GenAI literacy and online learning resilience among Chinese university students: the mediating role of self-efficacy

1 Introduction

In an era where digital technology is deeply embedded in education, higher education is undergoing unprecedented transformation. Universities are placing increasing emphasis on digital education transformation (Hemsley-Brown and Oplatka, 2015) and innovation in talent cultivation models (Chapleo, 2015). Education providers have come to recognize the importance of technology-empowered education and are actively exploring the use of emerging technologies to attract and retain more learners (Hemsley-Brown and Goonawardana, 2007). Online education represents not merely a shift in technological means of communication, but a new learning ecology. This mode of learning, which embodies the principles of sustainable education, has become deeply integrated into the daily learning of university students. For these students, adapting to the online learning ecology and enhancing their online learning resilience are crucial to their personal development (Abuhassna et al., 2020; Plungpongpan et al., 2016).

Generative Artificial Intelligence (GenAI) has become a key bridge for interaction between universities and students. A growing number of university students have started using GenAI tools to support their learning. It not only enables richer and more engaging forms of interaction but also goes beyond the boundaries of traditional knowledge transmission, helping institutions build lasting learning connections with students (Thackeray et al., 2008). The traditional role of education is shifting: students are moving from passively receiving knowledge to engaging in more efficient, two-way interactions with GenAI. Learners now have more opportunities to express their views and co-create content. This approach makes education feel less rigid and more like a team effort, where everyone can take part and jointly shape their own learning experience (Lipsman et al., 2012; Mangold and Faulds, 2009).

With the support of GenAI, universities have become more open in listening and responding, creating spaces where students can explore together and collaboratively solve problems. This deeper level of engagement adds value to the online learning process. Students not only learn more creatively but also change their own learning habits (Sashi, 2012). Thus, GenAI plays a significant role in shaping how students learn and what they are capable of, especially in the open and inclusive environment of online learning, where students can more easily find learning paths that suit them. Administrators in higher education must recognize that GenAI will influence the learning support systems of universities and, in turn, affect students’ online learning resilience.

In the era of digital technology becoming deeply embedded in education, online learning has shifted from a supplementary form of instruction to a normative mode in higher education, and online learning resilience—defined as learners’ dynamic capacity to sustain academic pursuits when facing digital learning adversities such as information overload, technical failures, and teacher–student separation—has become a key construct for understanding university students’ learning effectiveness in digital environments. However, existing research has largely focused on conceptualizing online learning resilience and elaborating its significance, while the antecedents that shape this capacity remain underexplored. Meanwhile, generative artificial intelligence (GenAI), with its capabilities in automated content generation, personalized learning support, and real-time interactive feedback, is profoundly reshaping how university students access resources and construct knowledge, yet prior studies have predominantly attended to GenAI's effects on learning efficiency and outcomes, with limited systematic empirical attention to how GenAI literacy—the ability to understand, use, critically evaluate GenAI tools, and maintain ethical awareness in their application—influences students’ adaptation and persistence in digital learning predicaments. The intersection of these two research gaps constitutes the core question of this study: whether and how GenAI literacy affects online learning resilience through the mediating role of self-efficacy. Accordingly, this study constructs an integrative theoretical framework with GenAI literacy as the antecedent, online learning resilience as the outcome, and self-efficacy as the mediator, aiming to provide a theoretical explanation for understanding university students’ adaptive mechanisms in digital-age online learning and to offer empirical evidence for higher education institutions to optimize GenAI literacy education and enhance students’ online learning resilience.

This study is structured as follows: (1) literature review and theoretical analysis; (2) research methods and hypothesis testing results; (3) discussion and conclusions.

2 Literature review

2.1 GenAI development

GenAI, centered on Large Language Models (LLMs), marks a shift from “analytical” to “creative” AI. Its importance lies not only in the exponential increase in analytical efficiency, but also in its reorganization of human cognitive patterns and social collaboration, according to Taeihagh (2025), who notes that the rapid proliferation of GenAI is reshaping the global innovation landscape at an unprecedented rate. Through automated content generation, code writing, and complex data synthesis, GenAI dramatically lowers the barriers to knowledge production (Wang and Zhao, 2024). According to Sauvola et al. (2024), GenAI is a “disruptive challenge” to traditional software development and creative industries. At the same time, the ensuing ethical risks, privacy violations and other issues (Pezzulo et al., 2024) have made it necessary for all sectors of society to embrace the dividends of the technology and at the same time face the pressing issues of governance and regulation (Dong and Chen, 2024).

GenAI has shown strong potential for application in several domains. research by Jauhiainen and Guerra (2023) shows that GenAI significantly enhances developer productivity. In healthcare, it is not only used to assist diagnosis, but also plays a role in simulation and assisted decision-making in medical research (Feng et al., 2024), e.g., Mastrianni et al. (2025) explored the supportive role of GenAI in genetic data analysis. In the transportation domain, Yan and Li (2025) systematically elucidated how GenAI can address the challenges of data sparsity and uncertainty in transportation decision-making. In addition, in the field of marketing and media, Duivenvoorde (2025) analyzes the empowerment of GenAI for creative marketing, while revealing the risks of deep falsification, consumer vulnerability exploitation, and so on.

In the field of education, the development of GenAI is triggering a profound revolution in teaching and learning. This transformation is no longer limited to instrumental aids, but permeates all parts of the education ecosystem. On the one hand, GenAI greatly empowers teachers’ professional development and instructional design (Tan, 2025), and through the introduction of platforms such as the TEAM model, teachers are able to use GenAI to conduct more in-depth case studies to optimize lesson plans and achieve precise teaching (Lee., 2026; Liang et al., 2024). On the other hand, for students, GenAI is reshaping the learning path, offering unlimited possibilities for personalized learning (Yang and Lin, 2025). However, students’ attitudes and adoption behaviors toward GenAI are influenced by a variety of psychological factors, and Kong and Yang (2024) suggest that fostering positive attitudes toward technology is a prerequisite for maximizing the potential of GenAI, while Fock and Siller (2025) suggest that education for Human Flourishing should develop students’ ability to collaborate with AI, not just manipulate technology. The goal of education should also shift from simple knowledge transfer to developing students’ technological literacy in an AI environment. Incorporating GenAI literacy into the core literacy system will help students adapt.

2.2 University students’ GenAI literacy

With the widespread application of generative artificial intelligence (GenAI) in education, how university students understand, use, and critically evaluate this emerging technology has become a significant topic in educational technology research. GenAI literacy refers to an individual's comprehensive ability to understand the fundamental principles of GenAI, acquire the skills to effectively use AI tools, critically assess AI-generated content, and maintain ethical awareness in their application (Liu et al., 2025). Compared to traditional AI literacy, GenAI literacy places greater emphasis on core dimensions such as prompt optimization, content evaluation, and creative application in human–computer interaction (Wang et al., 2025).

Bloom's Taxonomy of Educational Objectives (Anderson and Krathwohl, 2001) provides a theoretical framework for analyzing the structural layers of literacy, categorizing cognitive processes into six hierarchical levels: remember, understand, apply, analyze, evaluate, and create. Existing research on the dimensional structure of AI literacy has largely followed this theoretical trajectory, organizing competency elements from lower-order cognitive understanding to higher-order evaluation and creation. Ma and Chen (2024) developed an AI literacy scale comprising four dimensions—awareness, use, evaluation, and ethics—based on a sample of Chinese university students, and validated its applicability in educational contexts within developing countries. Weikang et al. (2025) synthesized six core constructs of AI literacy through a scoping review, namely identifying AI, understanding AI, AI ethics, AI empowerment, AI self-competence, and applying AI. Liu et al. (2025) further constructed a five-dimensional GenAI literacy measurement framework encompassing basic technical competence, prompt optimization, content evaluation, innovative application, and ethical compliance awareness.

Although these frameworks differ in the number of dimensions and their specific emphases, they all reflect a cognitive progression from “cognitive understanding” to “applied practice” and then to “evaluation and creation.” Given that some frameworks cover a relatively broad scope with certain conceptual overlaps among constructs, the present study draws upon the core structure of Ma and Chen’s (2024) four-dimensional framework, with adaptive modifications tailored to the specific characteristics of GenAI. This ultimately yielded a measurement framework comprising four dimensions: cognitive awareness, usage ability, evaluation ability, and ethical awareness. Together, these four dimensions span the full competency spectrum from cognitive understanding to applied practice and value judgment, ensuring both theoretical parsimony and empirical operability, while the contextual adaptations made to each dimension ensure adequate alignment with the GenAI learning environment.

2.3 Online learning resilience

The concept of resilience and its theoretical underpinnings originated in physics and ecology, and have since expanded into diverse fields including psychology, economics, and urban governance. At its core, resilience concerns how a system maintains its functions, recovers its state, or even achieves transformation in the face of disturbances, shocks, or pressures. In academic contexts, resilience is defined as students’ capacity to sustain high levels of academic engagement and the likelihood of success when encountering academic setbacks, stress, or failure. The academic resilience framework proposed by Martin and Marsh (2006) is among the most influential models in this field, conceptualizing academic resilience as a tripartite “cognitive–affective–behavioral” response system. At the cognitive level, it involves students’ interpretations and attributions of academic difficulties; at the affective level, it concerns emotion regulation capacity; and at the behavioral level, it manifests in sustained effort and active help-seeking. Based on this theoretical framework, Cassidy (2016) developed the Academic Resilience Scale (ARS-30), which, through exploratory factor analysis, yielded three stable factor structures—persistence, reflection and adaptive help-seeking, and negative affect and emotional response—corresponding respectively to the behavioral, cognitive, and affective dimensions of resilience theory.

For higher education, digital transformation is no longer an option but a core pathway. As online education grows globally, students’ educational journeys are shifting from the traditional physical classroom to the more open, connected, but also uncertain cyberspace (Hikmat et al., 2025). As university students embark on their digital learning journeys, the primary challenge is not only the acquisition of knowledge, but also how to psychologically adapt in such digital environments that lack immediate physical interactions and rely heavily on self-management (Warshawski, 2022). Online education is an inexperienced field for both students and educators, and as a result, students may have difficulty adapting to online learning, feel stressed and reduce their time commitment (Naidoo et al., 2022).

Online learning resilience is a natural extension of the ARS-30 theoretical framework when applied to online learning contexts. It is a dynamic adaptive capacity that enables students to rapidly restore psychological equilibrium and sustain academic pursuits when confronted with information overload, technical obstacles, and learning interruptions. In this study, online learning resilience was measured along the same three dimensions—learning persistence, reflection and help-seeking, and emotional reaction—which correspond respectively to students’ behavioral responses, cognitive strategies, and affective regulation when facing adversities in online learning.

Moreover, for learners, the reasons that determine their ability to persist in online learning are complex. Existing studies have explored to some extent how learners perceive the interactivity of the online environment, how they evaluate the atmosphere of the online learning platform, and their sense of mastery in the face of technological barriers, which are subjective perceptions that determine whether students choose to retreat negatively or respond positively when experiencing stress (Pan and Zhu, 2025). In addition, sustaining online learning resilience is also often closely related to an individual's ability to self-regulate (Ge, 2025). Good online social interactions and emotional feedback can act as a psychological buffer to enhance individuals’ resilience performance (Li et al., 2022). However, existing studies are still insufficient in exploring the mechanisms of sustained development of resilience in normalized online learning environments.

2.4 University students’ self-efficacy

Self-efficacy refers to an individual's beliefs about his or her ability to organize and execute a particular course of action to achieve a desired outcome (Bartimote-Aufflick et al., 2016). In higher education contexts, University students’ self-efficacy is widely understood as the degree of confidence students have in their ability to successfully complete learning tasks and reach academic goals in academic contexts (Greco et al., 2022). Research has shown that self-efficacy is an important variable in predicting academic achievement, academic engagement, and psychological adjustment among university students (Al-Abyadh and Azeem, 2022; van Zyl and Dhurup, 2018).

Self-efficacy in higher education is influenced by a combination of physical environment and non-physical psychological factors. While campus infrastructure and digital learning resources are important, students’ academic identity, psychological well-being, and overall socio-emotional experience are equally essential (Mao et al., 2023). Additionally, peer modeling, level of mastery of technological tools, and support from family and instructors can significantly influence the formation of college students’ sense of efficacy (Chuang et al., 2015; Zander et al., 2018). Taken together, moderately challenging and well-supported conditions tend to produce students who value deep learning and possess a high sense of academic efficacy (Al-Abyadh and Azeem, 2022), which is viewed as a key marker of a student's academic quality and influences subsequent choices of learning styles.

Existing research has amply revealed the far-reaching impact of GenAI from technological change to managerial reshaping (Sauvola et al., 2024; Taeihagh, 2025) and explored its positive effects on teacher development, reshaping of students’ learning paths, and restructuring of course evaluation models in the field of education (Fock and Siller, 2025; Tan, 2025; Yang and Lin, 2025). GenAI literacy among university students can also be reflected in multiple dimensions (Liu et al., 2025; Ma and Chen, 2024). However, whether and to what extent GenAI literacy can enhance college students’ ability to cope with digital learning dilemmas remains an urgent question to be answered in the context of online learning, which has become the norm in higher education. In addition, research on university students’ self-efficacy has developed more mature theoretical frameworks and measurement tools (Chuang et al., 2015; Greco et al., 2022), and has confirmed its positive impact on academic achievement, engagement in learning, and psychological resilience (Al-Abyadh and Azeem, 2022; Mao et al., 2023). However, at a time when GenAI is rapidly integrating into learning, the formation of self-efficacy and how it further acts on online learning resilience when college students use GenAI tools to accomplish learning tasks remains to be explored in depth.

This study attempts to construct an integrative theoretical framework that systematically examines the role of GenAI literacy as an antecedent variable, online learning resilience as an outcome variable, and self-efficacy as a mediating variable. Through this research design, this study aims to provide a more complete theoretical explanation for understanding the adaptive mechanism of university students’ online learning in the digital era, and to provide practical guidance for colleges and universities to carry out GenAI literacy education and enhance university students’ online learning resilience.

3 Research hypothesis

Social Cognitive Theory posits that human behavior is the product of the interplay among personal factors, behavior, and the environment—a formulation known as “triadic reciprocal determinism"—which illuminates individuals’ psychological dynamics and behavioral logic in changing contexts (Bandura, 1986). The core proposition of the theory is that individuals’ capabilities and psychological beliefs jointly determine their behavioral performance, while the environment exerts an indirect influence on behavior by affecting the individual. As a novel competency for adapting to digital-intelligent environments, GenAI literacy can directly drive persistent behavior in online learning, and can also be indirectly translated into behavioral motivation by shaping self-efficacy—a psychological belief about one's own capabilities. Accordingly, guided by SCT, this study constructs a logical framework of “competence → psychology → behavior” within the digital-intelligent learning context (See Figure 1 for details).

3.1 University students’ GenAI literacy and online learning resilience

Taeihagh (2025) notes that with the rapid proliferation of generative artificial intelligence (GenAI) technologies, these new tools will reshape students’ learning trajectories. As digital learning environments become the dominant form of higher education, individuals’ GenAI literacy is crucial for influencing their learning perceptions and addressing complex challenges. From the perspective of the relationship between individual competence and behavioral performance, GenAI literacy constitutes the most fundamental competency for students in digital-intelligent learning contexts—it determines whether students can effectively utilize technological resources in the environment to cope with uncertainties and difficulties in learning.

GenAI literacy is not only a core driver for stimulating learning curiosity and maintaining academic engagement; by enhancing their AI literacy, students can also strengthen their belief in learning and improve their actual outcomes (Kong and Yang, 2024). Generative AI platforms provide students with more convenient ways to break down complex tasks, discuss academic issues, and connect instantly with intelligent scaffolding. This shift in learning communication makes knowledge acquisition faster and more effective than traditional search engines or static textbooks, significantly reducing the time and cognitive effort required to overcome academic bottlenecks (Pan and Zhu, 2025). Furthermore, the personalized nature of GenAI-generated content serves as a bridge between users and technology; through these interactions, what was once a tedious online learning process becomes more relatable and memorable (Warshawski, 2022).

GenAI possesses the potential to fundamentally transform learning; it is not merely a learning tool but an open channel for information sharing (Tan, 2025). If students can proficiently utilize GenAI, they can gain access to more success stories and accounts of overcoming setbacks in online learning, which may enhance the appeal of online learning. Kong and Yang (2024) further note that GenAI can amplify the inherent openness and sharing attributes of online learning, helping to build confidence among students facing learning plateaus. This also reflects the inclusivity and sense of support that digital courses provide to all learners. High-quality interactions facilitated by GenAI help college students feel a greater sense of control and connection within the online environment, making them more likely to persist when encountering difficulties (Pan and Zhu, 2025). Precisely because GenAI literacy provides a solid capability foundation for students’ online learning behaviors, when faced with technical failures or comprehension difficulties, students with such literacy tend to persist rather than give up easily.

Therefore, we propose the following hypothesis:

H1: University students’ GenAI literacy is positively correlated with online learning resilience.

3.2 University students’ GenAI literacy and self-efficacy

According to Social Cognitive Theory, self-efficacy does not emerge in a vacuum but is rooted in individuals’ successful experiences in specific domains—namely, enactive mastery experiences. GenAI literacy precisely creates numerous such successful experiences for students: by effectively using AI tools to complete learning tasks, students repeatedly receive positive feedback that “I can solve problems with the help of new technologies,” and these experiences gradually accumulate and are internalized into stable judgments about their own academic capabilities. In other words, GenAI literacy enables students to continually obtain validation of their capabilities through interaction with the digital-intelligent environment, which is the fundamental pathway for the formation and strengthening of self-efficacy. An individual's competence is crucial for building academic confidence, particularly for college students navigating the transformation of digital education. In learning environments driven by big data and artificial intelligence, students’ assessments of their own abilities are often influenced by their proficiency in using cutting-edge tools (Greco et al., 2022). Strong GenAI literacy can engage students more actively in learning, thereby helping them progressively achieve their academic goals and personal development aspirations, and fostering a higher sense of self-efficacy (Wang et al., 2025). GenAI literacy lays the cognitive foundation for enhancing academic self-efficacy by bridging the gap between expectations and capabilities.

Furthermore, in high-quality interactions between students and GenAI, clear communication and effective feedback can significantly strengthen their belief in their own academic abilities, thereby enhancing self-efficacy. Kong and Yang (2024) note that communicating with GenAI in a clear and effective manner helps foster a sense of academic belonging and task competence. Research by Wang et al. (2023) further confirms that the effective utilization of digital resources is a key factor in validating students’ learning abilities and shaping positive self-evaluations. GenAI is not merely an information retrieval tool but a collaborative partner capable of providing instant feedback and guiding reflection. When students receive answers or learning support that meet their expectations through precise prompts, they attribute this success to their own abilities, thereby reinforcing the mental representation that “I am capable” and driving the continuous enhancement of academic self-efficacy.

GenAI literacy is not merely a technical skill but also a form of psychological capital; through the accumulation of students’ successful experiences using AI, it is internalized as enduring academic self-efficacy. GenAI literacy is a key component of learning psychology; as a “cognitive cue,” it helps learners predict academic outcomes, develop inquiry behaviors, and maintain academic persistence (Bartimote-Aufflick et al., 2016). Students evaluate their past experiences of successfully applying GenAI, and these evaluations directly influence their academic self-efficacy (Alrashidi and Alshammari, 2025; Xu et al., 2022). When students repeatedly experience the positive cycle of “successfully using GenAI to solve academic challenges,” this experience transforms into a stable psychological resource, enabling them to remain confident when facing new learning challenges.

Based on the above literature and theoretical background, we propose the following hypothesis:

H2: University students’ GenAI literacy positively influences their academic self-efficacy.

3.3 University students’ self-efficacy and online learning resilience

The influence of self-efficacy on behavior is not merely a cognitive perception but permeates the entire process of behavior—it determines whether individuals will attempt a task when facing difficulties, how much effort they are willing to invest, and how long they can persist. In the context of online learning, this influence is particularly prominent. When encountering adversities such as technical failures, information overload, or separation from instructors, students with high self-efficacy do not interpret difficulties as a sign of personal inadequacy, but rather as temporary obstacles that can be overcome through their own efforts. It is precisely this difference in interpretation that leads to fundamentally different behavioral responses: those with low self-efficacy tend to withdraw or give up, whereas those with high self-efficacy proactively adjust their learning strategies, seek alternatives, and recover more quickly from setbacks. In short, self-efficacy provides students with a sense of psychological assurance that “I am capable of coping,” and this assurance directly shapes their specific behavioral choices and the degree of persistence in the face of adversity. Learners first form opinions and attitudes about their abilities, which then further shape their learning behaviors (Woreta et al., 2025). In online learning environments, college students face unique challenges such as technological failures, information overload, and separation of teachers and students. Students with high self-efficacy are more inclined to view difficulties as manageable and surmountable rather than threatening. This positive cognitive appraisal motivates them to proactively adopt alternative learning strategies (e.g., changing platforms, finding supplemental resources), adjusting their learning methods, and resuming their efforts more quickly after a setback. If a student holds a positive appraisal of their abilities in a digital environment, they will demonstrate persistence through goal setting, strategy selection, and emotional arousal when faced with an online learning dilemma (Alrashidi and Alshammari, 2025).

Self-determination theory further reveals the underlying dynamics from the perspective of basic psychological needs: individuals with higher self-efficacy perceive that they are able to effectively master learning content and cope with technological challenges, and the satisfaction of this competence need enhances their autonomy and intrinsic motivation, which in turn serves as the most stable source of sustained behavior. In higher education, self-efficacy reflects the fact that college students show preference and adaptation to online learning modes due to their competencies. This intrinsic motivation is an important driver of enduring learning behaviors and resilience. In the face of adversity such as network interruptions, time conflicts, or lack of face-to-face supervision, students with high self-efficacy are intrinsically motivated to seek social support (e.g., seeking help from classmates online, contacting the instructor), self-regulate their pace of learning, and find a sense of meaning in the midst of difficulties (Li et al., 2022; Warshawski, 2022).

Therefore, we propose the following hypothesis:

H3: University students’ self-efficacy positively affects their online learning resilience.

3.4 The mediating role of self-efficacy

Individual competence is undoubtedly important, but competence itself is not the sole driver of persistent behavior in the face of adversity, nor does it automatically translate into action—it must be “activated” at the psychological level, that is, transformed into an individual's belief in their own capacity to employ that competence. This is precisely the core insight that distinguishes SCT from a simplistic linear model in which “competence directly determines behavior.” In this study, if the competency advantages represented by GenAI literacy are to translate into sustained behavior when confronting online learning difficulties, a critical step of “psychological internalization” must occur: the enactive mastery experiences accumulated through students’ repeated successful use of GenAI to solve learning problems need to be attributed to their own capabilities and thereby transformed into stable self-efficacy beliefs. Only after this belief is formed will students choose to persist rather than withdraw when encountering real difficulties. Existing research findings suggest a complex psychological transmission mechanism between the availability of technological tools and students’ academic resilience (Bartimote-Aufflick et al., 2016). GenAI literacy is not synonymous with resilience, but rather requires a process of ‘psychological internalization’. Attribution theory suggests that the way individuals attribute success or failure outcomes significantly influences their subsequent motivation and behavior. High levels of GenAI literacy can provide students with a “sense of mastery” cue, which is solidified by increased self-efficacy, making students more inclined to attribute failures in e-learning (e.g., missing the point due to video lag, unsatisfactory scores on an online test) to temporary or controllable factors, which can lead to greater motivation and behavioral changes when they encounter bottlenecks in e-learning. more aggressive defensive and recovery behaviors when experiencing online learning bottlenecks (Alrashidi and Alshammari 2025).

As mentioned earlier, GenAI enables immediate feedback that may shapes students’ perceptions of competence in real time through the quality of effective technological interactions (Greco et al., 2022), thus enabling students to objectively evaluate immediate difficulties and avoid falling into self-doubt, thereby protecting their learning self-esteem and future expectations and reducing feelings of helplessness and the tendency to give up. In contrast, low self-efficacy individuals are prone to attribute failures to a lack of competence, which creates a negative cycle and weakens resilience. This is a reflection of the ability of high GenAI literate individuals to utilize AI resources for self-regulation (Ng et al., 2024). Taken together, the complete transmission pathway of this study is as follows: GenAI literacy creates abundant enactive mastery experiences for students; these experiences, attributed to students’ own capabilities, are internalized into self-efficacy beliefs, which in turn motivate students’ persistence and adaptive behaviors when confronting difficulties. Within this chain, self-efficacy plays an irreplaceable mediating role.

Therefore, we propose the following hypothesis:

H4: University students’ self-efficacy mediates the relationship between GenAI literacy and online learning resilience.

The theoretical model of this study is as Figure 2.

4 Data and results

4.1 Data collection and sampling

In this study, college students from five colleges and universities in the Xinjiang region of China were selected for the study using non-probability sampling methods, especially convenience sampling. The reason for choosing these five institutions is that they are in the critical period of digital transformation, and their students’ performances in generative AI literacy and online learning resilience are highly homogeneous, which can provide good empirical samples for the theoretical model of this study. To ensure the scientific rigor of the questionnaire survey, this study employed a screening question—"Do you know about and have you used GenAI tools?"—as a control measure to minimize potential interference in the measurement of the GenAI literacy scale, thereby enabling it to effectively reflect individual differences in students’ literacy levels. Between 2024 and 2025, a total of 350 questionnaires were distributed, and 306 valid responses were collected, far exceeding the minimum sample size required for a medium effect size according to G\*Power, and also meeting the requirements for the SEM approach adopted in this study. Among the participants, male students accounted for 46.4% and female students accounted for 53.6%, the age distribution was dominated by 18–22 years old (81.4%), humanities and social sciences majors accounted for 47.1%, and science and technology majors accounted for 52.9% (see Table 1 for details).

Table 1. Basic information description.

4.2 Measuring tool

The questionnaire consisted of two main sections: (1) Demographic information: this section collected basic demographic background information about the participants. (2) Rating scales: this section focused on measures of GenAI literacy, self-efficacy, and online learning resilience. All scales were developed based on the existing established literature and revised appropriately for this research context (see Table 2 for details).

Table 2. Rating scale.

GenAI literacy was measured using 12 items. The scale was adapted from the four-dimensional framework of Ma and Chen (2024) and contextually adjusted to the specific characteristics of GenAI. The cognitive awareness dimension (CAW1–CAW3) focuses on students’ ability to distinguish GenAI from traditional tools and understand its learning-supportive functions, with CAW2 being reverse-coded. The usage ability dimension (UAB1–UAB3) concerns students’ operational proficiency and perceived efficiency in using GenAI tools to support learning, with UAB2 being reverse-coded. The evaluation ability dimension (EAB1–EAB3) measures students’ critical assessment of GenAI-generated content and their ability to select appropriate tools. The ethical awareness dimension (EAW1–EAW3) focuses on academic integrity, information security, and risk vigilance, with EAW2 being reverse-coded.

The online learning resilience scale was adapted from Cassidy’s (2016) ARS-30 framework and contextualized to the online learning setting. The learning persistence dimension (LEP1–LEP3) measures students’ persevering behaviors when encountering technical failures or comprehension difficulties. The reflection and help-seeking dimension (REH1–REH3) measures students’ tendency toward self-reflection and active help-seeking. The emotional reaction dimension (EMO1–EMO3) measures students’ emotion regulation capacity when facing learning setbacks. These three dimensions correspond respectively to the behavioral, cognitive, and affective levels in resilience theory.

The self-efficacy scale adopted established measures from Pintrich et al. (1991) and Zimmerman and Kitsantas (2005), with a unidimensional structure consisting of five items (SEF1–SEF5). The scale measures students’ confidence in their own learning abilities in academic contexts, with all items being positively worded.

All scales were scored on a 5-point Likert scale from 1 (strongly disagree) to 5 (strongly agree).

Finally, all the constructs in this study are reflective constructs. In reflective conceptualization, indicators (items) are the result of conceptualization, i.e., indicators are determined by conceptualization. This means that the existence of the conceptualization precedes the indicators, the changes in the indicators reflect the changes in the conceptualization, and there is internal consistency among the indicators. For example, a high level of GenAI literacy will be manifested by students’ high scores in the dimensions of cognitive awareness, usage ability, evaluation ability and ethical awareness.

4.3 Data analysis

In this study, covariance-based structural equation modeling (CB-SEM) was used and AMOS 24.0 software was selected for model validation and hypothesis testing. In the process of data analysis, this study used a two-stage analysis method. The first stage focuses on assessing the validity and reliability of the measurement model. For the reflective constructs, this study examined the factor loadings, Cronbach's alpha coefficients, combined reliability (CR), and average variance extracted (AVE). Based on this, discriminant validity tests as well as the HTMT criterion were used to assess the discriminant validity between the constructs. The second phase focused on path analysis and hypothesis testing of the structural model.

First stage evaluation:

The assessment results of the measurement model in the first stage are detailed in Table 3, which focuses on testing whether the reliability and validity indicators of the constructs meet acceptable standards. Since all the constructs in this study are reflective constructs, factor loadings, internal consistency reliability, convergent validity and discriminant validity were systematically tested.

Table 3. Measurement model evaluation.

In the test of reflective constructs, all indicators met the recommended criteria:

The factor loadings were all greater than 0.6;

Cronbach's alpha coefficient and combined reliability (CR) were greater than 0.8;

Average Variance Extracted (AVE) were all greater than 0.5;

The variance inflation factors (VIF) are all less than 3, indicating that there is no serious multicollinearity problem.

Given that all respondents to the questionnaire were university students, the single-source nature of the data may raise concerns about common method bias. To mitigate this potential bias, several procedural measures were implemented during data collection: respondents completed the questionnaire anonymously, items were worded concisely, and the ordering of questionnaire sections was rearranged. In addition, two statistical approaches were employed to test for common method bias. First, Harman's single-factor test was conducted through exploratory factor analysis on all variables, which extracted eight factors, with the first factor accounting for 31.013% of the total variance—well below the 40% threshold. Second, a single-factor confirmatory factor analysis was performed. Compared with the original measurement model, the single-factor model failed to meet the recommended fit indices and fell considerably short of the acceptable criteria. These results suggest that common method bias does not pose a serious threat in this study.

On this basis, the discriminant validity test of Fornell and Larcker (1981) and the HTMT criterion were used to further validate the discriminant validity of the inter-conceptualization (see Tables 4, 5 for details). The test results show that all measures have good convergent validity and reliability, and there is sufficient discriminant validity among the constructs to meet the prerequisites of structural equation modeling analysis.

Table 4. Discriminant validity (Fornell-Larcker test).

Table 5. The HTMT criterion.

The results of the path analysis of the structural model in the second stage will be used to verify the hypotheses proposed in this study, and the specific results will be presented in detail in the subsequent chapters.

Second stage evaluation:

In this study, both GenAI literacy (GAI) and online learning resilience (NLR) are multidimensional abstract constructs that are defined by their respective first-order dimensions; accordingly, a second-order reflective modeling strategy was adopted. On the one hand, the second-order model subsumes correlated first-order dimensions under a higher-order latent construct, thereby more accurately reflecting the theoretical nature of the construct (Gerbing and Anderson, 1988). On the other hand, incorporating multiple first-order dimensions directly into the structural model would result in an excessive number of paths and unnecessarily high model complexity; the second-order approach effectively simplifies the model structure and reduces the number of parameters to be estimated. Statistically, the correlation coefficients among the four dimensions of GAI and among the three dimensions of NLR were all at moderate levels, meeting the basic requirement for second-order factor analysis regarding the interrelatedness among first-order dimensions.

In AMOS, estimation of second-order constructs was performed using a two-stage approach: first, factor scores for each first-order dimension were obtained through first-order confirmatory factor analysis; these factor scores were then used as observed indicators for the second-order constructs in model estimation. Model identification was achieved by fixing the path coefficient from the second-order factor to the first dimension to 1 (GAI fixed to CAW, NLR fixed to LEP). Through this method, the contribution of each first-order dimension to the second-order construct was obtained. In this study, all dimensional path coefficients were significant and similar in magnitude, confirming the reasonableness of the dimensional structure. This study constructed a model of the influence relationship between GenAI literacy, self-efficacy, and online learning resilience, and the path of influence structural equations between the variables is shown in Figure 3.

As can be seen from Table 6, the ratio of chi-square value to degrees of freedom is 1.287 < 3.000, the RMSEA value is 0.031 < 0.080, the SRMR value is 0.050 < 0.080, the fitting indexes of GFI, CFI, IFI, TLI are all greater than 0.90, and PNFI is greater than 0.50. All fitting indexes meet the recommended standards, which indicates that the model fits well in general. The situation is better, the fit between the hypothesized theoretical model and the actual data is higher, and the model results are more convincing.

Table 6. Optimization analysis of structural model fitting.

Since both the independent variable GAI and the dependent variable NLR in this study were treated as second-order reflective constructs, a two-stage approach was used for model estimation. In the first stage, factor scores for each first-order dimension were obtained through validated factor analysis; in the second stage, these first-order dimension factor scores were used as reflective indicators of second-order constructs for model estimation. Table 7 presents the path coefficients between the first-order dimensions and the second-order latent variables. It can be seen that the path coefficients of the four first-order dimensions of GAI—CAW, UAB, EAB and EAW-are all significant and of similar values (0.585–0.653), suggesting that these four dimensions play an equally important role in forming GAI; the path coefficients of the three first-order dimensions of NLR—LEP, REH and EMO—are all significant and have similar values (0.584–0.632), indicating that these three dimensions have an equally important contributions.

Table 7. Optimization analysis of structural model fitting.

The symbol “**” denotes statistical significance at the 0.001 level (P* < 0.001).

The results of the final structural model test are listed in Table 8. It can be seen that the standardized path coefficient of GAI on NLR is 0.666 (p < 0.001), which reaches the significance level, and H1 is supported, i.e., the higher the level of university students’ GenAI literacy, the more resilient they are in online learning. The standardized path coefficient of GAI on SEF is 0.759 (p < 0.001), which reaches the significance level, and H2 is supported, i.e., the university students’ GenAI literacy has a significant positive effect on their academic self-efficacy. The standardized path coefficient of SEF on NLR is 0.418 (p = 0.004), which reaches the level of significance, and H3 is supported, i.e., the self-efficacy of university students has a significant positive effect on their online learning resilience. By comparing the path coefficients, it can be found that GAI has the strongest effect on SEF, followed by its direct effect on NLR, and again by SEF on NLR.

Table 8. Basic path verification of the model.

The symbol “**” denotes statistical significance at the 0.001 level (P* < 0.001).

The mediating effect of self-efficacy was tested with reference to the Bootstrap method proposed by Hayes (2017). The method requires the selection of Bootstrap sample size of 5000, and if the mediation test result under the 95% confidence interval does not contain 0, it means that the corresponding indirect effect significantly exists, and the test results are listed in Table 9. It can be seen that the 95% confidence interval results of 5,000 iterations of Bootstrap mediated effects test do not contain 0, and all pass the significance test. In the indirect effect test, the indirect effect value of GAI on OLR through SEF is 0.292, with a confidence interval of [0.220, 0.370] and does not contain 0, indicating the existence of a significant indirect effect; in the direct effect test, the value of the direct effect of GAI on OLR is 0.344, with a confidence interval of [0.235, 0.452] and does not contain 0, indicating the existence of a significant direct effect.

Table 9. Bootstrap mediation effect test.

The paths of influence of each first-order dimension of GenAI literacy on each dimension of online learning resilience through self-efficacy were further examined. The results showed that the indirect effects of CAW through LEP, REH, and EMO were 0.153, 0.137, and 0.179, respectively, with confidence intervals that did not include 0. The indirect effects of UAB through SEF were 0.259, 0.233, and 0.249, respectively, with confidence intervals that did not include 0. The indirect effects of EAB through SEF on LEP, REH, EMO through SEF were 0.150, 0.141, and 0.149, respectively, with confidence intervals that did not include 0; and the indirect effects of EAW through SEF on LEP, REH, and EMO were 0.138, 0.129, and 0.160, respectively, with confidence intervals that did not include 0. All indirect effect paths held significantly. In summary, the bootstrap results confirm that self-efficacy significantly mediates the relationship between GenAI literacy and online learning resilience, with both direct (β = 0.344, p < 0.001) and indirect effects (β = 0.292, p < 0.001) being statistically significant. These findings provide empirical support for H4.

5 Discussion

This study positions GenAI literacy as a dynamic competency rather than a mere technical operational skill, centering on the adaptability and generativity of university students during their continuous interaction with AI tools. The study found that the four dimensions of GenAI literacy—cognitive awareness, usage ability, evaluation ability and ethical awareness—together constitute a framework of foundational competencies for college students to meet the challenges of digital learning. This ability to proactively identify, access, integrate, and optimize digital resources based on environmental changes distinguishes GenAI literacy from traditional static technology manipulation abilities. The results suggest a “literacy-efficacy-resilience” pathway, with self-efficacy serving as a significant mediating mechanism. This pattern is consistent with self-determination theory, which proposes that satisfying individuals’ intrinsic psychological needs fosters intrinsic motivation and behavioral persistence. The findings also align with and extend Social Cognitive Theory by suggesting that self-efficacy serves as a key psychological link between technical competence and learning resilience.

This mechanism also offers a perspective on understanding the “technology adoption gap” in digital transformation: why many university students have a wealth of digital tools at their disposal, but struggle to keep up with the pressure to learn. For example, while Ge (2025) emphasized the role of AI literacy in monitoring learning processes and reducing cognitive load, the present findings suggest that self-efficacy may function as a psychological transmission mechanism—that is, the association between technological literacy and learning resilience may be contingent upon learners’ positive beliefs about their own abilities. Without the catalyst of this intrinsic sense of efficacy, it may be difficult to sustain learning engagement in the face of adversity, even with a wealth of tools and operational skills. The observed direct and indirect effects are consistent with the view that learning resilience in digital environments is closely related to individual psychological capital.

In addition, the present study identifies self-efficacy as a key psychological resource connecting technological literacy and learning resilience, which is consistent with the perspectives of existing research. Kong and Yang (2024) state that fostering positive attitudes towards technology among students is a prerequisite for realizing the potential of GenAI. The present findings offer support for this view by suggesting that digital literacy may need to be internalized through self-efficacy to be associated with sustained resilience in digital learning contexts. GenAI is deeply embedded in complex online learning environments, and the realization of its educational value depends on students’ comprehensive understanding and mastery of the technology. Students with high GenAI literacy tend to integrate technological tools into personalized learning strategies more effectively, and self-efficacy may serve as a source of psychological motivation in this integration process, which appears to be associated with greater academic persistence and adaptability in digital learning environments.

The findings of this study are in line with UNESCO's (SDG 4) view that technology-enabled learning should focus on equity and accessibility. In the wave of digital transformation in higher education, GenAI literacy, as a key factor influencing students’ learning resilience, can help to narrow the learning opportunity divide due to differences in technological conditions. The findings also point to the notion that self-efficacy represents cultivatable psychological capital, and that efforts to promote digital education may benefit not only from introducing advanced technological tools but also from building an educational ecology centered on competency development to help students establish positive beliefs about their own learning abilities. The findings also encourage a shift in perspective in educational technology research—from a primary focus on technology platforms toward a greater attention to individual learners as active agents in human-AI collaboration, offering insights for exploring sustainable digital learning support systems.

6 Conclusions and recommendations

6.1 Conclusions

This study enriches the existing literature in the following ways. On the one hand, it clarifies the core logic of the transformation of technological literacy into learning resilience at the micro level. Existing studies have mostly emphasized the importance of GenAI literacy in identifying and grasping technological opportunities, but have neglected how GenAI development can positively impact college students’ online learning. This study provides empirical evidence that GenAI technology development alone does not naturally lead to university students’ online learning resilience, and that possessing technological literacy and constructing self-efficacy are key to transforming technological competence into enduring learning motivation.

On the other hand, from the perspective of sustainable educational provision, the results of this study suggest that GenAI literacy plays an important role in achieving scalability, accessibility and cultural adaptability of digital learning services. When students are able to effectively integrate technological tools with their own learning needs in order to maintain positive self-beliefs in the face of setbacks, GenAI literacy can truly contribute to educational equity and learning sustainability. This study also makes it clear that self-efficacy is a fundamental element to support sustainable digital learning, and that sustainable digital education cannot rely on the natural development of technology, but should also actively intervene in the construction of individual technological literacy and psychological capital.

6.2 Recommendations

The practical significance of this study is not only to make general educational recommendations, but also to highlight the unique requirements for sustainable digital education development in the context of GenAI development. As the learning process is deeply embedded in technological literacy, individual emotions and cognition, the relevant policies should not only focus on the technology itself, but also on the overall construction of the educational support system.

Institutions of higher education need to strengthen individual-level training, offer more forward-looking digital literacy courses, teach students to use GenAI tools efficiently, and help students improve their self-efficacy and adaptability in diverse learning situations through simulated human-computer collaborative learning programs. At the same time, colleges and universities should establish a systematic academic support mechanism to strengthen students’ confidence and resilience in the face of learning challenges through timely feedback, peer support and personalized guidance.

The government and education authorities should build a learning quality assurance framework adapted to the digital era, establish a GenAI literacy evaluation system for students, create a flexible policy environment for the development of digital education, and lower institutional barriers for students in the process of technology adaptation. At the same time, relevant policies to support digital learning standards, mutual recognition of curricula and data governance should be introduced to enhance the replicability and scalability of digital learning services in different educational contexts.

In addition, in order to achieve long-term sustainable development, education technology platforms and related enterprises should not only focus on the construction of technical infrastructure, but also provide professional psychological and learning support for students, monitor learning effectiveness and establish a service improvement mechanism based on user feedback. A platform that combines technological tools with continuous user support can more effectively provide students with a diversified, high-quality learning experience.

In short, a sustainable digital education ecology requires the collaborative efforts of universities, governments and digital platforms. The online learning resilience of university students is the result of the synergy of technology application, talent capacity, institutional framework and teaching support. Only when these elements support each other can GenAI literacy truly realize the development goals of digital inclusion and educational equity.

6.3 Limitations

Several limitations of this study should be acknowledged. First, the sample consisted of 306 university students recruited via convenience sampling from five institutions in the Xinjiang region of China. Although this sampling strategy is commonly employed in exploratory educational research and facilitates the collection of preliminary evidence under resource constraints, the inherent selection bias of non-probability sampling may compromise the representativeness and external validity of the findings. The Xinjiang region possesses distinct cultural characteristics, levels of economic development, and educational policy contexts that may differ from other regions in China. Consequently, caution is warranted when generalizing the conclusions of this study to the broader population of Chinese university students. Future research should employ probability sampling methods with larger sample sizes and cross-regional coverage to enhance the generalizability of the findings. Second, this study adopted a cross-sectional survey design, in which all variables were measured at a single time point. This design limits the capacity to draw causal inferences regarding the relationships among the variables. Future research could employ longitudinal tracking designs or experimental designs to more rigorously examine the causal directions among the variables through time-series data or intervention experiments. Third, although the scales used in this study were adapted from well-established literature and subjected to expert review and pilot testing to ensure content validity, GenAI technology remains in a phase of rapid development, with its functions and application scenarios continuously evolving. The dimensions currently measured by the scales—cognitive awareness, usage ability, evaluation ability, and ethical awareness—may not fully capture the emerging literacy requirements brought about by future developments in GenAI technology. In addition, students with different disciplinary backgrounds or varying levels of technological familiarity may differ in their understanding of GenAI-related items. Future research should continuously revise and refine the measurement instruments in light of the latest developments in GenAI technology, and explore differentiated measurement strategies tailored to diverse student populations.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Bibhya Sharma, University of the South Pacific, Fiji

Reviewed by: Ahmad Syamil, Binus University, Indonesia

Thean Pheng Lim, Tunku Abdul Rahman University College, Malaysia