1. Introduction
In recent years, the surge in computing power has driven the application and development of artificial intelligence technology in the field of education [
1], mainly including personalized tutoring, homework help, concept learning, standardized test preparation, discussion and collaboration, and mental health support [
2]. Existing literature emphasizes its ability to significantly enhance emotional communication in the learning process, providing a more personalized learning experience for each student [
3]. AI chatbot is an automated conversational system capable of interacting with humans using natural language. Functioning as a virtual personal assistant, it provides support for various tasks [
4]. AI chatbots can meet students’ needs for academic consultation [
5], and improve communication efficiency by responding to inquiries 24/7, overcoming human limitations [
6], AI chatbots support students in using their native language for communication, providing sufficient inclusivity [
7]. AI-driven chatbots can use predictive technology to provide early intervention support for students at risk or in rebellious periods [
8]. Although AI chatbots may have multiple positive impacts on learners, their current acceptance rate remains relatively low. The key factors influencing learners’ adoption of AI chatbots and their underlying relationships remain unclear.
Davis proposed the Technology Acceptance Model (TAM) to evaluate technology adoption in organizational settings [
9]. Sánchez-Prieto et al. used TAM to study students’ acceptance of AI-based assessment tools [
10]. In another study, Gupta applied TAM to investigate the determinants of teachers’ adoption of emerging technologies like AI in teaching [
11]. Venkatesh and Bala made improvements on the basis of the TAM2 model. They added two composite variables: personal differences and system characteristics, and proposed the TAM3 model, which is more comprehensive and applicable. Their theoretical advancement posits that these two variables play significant roles in shaping individuals’ acceptance and usage of information technology systems [
12]. Lin et al. used the TAM3 model to study the factors influencing students’ acceptance and use of handheld technology for digitizing MOOCs [
13]. Kim et al. focused on the perceived ease of communication and perceived usefulness in the TAM to determine students’ perceptions of AI teaching assistants [
14].
Some studies have extended TAM to investigate students’ perceptions of educational tools. Shamsi et al. expanded TAM by incorporating constructs such as subjective norms, enjoyment, facilitating conditions, trust, and security to explore how students utilize AI-driven conversational agents for learning [
15]. Ragheb et al. combined UTAUT with the social influence construct to study students’ acceptance of chatbot technology [
16]. Bilquise integrated TAM, the service robot acceptance (sRAM) model, and the self-determination theory (SDT) model to understand UAE students’ acceptance of academic advisory chatbots [
17]. It can be seen from the existing literature that as users’ expectations of technology have evolved, the TAM lacks the ability to provide deeper insights into behavioral intentions in educational contexts. Therefore, it is crucial to gain a more comprehensive understanding of users’ acceptance of intelligent chatbots by introducing external factors such as social influence.
This study is based on the TAM3 model by incorporating external factors such as social influence, while discarding elements from the original model, including job relevance, result demonstrability, perceptions of external control, computer playfulness, and objective usability, as these factors are less directly related to students’ immediate learning experiences and needs in an educational context. This study focuses on analyzing the impact of eight main factors on students’ behavioral intentions to use AI chatbots: Perceived Ease of Use (PEOU), Perceived Usefulness (PU), Self-Efficacy (AISE), Anxiety (AIA), Perceived Enjoyment (PE), Output Quality (OQ), Social Influence (SI), and Behavioral Intention (BI). These factors were selected to understand students’ beliefs and expectations regarding the acceptance of AI chatbots. Additionally, the moderating effect of education level in the TAM has been demonstrated in numerous previous studies on behavioral intention influences. Qaid et al. found that education level affects university lecturers’ attitudes towards using e-government services [
18].
Integrating chatbots into educational environments has the potential to create more efficient learning environments [
19]. AI tools can provide students with timely feedback anytime and anywhere, potentially increasing student success rates and engagement, especially among students from disadvantaged backgrounds [
20]. In terms of homework and learning assistance, AI chatbots can offer detailed feedback and suggestions on student assignments [
21]. For example, ChatGPT can serve as a useful learning companion, providing step-by-step solutions and guiding students through complex homework problems [
22]. ChatGPT can write essays at a level similar to a third-year medical student [
23]. In personalized learning, AI chatbots can provide students with personalized guidance, accurately identifying learning blind spots and enhancing learning outcomes [
24], to meet each student’s unique needs, helping them learn difficult concepts and improve their understanding [
25]. Khan et al. studied the impact of ChatGPT on medical education and clinical management, emphasizing its ability to provide students with tailored learning opportunities [
26]. The interactive and conversational nature of ChatGPT can increase student engagement and motivation, making learning more enjoyable and personalized [
27]. In terms of skill development, AI chatbots can help students improve their writing skills by offering syntax and grammar correction suggestions [
28], fostering problem-solving abilities [
29].
However, integrating AI chatbots into student education also presents challenges. First, information reliability and accuracy, AI chatbots may provide biased or inaccurate information [
27], potentially misleading students and hindering their learning progress. Especially in the field of medical education, ensuring the reliability and accuracy of information provided by chatbots is crucial [
26]. Moreover, biased training data can lead chatbots to echo distortions, stereotypes, or discriminatory advice. Second, AI chatbots challenge academic integrity, educators may find it difficult to determine whether students’ answers are original or AI-generated, thereby affecting the accuracy of rating and feedback. This raises concerns about academic integrity and fair assessment practices [
30]. Third, in terms of social relationships, unlike human teachers, AI chatbots lack the ability to sense emotions and provide real-time emotional support [
24]. Finally, ethical concerns arise, particularly around data privacy, security, and accountability. Since AI chatbots gather student data during interactions, strong safeguards are necessary. In medical education, it typically includes patient confidentiality and ethical considerations, making the ethical and proper use of chatbots significant [
31].
In summary, applying AI chatbots in the education field can provide students with personalized study assistance and improve educators’ efficiency. However, the public has growing concerns about the accuracy of information, academic integrity, and ethical considerations. Therefore, striking a balance between these advantages and challenges is crucial for integrating AI chatbots into education responsibly.
This study conducts an extensive review of the literature on the current state of AI chatbots, the application of the Technology Acceptance Model (TAM) in education, and the benefits and challenges of integrating AI chatbots into student learning. It further proposes to employ Structural Equation Modeling (SEM) to analyze key factors influencing Chinese university students’ acceptance of educational chatbots based on an extended TAM framework.
4. Discussion
Generative large models are deeply integrated across various fields, yet their impact on students at different educational levels and the factors influencing their acceptance remain unclear. This study constructs a research model using selected variables from the TAM3 model, ultimately determining that undergraduate and graduate students exhibit different levels of acceptance toward educational AI chatbots, indicating that educational level significantly influences students’ willingness to use educational AI chatbots. The study tested the model with survey data from China. Results supported more than half of the hypotheses, while four hypotheses were not supported. The discussion will separately address each hypothesis.
4.1. Supported Hypothesis
Research shows that AISE [
35] may enhance users’ confidence in chatbots, making them perceive the tool as easier to use. This relationship also applies to educational AI chatbots, where students’ confidence in technology may reduce usage barriers. The technical AISE of the students directly influences their judgment of the PEOU of AI chatbots. In higher education settings, students generally have more experience with AI tools, and their familiarity with AI technology lowers the operational threshold.
PE [
36] may enhance users’ PU of AI chatbots through positive experience. The enjoyment and interactivity of AI educational chatbots could increase students’ recognition of the tool’s value. A pleasant learning experience could make students more likely to believe that AI tools can improve their learning efficiency.
Previous studies have mentioned that PEOU [
40,
41] has a positive effect on PU. In educational settings, when students perceive AI chatbots as having good PEOU, they tend to perceive them as useful. This relationship is central to the TAM. When AI chatbots serve as educational tools, PEOU may often be directly equated with usefulness. This conclusion accords with the logical chain of the classic TAM, indicating that operational convenience directly influences students’ judgment of a tool’s utility. For example, ChatGPT’s natural language interaction reduces learning costs, and the efficiency gains from PEOU can directly translate into increased usefulness.
BI is positively influenced by AISE [
35], meaning that confident students are more likely to have the willingness to use AI chatbots. They may perceive AI tools as enhancing their capabilities rather than diminishing their thinking abilities, further increasing their readiness to adopt new technologies with the aid of AI tools, reflecting the pathway of “confidence driving BI.” High AISE may lower the psychological threshold for technology use, directly encouraging BI.
Chocarro et al. found that the PU of chatbots has a certain advantage in teachers’ intention to use technology [
44]. Similar results were observed in this study: if students perceive the usefulness of AI chatbots, they may be more inclined to autonomously use AI to learn knowledge that is typically difficult to grasp, effectively enhancing their self-motivation and positively influencing BI.
Research has shown that SI has a positive impact on students’ BI to use chatbots for teaching [
16]. In higher education settings, social pressure factors such as teacher-student interaction, peer recommendations, and academic stress significantly influence individual decision-making. Moreover, Chinese society exhibits a positive attitude towards technological development, and people are more willing to learn and use AI tools rather than resist them. Especially the higher education students, who are more inclined to use AI tools to enhance their competitiveness.
Based on the verified hypotheses, it can be concluded that more confident students in higher education show greater BI to use AI chatbots to enhance their abilities. Developers of AI educational chatbots should enhance AISE feedback and convenience in their interaction designs, such as incorporating gamification and personalization while simplifying the interface. These designs will collectively, directly or indirectly, strengthen students’ BI to use AI educational chatbots. It is crucial to create a favorable social environment and regulatory framework for AI technology as well. Additionally, establishing shareable social networks within educational AI chatbot designs will help increase students’ intention to use them. Integrating the above design principles may facilitate the establishment of a virtuous cycle of “AISE-BI-AISE,” enhancing students’ self-motivation and promoting exploratory learning.
4.2. Unverified Hypothesis
Previous studies have suggested that AIA has a negative impact on PEOU [
37], but this study contradicts this finding. A possible reason is that students are highly familiar with AI tools, with 93.96% of students reporting their experience as “generally familiar” or above. This may be the main reason why AIA did not significantly affect PEOU. Furthermore, current AI chatbots generally apply user-friendly interactive designs with timely feedback, which could also reduce students’ AIA. Furthermore, higher education students, the study’s focus, whose AIA levels toward new technologies are generally lower than those of the general population.
The findings did not support a positive correlation between PE [
36] and PEOU. A possible reason is that in higher education settings, students may prioritize the practicality of AI tools, making the impact of PE on PU relatively minor. Additionally, in terms of experimental measurement, the questionnaire item “Interacting with AI chatbots makes me happy” may not adequately capture the aspect of learning enjoyment in an educational context.
This study does not support the conclusion regarding OQ [
39] from previous research, possibly due to differing evaluation standards among students for OQ. The measurement items of OQ in questionnaire design focus more on universality rather than educational context specificity, which is closely related to this conclusion. Additionally, AI products in 2024 generally exist at a high hallucination rate (e.g., ChatGPT 3.5, OpenAI o3, ERNIE Bot), potentially affecting experimental result reliability.
Dwivedi et al. suggested improving the system’s PEOU to enhance users’ willingness to use it [
43]. However, this study contradicts this finding, possibly because PE is more dominant, and PEOU affects intention through mediating effects. The study’s participants placed greater importance on PEOU, as evidenced by the non-significant PEOU → BI path (β = −0.197,
p = 0.099), which contradicts core assumptions of TAM3. This may reflect the unique characteristics of educational AI applications: students appear to prioritize PU over PEOU. For instance, compared to traditional interactive software, AI products like ChatGPT demonstrate inherently high PEOU, requiring minimal training for effective use. However, this conclusion warrants further validation through cross-cultural samples.
Among the unsupported hypotheses, the three paths related to PEOU—H2, H3, and H8—were not supported. One possible explanation is that the current development of large AI models is in a state of intense competition. Higher education students frequently employ various AI-assisted learning tools, which feature user-friendly interactions and fast feedback. Their PEOU significantly surpasses that of ordinary search and auxiliary tools, yet the differences among various AI tools are minimal. Students have clear psychological expectations regarding AI-generated results, all of which greatly reduce their sensitivity to PEOU. Compared to PEOU, the high hallucination rate of AI in 2024 (e.g., ChatGPT 3.5, OpenAI o3, ERNIE Bot 3.5) remains a more serious issue, making PU a greater focus for higher education students.
This study focuses on the field of educational chatbots, employing the Technology Acceptance Model to explore factors influencing students’ intention to use educational chatbots and constructing a chatbot product acceptance model for higher education students. Based on empirical data, this theoretical model establishes a practical framework for educational AI chatbots, offering insights for implementation in higher education. It provides an innovative theoretical reference for the development of AI in higher education. In future research, the model can be translated into targeted design principles to guide the interface and functional design of educational chatbots, offering practical application value.
4.3. Theoretical Innovations
This study has made several innovative contributions at the theoretical level. First, although AI chatbots are widely applied in the educational field, research on the acceptance of AI tools with different forms and interaction methods among higher education students across varying educational levels remains insufficient. This study focuses on higher education students, providing an in-depth exploration of their acceptance of educational AI chatbots. Second, the results of this study support most hypotheses related to the TAM3 model under the given research conditions, offering theoretical references for the design and application of educational AI chatbots. Furthermore, this study incorporates educational level as a research variable. Through measurement invariance testing, it demonstrates that educational level exerts a moderating effect on higher education students’ acceptance of AI chatbots. Differences in educational levels indirectly influence students’ BI toward educational AI chatbots.
4.4. Practical Application Potential
This work has significant practical value in AI-driven scenarios: First, Hierarchical Design for Different Educational Levels: Students at different educational levels exhibit significantly varying BI in educational AI chatbots. Design considerations should address undergraduates’ search needs and graduate students’ research requirements. For example, undergraduates may prioritize ease of use and interface friendliness, so providing a pleasant user experience (e.g., scaffolded teaching support) can enhance their BI to adopt the technology. Graduate students may focus more on answer depth and OQ, so delivering precise and in-depth responses can foster their trust in educational AI chatbots. Second, Gamification Design for AI Chatbots: Based on research findings, enhancing PE, PEOU, and SI in educational AI chatbots can directly or indirectly increase students’ BI. Gamification design is recommended, including incorporating reward mechanisms, points, levels, leaderboards, and community-based social features. Showcasing usage experiences from peers and experts can further improve students’ PU and acceptance. Furthermore, Personalized Learning Support: Chatbots should provide personalized learning recommendations based on students’ knowledge levels, learning speeds, and preferences [
75]. Such recommendations can be realized by analyzing students’ behavior patterns and usage history through algorithms, ensuring a tailored experience for each student and providing them with maximum satisfaction.
4.5. Limitations of the Current Study
The limitations have to be described in multi-folds. First, due to the rapid development of artificial intelligence technology, especially the emergence of DeepSeek V3/R1 earlier this year, the experimental data collected and analyzed last year reflects user behavior concerning specific AI models prevalent during the data collection period (e.g., ChatGPT 3.5, OpenAI o3, ERNIE Bot). While AI technology develops rapidly, the patterns of human-AI interaction and adoption captured at this specific historical juncture retain theoretical and practical relevance. However, the target users of this study are current Chinese higher education students, it also provides a theoretical foundation for the development of customized AI chatbots in the higher education industry. Second, the total sample size of the questionnaire is relatively small, and 66.42% of the respondents are students from humanities and social sciences disciplines, potentially introducing disciplinary biases. Third, All samples were collected from Chinese universities, where technology acceptance behaviors within China’s educational environment may be influenced by Chinese culture, potentially leading to an amplified effect of Social Influence (SI). The conclusions of this study might vary across regions and cultural contexts and should not be directly generalized to other cultural settings. Fourth, The study employs a convenience sampling method, the sample may lack representativeness of a broadly defined population due to high selection bias. And the findings are susceptible to various interferences, making it difficult to generalize the results to the entire population. Fifth, The moderating effect of academic level may reflect differences in students’ exposure to AI technology across different educational stages, but it should be noted that this finding is based on a limited sample and should not be directly generalized to macro-social phenomena such as the digital divide.
These limitations will provide researchers with opportunities for further study. First, in future research, classifications and analyses can be conducted based on different student groups, such as cultural background, education level, quality of education, gender, age, and experience with AI usage, to enable more in-depth exploration. Second, the sole reliance on questionnaires in this study presents a limitation. Although education level was identified as a statistically significant moderator, the lack of detailed subgroup analysis prevented us from determining its specific direction or underlying mechanism. Therefore, future studies should employ mixed methods. For instance, combining physiological measurements (e.g., EDA, RESP) with qualitative approaches (e.g., focus groups, in-depth interviews) under a stratified sampling design would yield a deeper and more comprehensive understanding of educational AI chatbots. Finally, future research should follow up on AI models like DeepSeek V3/R1 to track the impact of technological advancements on the design principles of theoretical models.
5. Conclusions
This study extends the TAM3 framework to elucidate key drivers of Chinese university students’ acceptance of educational AI chatbots, revealing significant influences from self-efficacy, perceived usefulness, and social influence on behavioral intention, alongside moderation by academic level that highlights potential disparities in adoption. These insights underscore the dual-edged nature of AI in higher education: while chatbots offer opportunities for personalized learning and timely knowledge access, they also pose risks such as information inaccuracies, ethical dilemmas related to academic integrity, who may differ in self-efficacy and familiarity with AI tools.
To overcome these accessibility challenges in AI-disrupted learning environments, developers should prioritize inclusive design features, such as adaptive interfaces that reduce anxiety through simplified interactions, gamified elements to boost perceived enjoyment, and mechanisms ensuring output quality and bias mitigation. Generally, incorporating multilingual support and emotion-aware responses could address inclusive issues, promoting equitable adoption across diverse student populations. Educators and institutions are encouraged to integrate training initiatives that build self-efficacy, particularly for undergraduates, to bridge the digital divide and foster information literacy in the upcoming ear of generative AI.
Furthermore, the findings inform educational policies for responsible AI use, advocating frameworks that emphasize ethical guidelines, data privacy safeguards, and assessments of benefits versus risks. By guiding the development of AI-enhanced tools and policies that prioritize universal access, this research contributes to relieve AI-driven disruptions, ensuring that educational chatbots enhance knowledge acquisition, creative problem-solving, and overall inclusive in higher education. Future studies could explore longitudinal effects or cross-cultural variations to further refine these strategies.