (fast vs. slow, based on completion time quartiles) and by academic major for group-difference tests. Path coefficients were reported as STDYX standardized estimates, which reflect standardized effects for both independent and dependent variables.

Results

Common method bias test

To assess potential common method bias, a single-factor CFA was conducted. The results revealed a poor model fit: χ²/df = 26.03, RMSEA = 0.18, CFI = 0.60, TLI = 0.73, SRMR = 0.16. These indices suggest that the single-factor model does not fit the data well¹⁰⁰indicating that common method variance is unlikely to be a serious concern in this study.

Descriptive statistics

Demographic information of the participants is presented in Table 1. All participants were undergraduate students. In terms of gender, 463 respondents were male (50.3%) and 458 were female (49.7%). Regarding academic year, 204 participants were freshmen (1st year; 22.1%), 258 were sophomores (2nd year; 28.0%), 236 were juniors (3rd year; 25.6%), and 223 were seniors (4th year; 24.2%). As for academic disciplines, 292 students (31.7%) majored in Humanities, 299 (32.5%) in Natural Sciences & Engineering, and 330 (35.8%) in Social Sciences & Applied Disciplines.

Table 1.

The participants’ demographic information.

Characteristics	Categories	N	Percentage (%)
Gender	Male	463	50.3
	Female	458	49.7
Grade level	Freshman (1st year)	204	22.1
	Sophomore (2nd year)	258	28
	Junior (3rd year)	236	25.6
	Senior (4th year)	223	24.2
Major	Humanities	292	31.7
	Natural Sciences & Engineering	299	32.5
	Social Sciences & Applied Disciplines	330	35.8

Correlation analysis between the variables

Pearson correlation analysis was conducted to examine the relationships among digital competence (DC), perceived interactivity of learner–AI interactions (PILAI), information retrieval on the web self-efficacy (IROWSE), human–robot interaction self-efficacy (SE-HRI), and digital stress (DS). All correlation coefficients were statistically significant at the p <.01 level, as shown in Table 2. DC was positively correlated with PILAI (r =.540, p <.01), IROWSE (r =.457, p <.01), and SE-HRI (r =.604, p <.01), and negatively correlated with DS (r = –.488, p <.01). PILAI was positively correlated with IROWSE (r =.547, p <.01) and SE-HRI (r =.590, p <.01), and negatively correlated with DS (r = –.516, p <.01). IROWSE was positively correlated with SE-HRI (r =.554, p <.01) and negatively correlated with DS (r = –.645, p <.01). SE-HRI was negatively correlated with DS (r = –.567, p <.01).

Table 2.

Means, standard deviations, and correlations of the variables. Comment. ^**P <.01, SD = standard deviation, dc = digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress.

	Mean	SD	DC	PILAI	IROWSE	SE-HRI	DS
DC	2.48	1.06	1
PILAI	2.94	1.11	0.540**	1
IROWSE	2.71	0.97	0.457**	0.547**	1
SE-HRI	3.15	1.07	0.604**	0.590**	0.554**	1
DS	2.34	1.17	− 0.488**	− 0.516**	− 0.645**	− 0.567**	1

Test of chain mediation effects

To evaluate the hypothesized chain mediation model, SEM was conducted using Mplus version 8.3. As shown in Table 3, DC significantly predicted IROWSE (β = 0.66, SE = 0.02, z = 26.34, p <.001). IROWSE, in turn, significantly predicted SE-HRI (β = 0.40, SE = 0.03, z = 17.59, p <.001). DC also exerted a significant direct effect on SE-HRI (β = 0.54, SE = 0.02, z = 24.11, p <.001). With respect to the outcome variable, PILAI, DC had a significant direct effect (β = 0.26, SE = 0.03, z = 8.29, p <.001). In addition, both IROWSE (β = 0.32, SE = 0.03, z = 11.46, p <.001) and SE-HRI (β = 0.34, SE = 0.04, z = 9.73, p <.001) significantly predicted PILAI. The model explained 44% of the variance in IROWSE, 74% in SE-HRI, and 70% in PILAI. Gender, grade level, and academic discipline were included as covariates in the model. None of these control variables showed statistically significant effects on the outcome variables (all p values > 0.05).

Table 3.

Results of the chain mediation model. Comment. ^***P <.001, β = standardized coefficient, R² = variance explained in each outcome variable, dc = digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy the bootstrap test results (Table 4) revealed that IROWSE and SE-HRI partially mediated the relationship between DC and PILAI, with a total indirect effect of 0.51 (SE = 0.03, 95% CI: 0.46 to 0.56). This effect comprised three distinct pathways: (1) DC → IROWSE → PILAI (effect = 0.22, SE = 0.02, 95% CI: 0.18 to 0.26), (2) DC → SE-HRI → PILAI (effect = 0.20, SE = 0.02, 95% CI: 0.15 to 0.24), and (3) DC → IROWSE → SE-HRI → PILAI (effect = 0.09, SE = 0.01, 95% CI: 0.07 to 0.12). these findings suggest the presence of multiple mediating pathways through which DC contributes to students’ PILAI. Moreover, the significant direct effect of DC on PILAI (effect = 0.27, SE = 0.03, 95% CI: 0.20 to 0.33) further supports H1.

Outcome variable	Predictor	R 2	β	SE	z	P
IROWSE	DC	0.44	0.66	0.02	26.34***	< 0.001
	Gender		0.04	0.05	1.43	0.153
	Grade level		−0.04	0.02	−1.41	0.16
	Major		0.03	0.03	1.26	0.207
SE-HRI	DC	0.74	0.54	0.02	24.11***	< 0.001
	IROWSE		0.40	0.03	17.59***	< 0.001
	Gender		−0.01	0.04	−0.82	0.413
	Grade level		−0.01	0.02	−0.27	0.791
	Major		0.03	0.02	1.85	0.065
PILAI	DC	0.70	0.26	0.03	8.29***	< 0.001
	IROWSE		0.32	0.03	11.46***	< 0.001
	SE-HRI		0.34	0.04	9.73***	< 0.001
	Gender		0.03	0.04	1.78	0.076
	Grade level		−0.02	0.02	−1.05	0.295
	Major		0.01	0.03	0.44	0.658

Table 4.

Total, direct, and indirect effects of DC on PILAI. Comment. DC = Digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress.

Effect type	Estimate (B)	SE	Boot LLCI	Boot ULCI
Total effect	0.77	0.02	0.73	0.82
Direct effect	0.27	0.03	0.20	0.33
Total indirect effect	0.51	0.03	0.46	0.56
DC → IROWSE → PILAI	0.22	0.02	0.18	0.26
DC → SE-HRI → PILAI	0.20	0.02	0.15	0.24
DC → IROWSE → SE-HRI → PILAI	0.09	0.01	0.07	0.12

The moderating role of digital stress

To test Hypotheses 2 through 5, the serial mediation model was extended by incorporating interaction terms involving DS. As shown in Table 5, the interaction between IROWSE and DS was statistically significant (β = −0.09, SE = 0.03, z = − 2.54, p =.011), as was the interaction between SE-HRI and DS (β = −0.08, SE = 0.03, z = − 2.71, p =.007). These findings indicate that DS moderated the indirect relationships between DC and PILAI, consistent with a moderated mediation model.

Table 5.

Results of the moderated serial mediation model. Comment. ^*P <.05, ^**P <.01, ^***P <.001, β = standardized coefficient, R² = variance explained in each outcome variable, dc = digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress.

Outcome variable	Predictor	R 2	β	SE	z	P
IROWSE	DC	0.43	0.60	0.02	26.34***	< 0.001
	Gender		0.07	0.05	1.43	0.153
	Grade level		−0.03	0.02	−1.41	0.163
	Major		0.04	0.03	1.26	0.207
SE-HRI	DC	0.74	0.54	0.02	24.11***	< 0.001
	IROWSE		0.44	0.02	17.59***	< 0.001
	Gender		−0.03	0.04	−0.82	0.413
	Grade level		0.00	0.02	−0.27	0.790
	Major		0.04	0.02	1.85	0.065
PILAI	DC	0.71	0.22	0.03	6.72***	< 0.001
	IROWSE		0.29	0.04	6.81***	< 0.001
	SE-HRI		0.32	0.04	8.34***	< 0.001
	DS		−0.21	0.05	−4.08***	< 0.001
	IROWSE × DS (Int)		−0.09	0.03	−2.54*	0.011
	SE-HRI × DS (Int)		−0.08	0.03	−2.71**	0.007
	Gender		0.06	0.04	1.62	0.106
	Grade level		−0.02	0.02	−1.11	0.267
	Major		0.01	0.024	0.22	0.829

Specifically, the negative interaction between IROWSE and DS Suggests that the indirect effect of DC on PILAI via IROWSE was stronger when DS was low, providing Support for Hypothesis 2. Likewise, the moderation of the indirect path from DC to PILAI via SE-HRI by DS was also supported, as indicated by the significant interaction term, thereby confirming H3.

Furthermore, the indirect effect of IROWSE on PILAI through SE-HRI was also significantly moderated by DS. The interaction between SE-HRI and DS (β = −0.08, SE = 0.03, z = − 2.71, p =.007) indicates that this pathway was weakened under conditions of higher DS, supporting H4.

Finally, the overall pattern of results, including the significant direct paths from DC to IROWSE (β = 0.60, z = 26.34, p <.001), from IROWSE to SE-HRI (β = 0.44, z = 17.59, p <.001), and from SE-HRI to PILAI (β = 0.32, z = 8.34, p <.001), together with the two significant interaction effects is consistent with a moderated serial mediation model as illustrated in Fig. 2. The indirect effect of DC on PILAI through both IROWSE and SE-HRI was stronger when DS was low, which provides empirical support for H5.

Fig. 2.

Verification of the research model.

The final model explained 43% of the variance in IROWSE, 74% in SE-HRI, and 71% in PILAI. Gender, grade level, and academic discipline were included as covariates, and none of these variables demonstrated significant effects on the outcome variables (all p values > 0.05).

To explore the moderated mediation, we tested the conditional indirect effects of DC on PILAI at different levels of DS. As shown in Table 6, the indirect effect of DC on PILAI through IROWSE was stronger at lower levels of DS (β = 0.23, SE = 0.03, 95% CI [0.17, 0.30]) compared to mean (β = 0.17, SE = 0.02, 95% CI [0.13, 0.22]) and high levels of DS (β = 0.11, SE = 0.03, 95% CI [0.06, 0.17]). This pattern supports the hypothesis that DS dampens the positive impact of IROWSE on PILAI. Figure 3 visually illustrates this moderation, showing a steeper slope for individuals with lower DS.

Table 6.

The moderated mediation effect at different levels of digital stress. Comment. DC = Digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress.

Path	DS	Effect	BootSE	BootLLCI	BootULCI
DC → IROWSE → PILAI	M − 1SD	0.23	0.03	0.17	0.30
	M	0.17	0.02	0.13	0.22
	M + 1SD	0.11	0.03	0.06	0.17
DC → SE-HRI → PILAI	M −1SD	0.22	0.03	0.17	0.28
	M	0.17	0.02	0.13	0.21
	M + 1SD	0.12	0.02	0.08	0.17
DC → IROWSE → SE-HRI → PILAI	M − 1SD	0.11	0.02	0.08	0.14
	M	0.08	0.01	0.06	0.10
	M + 1SD	0.06	0.01	0.04	0.08

Fig. 3.

Digital stress as a moderator in the relationship between IROWSE and PILAI. Comment. IROWSE = Information retrieval on the web self-efficacy, PILAI = Perceived interactivity of learner–AI interactions.

Similarly, the indirect effect of DC on PILAI through SE-HRI was stronger under low DS (β = 0.22, SE = 0.03, 95% CI [0.17, 0.28]) than at the mean (β = 0.17, SE = 0.02, 95% CI [0.13, 0.21]) or high DS levels (β = 0.12, SE = 0.02, 95% CI [0.08, 0.17]). This indicates that DS also weakens the mediating role of SE-HRI, as depicted in Fig. 4, where the slope for the low-DS group is again more pronounced.

Fig. 4.

Digital stress as a moderator in the relationship between SE-HRI and PILAI.Comment.SE-HRI = Human–robot interaction self-efficacy, PILAI = Perceived interactivity of learner–AI interactions.

Finally, the serial indirect effect (DC → IROWSE → SE-HRI → PILAI) was strongest when digital stress was low (β = 0.11, SE = 0.02, 95% CI [0.08, 0.14]), compared to the mean (β = 0.08, SE = 0.01, 95% CI [0.06, 0.10]) and high DS levels (β = 0.06, SE = 0.01, 95% CI [0.04, 0.08]). These results further support the moderated serial mediation, demonstrating that DS consistently attenuates the strength of all indirect effects.

Robustness checks

To examine whether lengthy scales induced systematic fatigue effects, a multi-group SEM was conducted comparing fast responders (≤ 25th percentile of completion time) and slow responders (≥ 75th percentile). As presented in Table 7, all structural paths were statistically significant in both groups (p <.001). Parameter-difference tests indicated that none of the group differences reached statistical significance. Similarly, the total indirect effect of digital competence on PILAI via IROWSE and SE-HRI was highly comparable across groups (β_fast = 0.55, β_slow = 0.54). These findings suggest that the pattern and magnitude of effects were broadly consistent between fast and slow responders, indicating that potential participant fatigue was suggesting that participant fatigue was unlikely to have had a major influence on the data.

Table 7.

Multi-group SEM results for fast vs. slow responders. Comment. DC = Digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress. β = standardized coefficients; se = standard error; ^***p <.001.

Path	Fast group (≤ 25%) β (SE)	Slow group (≥ 75%) β (SE)	Difference (Δβ)
DC → IROWSE	0.63*** (0.02)	0.62*** (0.02)	0.01
DC → SE-HRI	0.82*** (0.02)	0.83*** (0.02)	−0.01
IROWSE → PILAI	0.39*** (0.03)	0.40*** (0.03)	−0.01
SE-HRI → PILAI	0.36*** (0.04)	0.37*** (0.04)	−0.01
DC → PILAI	0.26*** (0.03)	0.25*** (0.03)	0.01
Indirect DC →IROWSE/SE-HRI →PILAI	0.55*** (0.03)	0.54*** (0.03)	0.01

Because snowball sampling tends to expand within existing peer networks, concerns May arise that the recruited participants are overly homogeneous in their learning experiences. In the present study, however, participants were drawn from diverse academic fields, which differ in curricular design, disciplinary practices, and the ways AI tools are applied for academic tasks. Examining whether the structural associations varied across these disciplinary groups therefore provided an opportunity to evaluate whether the findings generalized beyond a narrow subset of students. To this end, a multi-group SEM was performed across three categories (Humanities, natural sciences & engineering, and social sciences & applied disciplines; see Table 1 for subgroup sizes). As shown in Table 8, all structural paths were statistically significant within each subgroup (p <.001). The estimated parameters were highly similar across majors, and no meaningful group differences were detected in pairwise comparisons. The total indirect effect of digital competence on PILAI through IROWSE and SE-HRI was also consistent across subgroups (βs = 0.54–0.55). Taken together, these results indicate that the observed associations were stable across academic fields within the present sample. Given the cross-sectional, self-report nature of the data, all paths are interpreted as associations rather than causal effects.

Table 8.

Multi-group SEM results by major. Comment. DC = Digital competence, pilai = perceived interactivity of learner–AI interactions, irowse = information retrieval on the web self-efficacy, SE-HRI = Human–robot interaction self-efficacy, ds = digital stress. β = standardized coefficients; se = standard error; ^***p <.001.

Path	Humanities β (SE)	Natural Sciences & Engineering β (SE)	Social Sciences & Applied Disciplines β (SE)
DC → IROWSE	0.68***	0.69***	0.68***
DC → SE-HRI	0.82***	0.82***	0.81***
IROWSE → PILAI	0.35***	0.36***	0.34***
SE-HRI → PILAI	0.36***	0.35***	0.35***
DC → PILAI	0.25***	0.26***	0.25***
Indirect DC →IROWSE/SE-HRI →PILAI	0.54***	0.55***	0.54***

Discussion

The results suggest a positive association between digital competence and PILAI. Within the framework of Constructivist Learning Theory, learning is described as the active construction of knowledge through interactions with the environment, in which learners’ existing competencies may be associated with the quality and depth of these interactions¹⁰¹. In this context, digital competence can be understood as an enabling resource that supports goal-directed exploration, the identification of relevant information, and the integration of new inputs from AI systems into existing knowledge structures. Learners with comparatively higher digital competence may be more likely to interpret AI feedback accurately, may recognize patterns more readily, and may regulate interaction strategies in ways aligned with their learning objectives²⁸. This interpretation is consistent with Constructivist Learning Theory’s emphasis on learner agency and adaptive regulation as mechanisms that sustain engagement and facilitate interactive learning experiences¹⁰². Empirical studies have reported comparable patterns, showing that higher digital competence is associated with more effective use of intelligent tutoring systems and greater engagement in technology-rich learning environments^42,103. The observed association in this study may reflect the role of digital competence as a set of cognitive and metacognitive resources that enable learners to engage more effectively with the affordances of AI-mediated learning environments¹⁰⁴.

The results provide support for H2, indicating that IROWSE mediates the relationship between digital competence and PILAI, with this pathway moderated by digital stress. Within the TAM, digital competence can be linked to perceptions of ease of use, which may facilitate more effective navigation and utilization of AI systems. IROWSE, as an indicator of confidence in information retrieval, may be associated with perceptions of usefulness in these interactions. When both ease of use and perceived usefulness are present, learners may be more likely to sustain constructive engagement with AI-mediated environments, which may be associated with higher levels of perceived interactivity¹⁰⁵. CLT provides an additional lens by considering how cognitive resources are managed during task performance. From this perspective, digital competence supports the efficient allocation of cognitive resources to task-relevant processes, enabling goal-directed information retrieval and integration of AI feedback¹⁰⁶. Digital stress may function as an extraneous cognitive load that competes for limited cognitive resources¹⁰⁷. Under conditions of elevated digital stress, attentional capacity may be allocated to managing demands such as processing high volumes of information or addressing interface-related difficulties¹⁰⁸. This diversion may correspond with a reduced ability to apply information retrieval strategies effectively, thereby weakening the mediating role of IROWSE in the pathway from digital competence to perceived interactivity. Under lower stress conditions, more cognitive resources may remain available for core task activities, allowing IROWSE to play a stronger role in linking digital competence with perceptions of interactivity¹⁰⁹.

The results also support H3, indicating that SE-HRI mediates the relationship between digital competence and PILAI, with this mediation moderated by digital stress. From the perspective of Constructivist Learning Theory, digital competence may be viewed as a resource that supports learners’ engagement in active, goal-directed interactions with AI systems. Learners with higher levels of digital competence may be more able to navigate system functions, interpret operational logic, and apply adaptive strategies¹¹⁰. These capabilities may be associated with the development of SE-HRI, as repeated and effective interactions with AI tools could correspond with greater confidence in managing system features and interpreting feedback¹¹¹. Constructivist Learning Theory’s emphasis on learner agency and the active construction of knowledge supports this interpretation¹⁰¹higher competence may provide a basis for more purposeful and self-regulated engagement with external systems, which may be associated with higher perceptions of interactivity. However, the findings also indicate that the association between SE-HRI and PILAI is weaker under higher levels of digital stress. In CLT terms, digital stress may represent an extraneous cognitive load that competes with the working memory resources required for processing AI feedback, experimenting with new features, and sustaining task engagement⁶³. Under elevated stress, a greater proportion of cognitive resources may be allocated to managing stress-related demands, such as resolving operational difficulties or coping with interface complexity, leaving fewer resources available for the metacognitive and regulatory processes that underpin SE-HRI¹¹². This theoretical framing helps explain why, even when digital competence is high, the translation of self-efficacy in AI interactions into perceptions of interactivity may be constrained under high stress conditions. Conversely, when stress levels are lower, fewer resources are diverted to extraneous demands, allowing SE-HRI to play a stronger role in linking digital competence with perceived interactivity in AI-mediated learning environments.

The results for H4 indicate a sequential mediating pathway from IROWSE to SE-HRI, which subsequently predicts PILAI. This pattern is consistent with the IET, which posits that the transfer of learning is more likely when tasks share overlapping cognitive or procedural elements⁸⁰. In the present context, IROWSE and SE-HRI both involve goal-directed exploration, adaptive regulation of strategies, and interpretation of system feedback. Such overlap may allow learners to draw on strategies and confidence developed in web-based information retrieval when interacting with AI systems, thereby facilitating a partial transfer of efficacy beliefs and interaction skills. This theoretical lens provides a plausible account of the observed association, suggesting that the metacognitive and procedural competencies underlying IROWSE may align closely enough with those required for SE-HRI to enable continuity in learners’ self-efficacy across related digital tasks. However, the findings also suggest that this transfer is sensitive to the level of digital stress. Under elevated digital stress, the progression from IROWSE to SE-HRI appears weaker. From a CLT perspective, digital stress may act as an extraneous cognitive load, consuming working memory resources that could otherwise be devoted to the regulation of interaction strategies and the interpretation of feedback⁶³. Emotional distractions and reduced perceived control under high-stress conditions may further limit the activation of prior self-efficacy beliefs in new but cognitively similar contexts¹¹³. This combination of cognitive and emotional constraints may restrict learners’ capacity to maintain or build upon previously acquired competencies¹¹⁴which could help explain the attenuated sequential pathway observed in the high-stress condition.

The results Support Hypothesis 5, which proposed a sequential mediation model in which IROWSE and SE-HRI jointly mediate the relationship between digital competence and PILAI, with this pathway moderated by digital stress. Within the framework of the CVTAE, learners’ emotional states during learning are shaped by their perceived control over the task and the value they attach to it¹¹⁵. Higher digital competence may contribute to stronger perceptions of control, enabling learners to approach information retrieval tasks and AI-mediated interactions with greater confidence. As learners successfully complete these tasks, they may develop successive layers of self-efficacy, initially in information retrieval and subsequently in human–AI interaction. These efficacy beliefs, when coupled with a sense of value in the learning activity, are theorized in CVTAE to facilitate the activation of cognitive and motivational resources that support sustained engagement¹¹⁶. In the present context, the sequential mediation from digital competence to PILAI through IROWSE and SE-HRI may be viewed as consistent with the CVTAE perspective, which posits that perceived control and value can operate in an accumulative manner, wherein earlier experiences of mastery may provide a basis for the development of subsequent, more complex competencies. Under conditions of low digital stress, the cognitive and emotional resources predicted by CVTAE may remain more readily available, allowing learners to apply prior efficacy beliefs to new but related tasks¹¹⁷. By contrast, under high levels of digital stress, these resources may be reduced, as stress has been associated with lower perceived control and with shifts in attention away from task goals. This potential reduction in available cognitive and motivational capacity may be related to a lower likelihood that efficacy beliefs developed in earlier stages of the sequence are mobilized in later, more complex interactions with AI systems¹¹⁸. This interpretation offers a theoretically grounded explanation for the observed moderation effect, highlighting how emotional and cognitive mechanisms may jointly influence the relationship between digital competence and perceived interactivity in AI-supported learning environments.

Robustness checks were undertaken to explore whether the findings were influenced by potential response biases or sampling procedures. When participants were divided by response speed, the structural relations among digital competence, self-efficacy measures, and PILAI remained statistically significant in both groups. The pattern of associations showed close similarity, and no systematic differences were detected. This outcome suggests that the use of relatively lengthy scales did not produce distortions in the estimated relations due to fatigue effects. In addition, the possibility that snowball sampling might have yielded an overly homogeneous sample was considered by comparing students from different academic fields. Despite disciplinary differences in curricula and the ways AI tools are typically integrated into learning tasks, the structural paths and indirect effects were consistent across Humanities, Natural Sciences & Engineering, and Social Sciences & Applied Disciplines. No major divergences emerged that would indicate discipline-specific influences on the model. Within the scope of the present data, the consistency across both response speed and academic field provides some reassurance that the observed associations were not solely an artifact of fatigue or restricted sampling.

Implications

The findings of this cross-sectional study should be interpreted as preliminary and associative rather than causal. Several considerations emerge for AI-mediated learning environments, but these must be weighed against the design’s limitations. First, the observed relationship between digital competence and PILAI suggests that skills such as formulating purposeful prompts, critically evaluating system feedback, and regulating learning processes may be relevant correlates of perceived interactivity. While these results do not establish temporal order, they point to possible directions for incorporating digital competence training into broader AI literacy initiatives in higher education. Longitudinal or experimental research would be needed to examine whether strengthening digital competence is associated with enhanced interaction outcomes over time.

Second, the association between IROWSE and SE-HRI may hint at a sequential pattern in self-efficacy beliefs, but the present design cannot verify such progression. Instructional activities that first target IROWSE and later extend to AI-based interactions could be considered for exploration in future work, yet only prospective or intervention-based studies can assess whether such sequencing is likely to be beneficial.

Third, the moderating role of digital stress suggests that higher digital stress levels may be related to a reduced likelihood of applying digital competence effectively in AI-supported tasks. Stress could reduce students’ ability to apply digital competence, or conversely, difficulties in interacting with AI systems might contribute to stress. Future research employing longitudinal designs, stress-reduction interventions, or behavioral measures of system use will be valuable to clarify these possibilities.

Limitations and future directions

Several limitations should be acknowledged when interpreting the findings. First, the study employed a cross-sectional design, which constrains the ability to establish temporal ordering among variables and precludes causal inferences regarding the proposed mediation and moderation processes. Because data were gathered at a single point in time, it remains uncertain whether digital competence precedes perceived interactivity or whether these perceptions subsequently influence learners’ self-assessments of competence. Longitudinal research could help trace changes in digital competence, self-efficacy, and perceived interactivity over time, particularly under repeated exposure to AI-supported learning. Experimental approaches may also help examine whether targeted instructional interventions, such as training in digital information retrieval or structured AI-interaction tasks, are associated with different outcomes under varying levels of digital stress.

Second, the sampling approach and participant profile may constrain the extent to which the findings can be generalized. Although additional analyses were conducted to examine possible fatigue effects and potential homogeneity due to snowball sampling, these risks cannot be fully excluded. Snowball sampling increases the likelihood that participants share overlapping networks and experiences, which may reduce variability in the data. Some heterogeneity was present in the current sample, as students came from different majors, represented multiple academic years, and were recruited from universities located in both northern and southern regions of China. Nevertheless, the sample still consisted exclusively of undergraduates, and the educational, institutional, and cultural contexts of these settings may have shaped their responses. For instance, access to digital resources, instructional practices, and cultural norms surrounding technology use may influence both self-efficacy and perceptions of interactivity. Learners in collectivist cultural contexts may also place greater weight on peer validation or external guidance, which could affect their confidence in AI-supported learning. Future research could adopt probability-based or stratified sampling strategies and conduct cross-cultural comparisons to examine whether the observed patterns hold across diverse educational systems and cultural settings.

Third, the study relied on measurement instruments with a relatively large number of items. While these scales were chosen for their established psychometric properties, the cumulative length may have contributed to respondent fatigue. Although robustness checks indicated that response speed did not materially alter the observed associations, the possibility of residual fatigue effects cannot be fully excluded. Future research could explore shorter validated measures or alternative administration procedures to reduce burden on participants while maintaining measurement quality.

Conclusion

This study examined the associations between digital competence and learners’ PILAI, with attention to the potential mediating roles of IROWSE and SE-HRI, as well as the potential moderating role of digital stress. The results suggest that digital competence may be related to PILAI both directly and indirectly through a sequential pathway involving self-efficacy beliefs. IROWSE and SE-HRI operated as linked mediators that were associated with learners’ cognitive and motivational engagement with AI systems. The indirect pathway appeared weaker under higher levels of digital stress, which was associated with lower relationship strengths across the mediation chain. The study draws on concepts from constructivist learning theory, the technology acceptance model, cognitive load theory, the identical elements theory, and the control-value theory of achievement emotions, suggesting a multidimensional perspective on how digital skills, confidence, and emotional factors may co-occur in AI-enhanced learning contexts. These findings may inform the design of digital learning environments and adaptive AI systems that account for variability in learners’ competence and stress tolerance.

Author contributions

Study concept and design: J.R. Data collection, analysis, and interpretation: J.R. and J.G. Writing – original draft: J.R. Writing – review and editing: J.G. and H.L. All authors have read and approved the final version of the manuscript for publication.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The study involving human participants was reviewed and approved by the Ethics Committee of Jiangxi University of Science and Technology (Ref: JXUST-2025-008). Informed consent was obtained electronically from all participants prior to their participation. The study was conducted in accordance with the Declaration of Helsinki.