From the functional perspective, numerous studies have demonstrated that AIGC tools hold significant potential as learning aids. Chen found that students using AIGC tools during writing exercises significantly improved their structural expression and task completion efficiency (Chen et al., 2024). Lukas noted that AIGC systems provide learners with real-time feedback, rewriting suggestions, and translation support, thereby alleviating language barriers and cognitive burdens (Jürgensmeier and Skiera, 2024). Crompton and Burke further emphasized that AIGC technologies enhance classroom interaction and promote personalized instruction, showing high adaptability and scalability, particularly in higher education and open learning environments (Crompton and Burke, 2024). Across these studies, PU and PEOU have consistently been validated as key cognitive factors influencing students’ BI, thus reinforcing the core theoretical logic of TAM.

However, as AIGC tools become increasingly embedded in critical teaching activities—such as course writing, academic translation, and assessment completion—associated ethical concerns have become more prominent. First, the risk of academic misconduct has significantly increased. Existing research indicates that students often struggle to distinguish between original and AI-generated content, leading to frequent incidents of plagiarism and academic dishonesty (Plata et al., 2023). Second, AIGC may diminish students’ critical thinking abilities. Zhai (2022) argued that overreliance on tool-generated content can lead to a “cognitive outsourcing” effect, impeding the development of independent cognitive construction and expressive capabilities. In addition, concerns related to unclear content authenticity, algorithmic opacity, and ambiguous accountability have emerged as critical ethical risks for both educators and platform administrators (Ateeq et al., 2024). For instance, Yakubu et al. (2025), in a UTAUT-based study on college students’ adoption of AIGC writing tools, included only “social influence” and “facilitating conditions” as key predictors, without considering negative cognitive variables such as ethical concern. From an ethical perspective, several studies have attempted to introduce MS as a key moderating variable in user behavior models, to capture the intensity of individuals’ psychological responses to ethical issues. For instance, Vance et al. (2012) found that MS significantly moderated users’ BI when facing information leakage risks in the context of information security behavior. Liu et al. (2021), in a study on AI algorithm transparency, noted that individuals with high MS are more inclined to respond ethically to algorithmic bias. A summary of representative literature on AIGC applications in education is provided in Table 1.

To understand AIGC adoption in education, this study integrates TAM and PMT—two key models explaining how users balance functional benefits with perceived ethical risks in decision-making. These theoretical perspectives are particularly valuable for understanding responsible technology adoption in sustainable educational environments. Originally proposed by Davis (1989), TAM is one of the most widely used frameworks in the information systems domain for predicting user behavior. It posits that two cognitive evaluations—PU and PEOU—are the key determinants of BI to use a new technology (Davis, 1989). PU refers to the belief that using the technology will improve one’s task performance, while PEOU reflects the perceived simplicity of the technology’s operation. Together, these perceptions shape users’ positive evaluations and jointly influence their adoption decisions. Owing to its parsimonious structure and strong predictive validity, TAM has been widely applied in diverse fields such as educational technology, healthcare systems, and e-government (Scherer et al., 2019; ElKheshin and Saleeb, 2020; Wang et al., 2022). In studies related to AIGC, scholars have confirmed the significant impact of PU and PEOU on user adoption. For example, Li (2023) identified PU as one of the strongest predictors of adoption intention among university students using AI-based writing tools; moreover, PEOU was found to positively influence PU, thereby validating the chained causal structure within the TAM framework.

Despite its emphasis on technological performance evaluations in shaping adoption decisions, TAM pays insufficient attention to users’ psychological defense mechanisms and ethical evaluation processes when facing emerging technologies (Islam et al., 2014). This limitation becomes particularly salient in the context of AIGC, which involves complex issues such as moral judgment, responsibility attribution, and value conflicts (Ajibade, 2018). To address this theoretical gap, this study introduces PMT as the foundation for the risk perception pathway. Initially developed by Rogers to explain individuals’ protective behaviors in response to health threats (Floyd et al., 2000), PMT has since been widely applied to areas such as information security (Sommestad et al., 2015), data privacy (Boerman et al., 2021), and AI technology adoption (Park et al., 2024). The core of PMT involves two cognitive stages—threat appraisal and coping appraisal—through which individuals assess risks and decide whether to adopt protective behaviors. In education, these appraisals influence whether AIGC tools are seen as supporting sustainable and equitable learning. Threat appraisal includes Perceived Severity (PS) and Perceived Vulnerability (PV), while coping appraisal comprises Self-Efficacy (SE) and Response Efficacy (RE), reflecting users’ confidence in handling risks and the perceived effectiveness of coping strategies.

Previous research has demonstrated the predictive validity of PMT in explaining user avoidance, resistance or behavioral adjustment in response to technological risks. For example, Park et al. (2024) found that in AI-based virtual service contexts, both PS and PV significantly influenced users’ adoption attitudes and BI, while SE and RE enhanced users’ perceptions of strategic effectiveness and increased their willingness to adopt such technologies. In the cybersecurity domain, Dodge et al. (2023) also identified SE and RE as key predictors of users’ adoption of recommended protective behaviors. Moreover, Su et al. (2022), in the tourism industry context, noted that PS and PV not only shaped practitioners’ risk awareness but also affected their SE and RE, which in turn strengthened their professional resilience and behavioral adjustment capabilities. Integrating PMT with TAM helps construct a comprehensive behavioral model that encompasses both positive motivations (functional adoption pathway) and negative motivations (risk defense pathway). In the context of educational applications of AIGC tools, users may simultaneously hold high functional expectations and ethical or risk-related concerns. Relying solely on TAM is insufficient to fully capture the internal psychological conflicts and dynamic trade-offs users experience (Hsiao and Tang, 2024). More importantly, the four cognitive antecedents in PMT may not only directly influence users’ risk perceptions but also indirectly affect their evaluations of the tool’s functionality, thereby impacting their BI (Hsu and Silalahi, 2024). This potential chain mechanism—from risk perception to functional evaluation to behavioral response—offers a theoretically grounded and practically relevant framework that surpasses TAM in explaining actual user behavior, especially when applied to sustainable technology adoption strategies in education. Accordingly, this study builds upon the primary pathway of TAM by further incorporating the four cognitive antecedents of PMT as predictors of PU and PEC. This dual-pathway adoption behavior model not only extends the explanatory scope of traditional technology acceptance theories but also provides a more systematic theoretical foundation for understanding user behavior in ethically sensitive educational settings involving AIGC tools (Hsu and Silalahi, 2024).

In the field of technology acceptance and user behavior research, ethical dimensions have long been situated at the periphery. Particularly in the context of educational artificial intelligence tools, users’ concerns over potential moral risks, value conflicts, and normative uncertainties are often simplified as “cognitive burdens” or “usage barriers.” However, as ethical issues surrounding artificial intelligence technologies gain increasing attention from both society and academia, a growing number of scholars have begun to focus on the mechanisms of ethical cognition during technology adoption, attempting to incorporate “ethical judgment” into behavioral intention prediction models (Kwon et al., 2020). Among these concepts, PEC—a risk-oriented ethical cognition—refers to an individual’s subjective perception of moral conflict, ambiguous responsibility, system manipulativeness, and potential adverse consequences during technology use. Initially prominent in research on data privacy, algorithmic fairness, and AI transparency, PEC has since been introduced into studies of user acceptance of AI systems. In the context of educational AI, Sain and Lawal (2024) found that students’ recognition of ethical risks associated with content generation tools significantly and negatively predicted their usage intentions, with heightened sensitivity observed in high-risk scenarios such as academic writing and course assessments. Specific manifestations of PEC in AIGC applications include lack of content originality, unclear attribution of responsibility, model bias risks, increased student dependency, and unbalanced evaluation mechanisms by instructors. Hsiao and Tang (2024) argued that the challenges posed by AIGC in educational settings are not merely technical but constitute a fundamental disruption to the legitimacy of knowledge production. Thus, conceptualizing PEC as a key cognitive factor influencing BI not only responds to the ethical realities introduced by AI technologies but also enhances the model’s capacity to capture complex psychological structures. Moreover, understanding ethical cognition is essential for ensuring the responsible integration of AIGC tools into education systems that aim to be sustainable and socially accountable.

Notably, the effect of ethical cognition on BI is not uniform across all user groups. Individuals may respond differently to the same ethical issue, and this variation often stems from differences in moral psychological traits. One such trait is MS—an individual’s ability to recognize and respond to moral cues in ethically charged situations—which serves as a critical moderating variable in the relationship between PEC and BI (Crowell et al., 2008). According to the Four-Component Model (Rest et al., 1999), MS constitutes the initial stage of activating moral judgment, moral motivation, and moral action; without the recognition of a moral issue in a given context, subsequent judgment processes are unlikely to be triggered. In user behavior studies, MS is often treated as a moderator that explains variations in behavioral responses to ethical cognition. Mower (2018) revealed that individuals with high MS are more inclined to adopt rejection or avoidance strategies when facing moral dilemmas Although originally applied in domains such as organizational behavior and medical ethics, MS—as a psychological trait at the individual level—has proven to be theoretically adaptable and empirically valid in AI ethics scenarios. This is particularly relevant in educational contexts, where both students and teachers may experience value conflicts triggered by the use of AIGC tools, and their levels of MS can significantly influence whether PEC translates into actual resistance to usage (Ashford, 2021).

Based on the proposed structural model and theoretical logic, this section presents a series of research hypotheses focused on the causal relationships among the core variables, thereby providing a theoretical foundation for the subsequent empirical analysis.

Building on TAM and PMT, this study developed a dual-pathway adoption model that incorporates two critical variables: PEC and MS. The model systematically investigates users’ adoption mechanisms of AIGC tools within educational settings. Structural equation modeling was used to test 14 hypothesized paths, all of which were found to be statistically significant, indicating a robust model fit and theoretical coherence. In the functional cognition pathway, PEOU had a significant positive effect on both PU and BI (supporting H9 and H10), while PU also significantly enhanced BI and further influenced CI (supporting H13). These findings confirm the applicability of the core TAM framework in the context of AIGC use in education.

In the risk cognition pathway, PS and PV negatively influenced PU but positively affected PEC. Conversely, SE and RE positively influenced PU while negatively affecting PEC. These results support the logic of the PMT framework by showing that both threat and coping appraisals shape users’ perceptions of AIGC tools’ functionality and ethical risks. Within the ethical pathway, PEC significantly and negatively impacted BI, indicating that ethical concerns can suppress users’ willingness to adopt AIGC tools in educational contexts. Moreover, MS significantly moderated the relationship between PEC and BI, with individuals high in MS being more likely to amplify the negative impact of ethical concerns on adoption behavior. Overall, the dual-pathway model successfully integrates functional, risk, and ethical cognitive dimensions, uncovering the multifaceted drivers of user behavior in educational AIGC scenarios. The findings also validate the explanatory power of MS as a key psychological trait, offering both theoretical and methodological contributions to future research on AI in education.

Among the positive predictors of PU, PEOU and RE showed the most significant effects. Users who perceive AIGC tools as intuitive, easy to operate, and user-friendly are more likely to positively evaluate their usefulness—this finding aligns with the core logic of TAM as proposed by Davis (1989) (supporting H9). Notably, however, RE also had a strong and statistically significant impact on PU, with an effect size nearly comparable to PEOU (supporting H4). This suggests that in educational contexts, users’ trust in institutional safeguards plays a critical role in their assessment of a tool’s value. In the domain of health technology adoption, prior studies have found that users’ perception of clearly defined rules and governance mechanisms significantly enhances both perceived utility and behavioral intention (Zhang et al., 2017). Similarly, in AI-driven education, higher RE is associated with stronger PU. Particularly in education—a domain characterized by high responsibility—the need for institutional clarity becomes more pronounced, amplifying the role of RE in shaping perceptions (Osman and Yatam, 2024). By contrast, although SE had a statistically significant positive effect on PU (supporting H3), its influence was relatively weaker. This may reflect users’ perception of AIGC tools as “functionally rich but logically complex”—even those with adequate experience may still lack confidence in managing output quality (Masry Herzallah and Makaldy, 2025). This finding echoes Kasneci et al. (2023) argument that although AIGC can enhance efficiency in education, excessive reliance on system logic may reduce users’ perceived control and agency.

Regarding negative influences on PU, the suppressive effects of PS and PV are statistically significant but relatively weak (supporting H1 and H2). This suggests that while users recognize the potential severe consequences of AIGC usage (e.g., hallucinations or bias), these risks do not substantially diminish their perception of the tool’s utility. This ‘utility-over-risk’ calculus implies that in educational settings, the demand for efficiency often outweighs concerns about potential severity. These results support PMT assumption that threat perception weakens positive evaluations (Ruan et al., 2020), and further indicate that users’ risk cognition has deeply permeated their value judgments of AIGC tools. This diverges from the traditional TAM assumption that PU is generally unaffected by negative variables, and can be explained by the contextual specificity of education: AIGC tools are directly linked to critical issues such as student assignments, fair evaluation, and content originality. When users recognize the potential risks—such as encouraging academic laziness or blurring responsibility—they may downgrade their value assessments even if they acknowledge the tool’s efficiency. This aligns with Kelly et al.’s (2023) findings that educational users often hold a dual perception of AIGC as “useful but potentially harmful,” especially in high-stakes scenarios such as examinations and grading, where PU is easily disrupted by ethical evaluations. In addition, the study found that RE had a significant negative effect on PEC. However, this effect was relatively modest (H8: β = −0.116). This indicates that simply believing in the effectiveness of external regulations or policy safeguards is not enough to fully eliminate users’ ethical anxieties. Since AIGC ethics involve complex value judgments, institutional responses alone may have a limited capacity to soothe users’ subjective ethical concerns compared to internal factors. RE not only significantly reduces PEC (Al-Sharafi et al., 2021) but also contributes positively to the formation of PU (Zhang et al., 2017). This mechanism can be attributed to RE’s function in building psychological safety: when users believe that there are clear rules and institutional protections, they are more likely to downplay ethical risks and elevate utility evaluations. This finding is consistent (Sain and Lawal, 2024), which shows that in ethically sensitive contexts, trust mechanisms at the organizational or platform level often surpass individual self-efficacy in driving technology acceptance.

Regarding the formation of ethical concerns, the analysis confirmed that threat appraisals (i.e., PS and PV) significantly heightened PEC (supporting H5 and H6). Theoretically, this suggests that risk perception acts as a cognitive trigger for ‘moral vigilance.’ When users believe that AIGC misuse leads to severe consequences (e.g., academic dishonesty) or feel personally susceptible to these risks, their psychological defense mechanisms are activated, manifesting as heightened ethical anxiety. Conversely, coping appraisals (i.e., SE and RE) acted as protective factors, significantly mitigating PEC (supporting H7 and H8). This indicates that a ‘sense of control’ acts as a buffer against ethical distress. When users feel competent in managing the tool SE or trust that external regulations are effective RE, they perceive the ethical risks as manageable rather than overwhelming, thereby lowering their overall level of concern.

Taken together, this study identifies PU as a critical intersection between the TAM and PMT pathways, reflecting a dual dynamic—being suppressed by risk variables while simultaneously promoted by structural trust. This structural tension suggests that in the adoption of educational AI tools, the cognitive weight of PU is shaped not only by perceptions of tool performance but also by users’ subjective assessments of whether the associated risks are manageable. Functional design alone may not be sufficient to drive adoption intentions; rather, psychological assurance must be built through institutional safeguards, user training, and ethical education to establish both a sense of control and trust (Masry Herzallah and Makaldy, 2025).

In this study, PEC emerged as a significant negative predictor of BI. Although its path coefficient (β = −0.145) is smaller than that of functional drivers such as PU (β = 0.349) and PEOU (β = 0.231), its theoretical significance implies that ethical concerns act as a distinct psychological barrier (supporting H12). This finding indicates that, in educational contexts, users’ adoption of AIGC tools is not solely driven by functional expectations, but is highly sensitive to ethical implications. Especially in high-risk academic scenarios such as coursework, academic writing, and examinations, users’ awareness of issues like “lack of originality” or “ambiguous responsibility” may directly suppress their intention to use these tools. This result aligns with Shin’s (2021) assertion that “AI-related ethical concerns can directly undermine user trust, thereby influencing decision-making,” and resonates with findings by Zhou et al. (2018) that “AI usage in educational settings is more constrained by moral norms and expectations.

From a path coefficient perspective, the effect of PEC on BI was notably stronger than that of PS, PV, SE, and RE—indicating that users’ risk cognition regarding the likelihood and severity of consequences often requires mediation through PEC to impact behavioral outcomes. In other words, PMT variables serve as upstream cognitive factors influencing PEC, rather than directly predicting behavioral intention. This structural relationship reinforces PEC’s theoretical role as a mediator in ethical cognition and illustrates that educational users are more concerned with whether the technology is “morally appropriate” rather than merely “risky.” A similar structure was confirmed by Jannat et al. (2024), who found that PMT variables influence behavior primarily through ethical anxiety.

In addition, the moderating effect of MS on the PEC → BI relationship was also empirically validated. Specifically, when MS was high, the negative effect of PEC on BI was significantly amplified; when MS was low, the effect was attenuated (supporting H14). This suggests that individuals respond differently to the same ethical issue, depending on their capacity to perceive value conflicts. High-MS individuals tend to internalize ethical concerns more strongly, translating them into avoidance behaviors (Strahovnik, 2018). This result is consistent with Rest et al.’s (1999) moral development theory, which posits moral sensitivity as a prerequisite for moral judgment, and echoes Vance et al.’s (2012) empirical findings that individuals with higher MS are more prone to triggering behavioral defense mechanisms. Notably, compared with general cognitive variables in TAM and PMT, MS represents a relatively stable psychological trait, making its moderating effect more context-independent. In education—where norms, fairness, and responsibility are highly emphasized—such trait-level effects are particularly salient. For instance, teachers or postgraduate students typically exhibit a stronger awareness of academic norms, and thus tend to score higher on MS than undergraduate students. This difference may help explain the observed variation in the PEC → BI inhibitory path across user subgroups. Taken together, PEC not only functions as an outcome of risk perception but also serves as a crucial bridge between moral judgment and behavioral response. Meanwhile, MS operates as a psychological threshold at the individual level—only when users possess sufficient ethical awareness can the moral conflicts represented by PEC translate into behavioral inhibition (Crowell et al., 2008). This finding further consolidates the moderating role of MS within the ethical pathway and adds depth to existing AI ethics adoption models.

Furthermore, our findings regarding the interplay between functional value and ethical concerns resonate with the broader discourse on AI-driven educational transformation. As noted by Naidoo (2023) and Rughiniș et al. (2025), the sustainable integration of AI requires balancing technological advancement with human-centric ethical considerations. Our model provides empirical evidence for this balance by demonstrating that adoption is not a linear function of utility alone. Rather, it is the result of a dynamic trade-off: while PU drives the ‘engine’ of adoption, PEC acts as a critical ‘brake’—particularly for morally sensitive users. Therefore, achieving the sustainable transformation envisioned in recent scholarship depends on establishing a ‘function-ethics equilibrium,’ where efficiency gains are matched by equally robust ethical safeguards RE to foster long-term trust.

The findings of this study offer concrete guidelines for educational AIGC stakeholders. First, for platform designers, the significant impact of RE on lowering PEC suggests that “visible ethical guardrails” are essential. Since users’ trust relies heavily on external safeguards, designers should embed features such as real-time plagiarism checks, clear data usage transparency badges, and one-click citation generation directly into the interface. These design cues can psychologically reassure users that the system is governed by safe protocols, thereby mitigating ethical anxiety.

Second, for educators and policymakers, the moderating role of MS necessitates a differentiated approach to guidance. Our results show that users with high MS are more prone to avoiding AIGC due to ethical fears. For this group, institutions should provide clear “safe-use lists” and definitive integrity policies to alleviate their uncertainty. Conversely, for users with low MS who may not naturally perceive ethical risks, educational programs should focus on “ethical awakening”—using case studies of AI misuse to heighten their sensitivity and prevent reckless adoption.

Third, regarding the link between risk perception and usefulness, developers must prioritize “explainability” to lower cognitive costs. Since high perceived severity reduces utility by forcing users to verify outputs, future tools should provide confidence scores or source attribution. This would reduce the ‘cognitive verification cost,’ thereby enhancing both perceived usefulness and adoption intention.

Despite validating a dual-path model of AIGC adoption in education, this study has several limitations. First, the sample pooled university students, teachers, and platform employees. Although our regression analysis indicated no significant impact of occupation on behavioral intention, the sample size of specific subgroups was insufficient to conduct a rigorous Multi-Group Analysis (MGA) in SEM. Different stakeholders may indeed possess distinct ethical concerns. Future research should aim for larger, balanced sample sizes to systematically compare group differences and include broader user groups like K–12 teachers or corporate professionals. Second, the cross-sectional design prevents analysis of behavioral change over time; longitudinal or experimental methods are recommended. Third, the model emphasizes individual cognition, overlooking macro-level influences such as social norms or institutional policies. Lastly, although reliability and validity were confirmed, construct applicability across cultural settings requires further validation. Future studies should explore broader contexts and adopt mixed methods to enrich theoretical and empirical contributions. Furthermore, methodological advancements in generative AI offer new avenues for instrument development. As noted in recent scholarship, Large Language Models (LLMs) show promise in automating the generation and cross-cultural adaptation of psychometric items (Grobelny et al., 2025; Marmolejo-Ramos et al., 2025). Future researchers could leverage these tools to further refine the validity and efficiency of scales used in educational technology acceptance.