The Stimulus-Organism-Response (SOR) model was proposed by Mehrabian (1974), as a theoretical framework for exploring the relationship between external stimuli and organismic responses. This model posits that behavioral performance is not merely a stimulus–response paradigm, but rather, it is achieved through cognitive processing by the organism to elicit a specific response. Specifically, stimuli (S) is defined as the various types of external factors in the environment that influence the internal mental state or cognitive processes of the organism (O), ultimately leading to a specific behavioral response (R). These stimuli activate various internal processes, including cognition, affect, and evaluation, ultimately determining the organism’s behavioral response (Xie et al., 2023). The concept of an organism’s perception establishes a link between stimulus and response, thereby explaining the process by which an organism is stimulated and responds.

While established models like TAM effectively explain technology acceptance through perceptual factors like usefulness and ease of use (Al-Adwan et al., 2023; Zou et al., 2025), the SOR model offers a more comprehensive perspective. It captures not only external technological characteristics but also the crucial mediating role of users’ internal states—particularly their cognitive processing (Fu et al., 2025). This model provides a more nuanced understanding of how external environmental stimuli interact with individual cognition and evaluation to shape behavioral responses (Chen, 2023; Liu Y. F. et al., 2023). Consequently, the SOR model is increasingly applied within online learning, including e-learning platforms (Fu et al., 2025), AI teaching assistants (Peng and Wan, 2024), and mobile-assisted language learning (Lee and Xiong, 2023).

This range of applications demonstrates the SOR model’s validity for GAI online learning contexts. It not only aids in understanding how GAI features (stimuli) and users’ internal psychological processes (organism), but also provides actionable recommendations for optimizing the GAI online learning experience and enhancing users’ infusion use.

Our study employed a mixed-methods approach to explore the infusion use of GAI in online learning scenarios, following established methodological guidelines (Venkatesh et al., 2016; Creswell and Creswell, 2018). The steps for an exploratory sequential design are as follows: First, the qualitative stage involved 26 semi-structured interviews to explore factors influencing users’ infusion use of GAI. We proposed our research model and hypotheses based on the results of stage 1. Then, stage 2 tested the hypotheses through an online survey. This study is conducted in two stages, as illustrated in Figure 1.

The mixed-methods approach is particularly appropriate for our study. First, this design offers advantages over single-method approaches by simultaneously addressing both confirmatory and explanatory research questions (Creswell and Creswell, 2018). It aligns perfectly with the dual nature of our investigation. Second, the application of GAI in online education context presents a degree of novelty, making it difficult for existing theories to provide a thorough description and explanation of the issues (Venkatesh et al., 2016).

This study employed semi-structured interviews conducted through face-to-face interviews and online video conferences. The research group first screened participants with prior experience using GAI, who usually use GAI as their main learning tool and can express their opinions based on their experiences. Participants were recruited using purposive sampling techniques to ensure their suitability for the study topic. All participants were requested to indicate their willingness to participate in semi-structured interviews. Among them, 13 were female (50%). The average age of participants was 29.77 (SD = 8.75). Most participants were young and middle-aged individuals under 35, aligning with QuestMobile’s finding that AI users are predominantly concentrated in this age group.²

Before the interviews, the researcher established several guidelines with participants, including encouraging open sharing, preventing interruptions, and ensuring the anonymity of all information, to foster an open and secure communication environment. During the interviews, the researcher initially collected basic demographic information and explained GAI. Participants were then asked to describe specific examples of using GAI for learning. The interviews focused on the role of GAI in the participants’ learning, their most memorable experiences, challenges encountered in using GAI, and the subsequent impacts. The detailed interview protocol is shown in Appendix A.1. Each interview lasted 20–30 min, with all participants agreeing to audio recording. After the interviews, the audio recordings were transcribed verbatim into textual material for subsequent qualitative analysis. Data were collected between December 2024 and January 2025. Interviews ceased upon reaching data saturation, defined as the point when no significant new information emerged from the data (Shao et al., 2024). Demographic details of the participants are presented in Table 1.

First, the semi-structured interview transcripts were pre-processed to remove content unrelated to GAI in online learning scenarios. Second, semantically ambiguous or irrelevant content was eliminated. Next, responses reflecting interviewees’ misinterpretation of the questions were excluded. Finally, the transcripts were coded and labeled line-by-line according to the logical sequence of the interview content. According to Liu Y. L. et al. (2023), we adopted thematic analysis combining “top-down” framework analysis (Ritchie and Spencer, 2002) and “bottom-up” grounded theory (Glaser and Strauss, 2017). We analyzed the interview data by coding, theming, decontextualizing, and recontextualizing.

This study utilized NVivo 11.0 software to conduct a thorough analysis of the interview data, examining the content word-by-word, sentence-by-sentence, and paragraph-by-paragraph. The researchers were instructed to use original phrases from the interview transcripts for labeling during the coding process. Two graduate students subsequently organized the data based on the initial nodes. Throughout this process, the original information was continuously compared and revised, with all meaningful themes and concepts being precisely extracted.

To minimize researcher bias in the coding process, two graduate students independently coded the data through a back-to-back approach. Both coders were native Chinese speakers and familiar with GAI. Before starting the coding process, a centralized meeting was conducted to align the coders on the procedures and clarify relevant concepts and theories. After each coding round, the results from both coders were compared, the same initial concepts were merged, and the coding conflicts were discussed with experts. Finally, the concepts were summarized and reorganized. After repeated discussions, only concepts agreed upon by both coders were retained. The complete coding process is presented in Appendix A.2.

Through semi-structured interviews, we identified several factors influencing the infusion use of GAI, such as intelligence, explainability, response time, integrability, accuracy, source credibility, personalization, emotional support, perceived value, and satisfaction. Building upon these exploratory findings, we subsequently designed a quantitative study to empirically validate the hypothesized relationships between these variables. Analysis of interview text reveals that users primarily employ ChatGPT, Doubao, ERNIE Bot, Deepseek, Gemini, Kimi, and other GAI platforms, reflecting the diversity of GAI tools within online learning contexts.

Intelligence reflects GAI’s capability through environmental perception, adaptive learning, problem-solving, and goal attainment. This characteristic of continuous evolution through feedback leads users to recognize it as a genuine intelligence (Moussawi et al., 2023). As the most critical factors of AI technology (Bartneck et al., 2009), intelligence fundamentally relies on natural language processing technologies that enable AI to simulate human cognitive processes in language comprehension and production (McLean et al., 2021). The statement is echoed by the qualitative investigation. For example, interviewee (P6) mentioned: “Because I think ChatGPT records all my previous conversations, I think its answer will be more professional and more in line with my heart.” GAI exemplifies this intelligence through its extensive knowledge repository and professional response capabilities, delivering not merely accurate and compelling answers but also formulating solutions that satisfy users through concise and coherent language outputs (Priya and Sharma, 2023). During interactive processes, GAI exhibits a high level of attentiveness to the users’ needs while employing diverse response strategies, which significantly enhance users’ experience and foster positive attitudes toward the technology.

Priya and Sharma (2023) pointed out that intelligence manifests in three critical dimensions of information generation: effectiveness, efficiency, and reliability. These dimensions not only constitute the fundamental drivers of GAI advancement but also serve as critical factors shaping users’ perceptions. As demonstrated in some studies, a direct relationship between GAI’s intelligent performance and its functional capabilities (Maroufkhani et al., 2022; Priya and Sharma, 2023). This relationship improves users’ perceived value and satisfaction (Song et al., 2022; Lin and Lee, 2024; Song et al., 2024).

In the context of online learning, GAI’s intelligence capacity enables accurate comprehension of users’ inquiries and provision of elaborated responses, thereby enhancing users’ learning experience. Therefore, we hypothesize that:

Explainable AI (XAI) has been defined as systems designed to provide transparent decision processes and clear explanations, enabling users to understand system capabilities and limitations (Dwivedi et al., 2023). Research has demonstrated that this explainability feature enhances users’ trust and acceptance of recommendation algorithms while simultaneously enabling effective knowledge transfer, thereby increasing the adoptability of AI-generated suggestions (Zhang and Curley, 2018). Fundamentally, explainability refers to an AI system’s capacity to articulate its decision logic in a user-comprehensible format. This capability primarily aims to eliminate the “black box” nature of the AI decision-making process, consequently strengthening users’ confidence in the system (Shin, 2021). Within human-GAI interaction contexts, explainability facilitates rational evaluation of algorithmic outputs by providing decision-making rationales. The statement is echoed by the qualitative investigation. For example, interviewee (P23) mentioned: “AI can give more detailed content. This thought process is more in line with my idea of recognizing it and learning from it.” This mechanism significantly shapes users’ attitudes toward GAI (Cheung and Ho, 2025). Furthermore, by bridging a connection between the user’s perception and GAI’s operation, explainability not only deepens understanding of specific responses but also elevates the overall interaction quality. As users progressively acquire more comprehensive explanations of the system through successive cycles of inquiries, it enhances users’ perceived value and satisfaction. Therefore, we hypothesize that:

Response time, as a key indicator of AI’s service efficacy, reflects the timeliness of the system in processing users’ requests and providing feedback (DeLone and McLean, 2003; Liu Y. L. et al., 2023). AI, powered by sophisticated machine learning algorithms and natural language processing capabilities, can analyze vast datasets and generate precise responses within milliseconds. This superior responsiveness significantly improves the efficiency of human-computer interaction (Neiroukh et al., 2024). The statement is echoed by the qualitative investigation. For example, interviewee (P14) mentioned: “If you use MOOC or bilibili, or some other large and well-known platforms, one thing they have in common is that you need to watch videos, which may require some investment in your time cost. But instead, using the generative AI software, it can give you a result in a few seconds. I really like it!” Research has shown that prolonged response times not only reduce task efficiency but may also convey negative impressions about the system’s predictive capabilities. Users’ prevailing assumption that most prediction tasks should be inherently simple for AI systems (Efendić et al., 2020). Liu Y. L. et al. (2023) also pointed out that response time is one of the determining factors affecting users’ perceived value and satisfaction with an AI service. GAI’s timely response impacts the interaction quality, which subsequently strengthens users’ trust in the GAI (Pham et al., 2024). Especially in online learning scenarios, as the response speed increases, it not only sustains users’ engagement and concentration during GAI interactions but also fosters more positive attitudes toward the technology. Therefore, we hypothesize that:

Integrability means that GAI effectively facilitates the ability to combine information from different sources to respond to users’ problems (Chen et al., 2025). It depends on the task and contextual environment, which also reflects task-related properties (Nelson et al., 2005). Chen et al. (2025) argued that highly interdependent complex tasks are more dependent on the integrated system’s outputs than a task-independent system. GAI can be optimally adapted not only based on existing databases but also combined with historical user interaction data in the content generation process (Chen et al., 2025). Existing studies have shown that optimizing the knowledge acquisition pathways, thereby enhancing learning effectiveness and enabling learners to achieve superior outcomes (Korayim et al., 2025). Furthermore, the integrability facilitates rapid adaptation to environmental changes and effective utilization of pivotal opportunities (Ding, 2021). In online learning contexts, GAI generates systematic and related knowledge according to users’ needs, and this powerful integration capability enables users to flexibly respond to problems, thus fostering a more favorable attitude toward technology. Interviewee (P14) mentioned: “GAI will organize these words into more complete results and present to me, so it may be more convenient in the ability to integrate information. I think it is better than before we searched the web page.” Therefore, we hypothesize that:

Accuracy represents that the system provides up-to-date and relevant information to the users’ intended goals (Chung et al., 2020). It is one of the important foundations for users to use smart service products (Cheng and Jiang, 2022). This viewpoint is echoed by P9, who mentioned: “There are some official data or real-time information, and I do not 100% believe the answers it gives me.” During the interaction process, it is often necessary to donate considerable cognitive efforts to evaluate the precision and relevance of GAI-generated content (Chen et al., 2023c). When users confirm that AI recommendations sufficiently address their requirements, this validation triggers a “cognitive resonance” phenomenon—the perception that the system genuinely comprehends their underlying needs—thereby substantially increasing recommendation acceptance (Li et al., 2021). Accuracy not only makes users feel that their needs are fully valued but, more importantly, provides effective solutions (Yuan et al., 2022). This positive experience strengthens users’ recognition of AI technology capabilities (Gursoy et al., 2019), and plays an important role in users’ satisfaction (Walle et al., 2023). In the context of online learning, accuracy is of equal importance. GAI should ensure the solutions and methods are correct and feasible to establish perceived value, fostering positive user attitudes (Chen et al., 2023b). Therefore, we hypothesize that:

Source credibility defined as individuals’ perception of sources as trustworthy and expertized (Hovland and Weiss, 1951). Camilleri (2024) stated that users often rely on pre-existing perceptions about information sources rather than objectively assessing content quality. It is a prerequisite for users to assess information’s usefulness (Camilleri and Filieri, 2023). Compared to non-expert sources, information disseminated by experts is usually perceived higher reliability and credibility (Ismagilova et al., 2020). Users will be more inclined to accept the advice and knowledge provided by professional and authoritative sources (Hovland and Weiss, 1951; Wang and Scheinbaum, 2018). In the context of online learning, users expect the source reliability from GAI. This viewpoint is echoed by P21, who mentioned: “If the results of GAI are different from those I searched in Baidu, I may not be able to judge which information is true.” Only when they confirm information originates from credible and authoritative sources do they feel confident acquiring knowledge through GAI interactions. This credibility reinforces users’ perception of information utility (Camilleri and Kozak, 2022), as well as contributing to positive attitudes toward GAI. Therefore, we hypothesize that:

Personalization provides customized services to users based on their needs, preferences, and intent (Ameen et al., 2021). Personalization in AI-based services has been defined as the service capability to provide specialized services based on a user’s personal information and contextual usage (Liu and Tao, 2022; Kim and Hur, 2024). Highly personalized AI not only formulates precise inquiries to identify individual needs but also simulates users’ decision-making processes (Kim and Hur, 2024). Therefore, it generates customized content that better meets the users’ expectations. Pappas et al. (2017) demonstrated that users are more inclined to higher relevant information. Aw et al. (2024) further noted that higher levels of personalization in mobile AR shopping apps can create a stronger sense of realism and coherence between virtual and real dimensions, which can lead to a significant increase in users’ immersion. In online learning scenarios, when GAI provides personalized answers to users’ specific questions, it optimizes satisfaction (Li and Zhang, 2023), and effectively enhances users’ perceived value. Such personalized learning support better matches the educational content with the learner’s knowledge level and cognitive style, thus creating a more efficient and enjoyable learning experience (Baillifard et al., 2025). Interviewee (P6) mentioned: “Because many of my questions will be very professional, I would prefer to have such a personalized and professional AI.” Therefore, we hypothesize that:

Emotional support, as a crucial dimension of social support, fulfills individuals’ psychological needs by conveying empathy, emotional validation, and encouragement (Meng and Dai, 2021). Existing research suggested that human-provided emotional support effectively enhanced individuals’ role meaning, thereby increasing well-being (Pai, 2023), as well as significantly improving service evaluations (Menon and Dubé, 2007) and effectively alleviating psychological stress (Meng and Dai, 2021). GAI can simulate human emotional communication, and are gradually taking on the role of emotional support (Gelbrich et al., 2021). This viewpoint is echoed by interviewee P16, who mentioned: “I think GAI is completely different from human’s feedback. Maybe humans will tell you that the employment environment is not very good for finding a job, and the situation is not very optimistic now. But GAI is relatively neutral. It may just give me some vitality.” Lee et al. (2022) pointed out that AI chatbots equipped with emotional intelligence dialogue systems can provide emotional understanding and encouragement to users. This interaction facilitates deeper communication and builds emotional connections. In our interviews, participants also reported that when they experienced academic stress and confided in GAI, they received both emotional consolation and personalized academic guidance. In online learning scenarios, GAI’s emotional support creates a friend-like interactive experience, and this humanized interaction significantly enhances users’ perceived value and satisfaction. Therefore, we hypothesize that:

Perceived value represents users’ overall evaluation of a product or service’s usefulness (Chen et al., 2023b). Within human-GAI interaction contexts, this construct is users’ assessment of GAI’s functionality, such as its 24/7 availability, problem-solving efficiency, and generation of content (Carvalho and Ivanov, 2024; Xu et al., 2024). As Lee et al. (2007) stated that perceived value is more significant in increasing satisfaction than service quality. Perceived value is different in scenarios, such as technology, service delivery, and tangible commitments (De Kervenoael et al., 2020). In AI service applications, robots can effectively balance temporal efficiency, economic considerations, and user experience, which directly influence users’ perceived value (De Kervenoael et al., 2020). Interviewee (P10) mentioned: “I may still need to spend time identifying what is right and wrong. But it will take less time than Baidu. This effectively offsets the time spent identifying the answer and is important for my experience of using it afterwards.” Infusion use is an active, repetitive, and long-term deep use behavior of GAI (Hu et al., 2024), which represents a more sustained engagement than continuous use behavior. In new technology research, perceived value is a critical factor that influences users’ long-term use behavior (Lavado-Nalvaiz et al., 2022; Maroufkhani et al., 2022). Therefore, we hypothesize that:

Satisfaction is an indicator of service quality assessment, quantifying the variance between users’ actual service experience and their expectations (Xie et al., 2023). This concept encompasses not only the immediate usage pleasure but also a comprehensive assessment process that involves the comparison of users’ past experiences with their current expectations (Poushneh et al., 2024). When using AI services, satisfaction is one of the main factors that impact users’ subsequent behavior (Jiang et al., 2022; Chen et al., 2023b; Xie et al., 2024). This viewpoint is echoed by interviewee P25, who mentioned: “GAI, like Doubao, is helpful, because it will make me more and more fluent, and then the expression will become more and more natural, and the response will get better and better in all aspects. That’s why I’ve been sticking with it for oral speaking.”Specifically, when the GAI provides information that aligns with user requirements, this positive experience fosters favorable beliefs, ultimately promoting their sustained behavior (i.e., infusion use) (Ku and Chen, 2024). Therefore, we hypothesize that:

Figure 2 describes the theoretical model.

This study collected data through a questionnaire comprising three main sections. The first section outlines the research purpose, defines GAI, and presents two examples of its application in online learning scenarios. The second section measures eight antecedent factors, two mediators, and the outcome variable in the theoretical model. The third section captures participants’ demographic information, including gender, age, education level, profession, frequency and year of GAI usage in learning.

To ensure the reliability and validity of the questionnaire, the measurement items in this study were adapted from established scales in the literature, with appropriate modifications to the context of GAI in online learning scenarios. Among them, intelligence, response time, and explainability use the scales developed by scholars such as Mehmood et al. (2024), Darban (2024), and Liu Y. L. et al. (2023), respectively. The scale of integration, accuracy, and source credibility mainly refers to the study of Chen et al. (2025), Zhou and Wu (2024), Yuan et al. (2022), and Wilson and Baack (2023). Personalization and emotional support are mainly based on the scales developed by Chen et al. (2023a), Zhu et al. (2023), and Zhang et al. (2018). Perceived value is adapted from De Kervenoael et al. (2020) and Chen et al. (2023b). Satisfaction is based on Xu et al. (2023). Finally, infusion use is adapted from Hu et al. (2024). All questionnaire research data were collected by a seven-point Likert scale (1 = strongly disagree, 7 = strongly agree). Detailed measurement items are provided in Appendix B Table B1.

Before the formal survey, this study conducted a pilot test involving fifty participants and consulted experts to refine the wording and structure of the questionnaire items based on the participants’ feedback. The pilot test results indicated that the scale demonstrated strong reliability and validity.

Questionnaire data for this study were collected via Sojump,³ a widely used online survey platform in China with 260 million registered users, similar to Amazon MTurk (Wu et al., 2024). This platform has also been widely used in previous related studies (Ding et al., 2023; Del Ponte et al., 2024; Javed et al., 2024). Before the questionnaire, participants were provided with a brief introduction to GAI and online learning, along with two screenshots demonstrating the use of GAI in online learning contexts. After the questionnaire is completed, each participant will receive 3 RMB (about 0.413 $) as a reward. The questionnaire was distributed and collected in February 2025, yielding 386 participants who joined our study.

We obtained data through random sampling, but to ensure the quality of the data, this study implemented three data screening criteria. First, participants were required to have prior experience using GAI for learning purposes. The question, “Have you ever used GAI to assist in learning?” was set to exclude participants without experience (N = 17). Second, two attention tests were conducted to exclude the sample who failed to answer correctly (N = 23). Additionally, the reverse questions were included for the third item of explainability and the fourth item of accuracy to drop samples with inconsistent responses (N = 19). In sum, we included 327 participants in our data analysis. Following Chin’s (1998) guideline for PLS-SEM, we ensured the sample size exceeded both: (1) 10 times the number of items in the largest construct; (2) 10 times the number of independent variables. Second, we used G*Power 3.1.9 software to calculate the sample size. With a significance level α= 0.05, Power(1−β)= 0.95, and effect size = 0.15, the minimum sample size was 160. Our sample sizes satisfy these requirements. Among the participants, 160 were female (48.93%), with the majority aged 19–24 (41.59%). The data on year of usage indicated that users with 1–2 years of experience constitute the majority (51.07%). The demographics of the final sample are presented in Table 2.

We used Partial Least Square (PLS) to test our theoretical model. PLS-SEM has no strict requirements on sample size and quantity. In addition, it has strong predictive and interpretative ability (Hair et al., 2011), which is suitable for exploratory theoretical construction. The influence mechanism of this study first built a model of GAI’s infusion use in online learning scenarios, which belongs to exploratory research and is suitable for the PLS-SEM method. The data were analyzed using SmartPLS 4.0. We followed the two-step approach in examining the measurement and structural models.

First, this study systematically proposes a theoretical framework that GAI’s characteristics influence users’ infusion use in online learning scenarios. Previous studies have investigated intelligence and explainability in promoting positive GAI usage (Al-Emran et al., 2024; Darban, 2024; Theresiawati et al., 2025). However, these studies often focus on a single or limited characteristic. Our work comprehensively identifies eight key GAI characteristics influencing users’ infusion use—intelligence, explainability, response time, integrability, accuracy, source credibility, personalization, and emotional support. This integrated theoretical framework not only enriches our understanding of GAI attributes but also provides a systematic theoretical foundation for subsequent research.

Second, this study extends theoretical understanding of perceived value and satisfaction in online learning contexts. Previous research has examined either attitudes (i.e., satisfaction) or cognition (i.e., perceived value) as independent mediators influencing GAI usage in online learning (Chan and Zhou, 2023; Kim et al., 2025). However, the synergistic mechanism between these two psychological mediators in users’ behaviors has not been fully elucidated. Based on the SOR model, we demonstrate that the eight GAI features (stimuli) influence infusion use (response) through the parallel mediation of perceived value and satisfaction (organism). These findings provide deeper insights into how individual cognition and attitudes jointly evolve when responding to external stimuli.

Third, these findings contribute to the literature on deep usage of GAI in online learning contexts. With the advancement of GAI technology, increasing research focuses on users’ long-term usage behaviors in online learning (Ngo et al., 2024; Holzmann et al., 2025; Liu et al., 2025). As the ultimate stage in post-adoption, infusion use represents not only the deep integration of technology into the learning process but also the latent commercial value within the education domain (Hu et al., 2024). Although infusion use has been investigated in contexts such as information technology (Sundaram et al., 2007), customer relationship management (Chen et al., 2021), and smart objects (Hu et al., 2024), its examination in online learning environments remains limited. This study addresses this gap by providing a theoretical framework for understanding GAI infusion use, thereby supplementing and enriching research on user behavior in online learning contexts.

This study offers significant practical implications for GAI researchers and developers. First, GAI service providers should adopt a long-term strategic perspective to foster users’ infusion use, thereby unlocking greater business value. As demonstrated by globally successful products like ChatGPT, Gemini, and DeepSeek, deep engagement and comprehensive feature utilization are critical to commercial success. Service providers should focus on enhancing all eight identified dimensions—intelligence, explainability, response time, integrability, accuracy, source credibility, personalization, and emotional support—while prioritizing core user needs and establishing clear strategic objectives.

Second, educational institutions and teachers should establish multi-dimensional evaluations when selecting GAI learning tools. Beyond conventional metrics like “intelligence,” criteria should also consider other factors, such as explainability and emotional support. This enables users to receive both comprehensive knowledge and psychological encouragement during challenging learning phases. Additionally, leveraging GAI to facilitate personalized learning plans can transform GAI into a valuable educational partner.

Third, developers can enhance perceived value and satisfaction by emphasizing GAI’s practical benefits and distinctive advantages. Effective strategies include implementing intelligent summaries (e.g., “I have summarized the key points of this chapter for you, saving your research time”) and generating personalized learning progress reports that help users visualize their achievement. Furthermore, incorporating empathetic interactions during complex tasks (e.g., “This question is indeed challenging. Let us tackle it step by step.”) can create a pleasant and efficient learning experience.

Finally, configuration analysis suggests that when facing technical resource constraints, developers should prioritize five key dimensions: integrability, accuracy, source credibility, personalization, and emotional support. Specific implementations may include regularly updating knowledge bases to ensure content authority and relevance, enhancing information capture and summarization capabilities, dynamic user profiles, and emotional interaction.

This study has several limitations. First, this study is primarily based on the sample of Chinese users, and its conclusions may be influenced by specific cultural and contextual factors. This may limit the generalizability of the findings to some extent. Future research could incorporate more diverse participants across different countries and regions. It can help validate the transferability of our findings. Furthermore, exploring how cultural differences influence the use of GAI in online learning contexts could both enhance the theoretical robustness and practical applicability of the research. Second, this study relies on cross-sectional data collected. Future studies could employ longitudinal methods to capture the dynamic nature of users’ perceptions over time. Third, while this study is grounded in the SOR model, future work could integrate other relevant theories (e.g., diffusion of innovations theory, CASA paradigm) to further enrich the understanding of factors influencing users’ infusion use.