Self-Determination Theory (SDT), developed by Deci and Ryan (1980) and Ryan et al. (2022), offers a comprehensive psychological framework to understand human motivation. At its core, SDT posits that individuals are driven by the innate psychological needs for autonomy (the desire to self-regulate one’s actions), competence (the need to feel capable and effective), and relatedness (the sense of connection with others). When these needs are satisfied, individuals are more likely to develop intrinsic motivation, resulting in enhanced engagement, well-being, and personal growth.

In educational contexts, SDT has been extensively applied to understand how learning environments can support students’ self-regulated learning, motivation, and academic outcomes (Wang Y. et al., 2024; Wu and Chiu, 2025). The theory emphasizes that supportive learning conditions that foster autonomy and competence are crucial for sustaining students’ participation and deep learning.

With the increasing integration of digital technologies into education, SDT has found renewed relevance. In digital learning environments, satisfying the need for autonomy and competence is especially vital, as students often engage with learning tools and content in more self-directed ways. Studies have demonstrated that digitally mediated learning systems—including AI-enhanced tools—can effectively support these needs and improve engagement (Chiu, 2022).

In the context of AI-powered learning assistants such as chatbots, SDT offers a powerful lens to examine how these tools influence learner motivation. For example, an AI chatbot that enables learners to control their pace of study (autonomy) (Alamri et al., 2020), offers timely and constructive feedback (competence), and facilitates meaningful learning interactions (relatedness) may foster deeper engagement (Chiu et al., 2024). Thus, SDT provides a theoretically grounded foundation for examining how students’ perceptions of AI-enabled educational technologies shape their emotional and cognitive involvement in learning (Li et al., 2024).

Task-Technology Fit (TTF) theory, originally proposed by Goodhue and Thompson (1995), addresses how well a given technology aligns with the specific tasks users are trying to accomplish. According to the theory, the effectiveness of technology use is maximized when the technological features appropriately match the task requirements (Suhail et al., 2024). In other words, a high degree of task-technology fit enhances the likelihood that individuals will perceive the technology as useful and adopt it more readily, ultimately improving performance outcomes (Sharma et al., 2024).

The TTF model includes multiple constructs—task characteristics, technology characteristics, utilization, and performance impact—all of which converge on the central idea that technology should serve the actual needs of its users (Saifi et al., 2025; Wang C. et al., 2024). For example, if students perceive that an AI teaching assistant helps them complete coursework, access relevant information, or enhance comprehension efficiently, the fit between the tool and the academic task is considered high.

Therefore, TTF provides an important complementary framework to SDT in this study. While SDT focuses on internal motivational mechanisms, TTF emphasizes the external functional alignment between technology and educational needs. Together, these frameworks enable a comprehensive exploration of how AI teaching assistants influence student engagement by simultaneously addressing what learners feel (motivation) and how learners act (task performance fit).

Bringing together Self-Determination Theory (SDT), Task-Technology Fit (TTF), and the role of External Support, this study constructs a Triadic Engagement Model to explain how AI teaching assistants facilitate student engagement in design theory education.

Specifically, the model proposes three key pathways:

This integrative model moves beyond traditional technology adoption frameworks by highlighting the interplay between internal psychological needs, task-technology alignment, and social-structural support in shaping engagement with AI-powered tools in design education. Through this approach, the study addresses a critical gap in the AI-in-education literature by focusing not just on adoption or functionality, but also on the emotional, motivational, and technology dimensions of student experience. Based on the proposed Triadic Engagement Model, this study addresses one overarching question: How do psychological, technological, and contextual factors shape student engagement with AI teaching assistants in design theory education? through six specific sub-questions, each linked to particular theoretical variables and hypotheses. Table 1 presents the overview of research questions, theoretical basis, and hypotheses. Figure 2 shows the proposed research model and hypotheses.

The AI teaching assistant used in this study, named “MinArt AI,” was developed by the first author, a senior university lecturer with decades of experience teaching design theory courses such as History of Chinese and Foreign Design and History of Chinese Craft and Decorative Arts. The name “MinArt” reflects both geographical and cultural significance: “Min” refers to the Minjiang River, a major river in the Aba Prefecture where the university is located, often seen as a cultural symbol or “mother river” of the region. “Art” signifies the chatbot’s foundation in the arts. The long-term vision for MinArt AI is to evolve into a comprehensive educational agent that spans both fine arts and design domains, offering students extensive support across disciplines. Based on the existing commercial AI large language model ERNIE Bot, a commercial AI large language model, the author fine-tuned a domain-specific AI chatbot assistant using a curated corpus of design theory resources. Appendix A presents the technical documentation.

“MinArt AI” was deployed through a course web portal and made available to students 24/7. Its core functionalities included answering lecture-related questions, providing feedback on design concepts, and recommending supplementary examples. This direct integration positioned the AI assistant as an embedded element of the learning process rather than an ancillary tool. Figure 3 shows the homepage of the AI teaching assistant and a sample interaction in which students asked questions about the Red and Blue Chair. The AI teaching assistant is publicly accessible at https://mbd.baidu.com/ma/s/0gjlz9Z7.

The questionnaire items used in this study were developed based on prior research and measured using a five-point Likert scale. To ensure content validity, the survey underwent both a pre-test and a pilot test before the actual data collection. Since the present study was carried out in China, the questionnaire was developed using Back-translation (Brislin, 1986). Two bilingual experts from the College of Foreign Languages independently translated the survey from English to Chinese and then back into English to ensure accuracy in meaning and terminology. Appendix B presents the questionnaires.

This study received ethical approval from the University Research Ethics Committee. To ensure the privacy of respondents, several key issues were addressed before they completed the questionnaire. Participants were informed that all collected data would be used solely for academic purposes, participation was entirely voluntary, and their responses would remain anonymous. Informed consent was obtained once participants agreed to these terms.

The participants were undergraduate students majoring in various art and design disciplines, including visual communication, fine arts, and environmental design. They were enrolled across multiple year levels and degree programs and had all taken one or more compulsory design theory courses, such as History of Chinese and Foreign Design and Design Theory. All participants had prior experience using the “MinArt AI” teaching assistant during their coursework. Data were collected through a combination of online distribution via course communication platforms and offline dissemination using QR codes during class sessions. In total, 400 questionnaires were distributed.

After data collection, duplicate and ineligible responses were screened and removed. The following invalid responses were excluded. For online questionnaires: (1) Responses completed in less than 1 min; (2) Responses that failed the attention check at the end of the questionnaire. Following the exclusion of responses with absent values and outliers, 363 questionnaires were used for subsequent analysis (Ngoc Su et al., 2023).

According to the suggestion of Soper (2020), we determined the prior sample size for the structural equation model. This involves using an online power analysis application with the URL: https://www.danielsoper.com/statcalc/calculator.aspx?id=89. The inputs considered by our research model, which include the number of observed variables (43 items measuring constructs in our study) and latent variables (a total of 9 variables), expected effect size (medium effect is 0.25), expected probability (95% significance level), and statistical power level (80%). The online power analysis application determined that the minimum sample size required to detect the specified effect is 281 based on the structural complexity of the model. The sample size used in this study (N = 363) is larger than the recommended amount, indicating an adequate sample size. Besides, to gain deeper insights into students’ real-world experiences with AI teaching assistants, we conducted semi-structured interviews with 10 undergraduate students enrolled in the History of Chinese and Foreign Design course.

To explore RQ1–RQ4, PLS-SEM was conducted using SmartPLS 4.0 to examine how different technological, motivational, and contextual factors influence students’ engagement with the AI teaching assistant in design theory courses. PLS-SEM was chosen due to its suitability for handling complex models with multiple latent constructs and its robustness in small to medium sample sizes, which is critical for understanding the complexity of the student’s adoption of the technology situation (Al-Emran et al., 2025; Hair et al., 2017). This technique enabled us to test direct and mediating effects between constructs.

To complement the linear findings from PLS-SEM and capture possible nonlinear relationships, we applied ANN analysis using SPSS 27. ANN helps identify the relative importance of significant predictors and enhances model prediction accuracy (Kalinić et al., 2021; Leong et al., 2019). This is important in an educational setting because many interacting factors influence student behaviors and attitudes toward technology (Al-Emran et al., 2025). A two-layer feedforward network with 10-fold cross-validation was constructed to avoid overfitting and ensure the generalizability of results.

To explore RQ5, we adopted fuzzy-set Qualitative Comparative Analysis (fsQCA) to uncover multiple conjunctural paths that lead to high student engagement (Abbasi et al., 2022). This method is particularly suitable for identifying causal complexity and understanding how combinations of conditions may be sufficient for the outcome (Ragin, 2009).

To gain deeper insight into students’ experiences using AI teaching assistants in design theory education, the first author conducted semi-structured interviews with 10 undergraduate students who had interacted deeply with the AI teaching assistant throughout the semester. The participants were purposefully selected from multiple class cohorts to ensure diversity in academic background and AI teaching assistant usage frequency. The interviews were guided by open-ended prompts, such as: “In what ways has the AI teaching assistant helped you understand design theory content?,” “What limitations have you encountered?,” and “How does it compare to traditional teaching methods?” All interviews were conducted in a quiet setting, audio-recorded with participant consent, and transcribed verbatim by the first author to maintain data accuracy.

The qualitative data were then analyzed following Braun and Clarke’s (2019) six-step reflexive thematic analysis. Both authors independently read and coded all transcripts, identifying key phrases and patterns. Through iterative discussion, the codes were refined, grouped, and developed into broader themes. Any discrepancies in interpretation were resolved through dialogue until consensus was reached, ensuring analytical rigor and intersubjective reliability.

Considering the CMV that the questionnaire survey method could have potentially caused in this study, the Harman single-factor test was applied as a post-hoc statistical control (Al-Emran et al., 2025; Howard et al., 2024). The subsequent principal component factor analysis indicated that the first factor accounted for only 34.601% of the variance, which is well below the 50% threshold (Podsakoff et al., 2003). Furthermore, the variance inflation factors (VIF) of all latent constructs were calculated to test the risk of multi-collinearity, and all of these were inferior to the vigilance value of 5 (Hair et al., 2017). Therefore, CMV is not considered a significant issue in this study.

To ensure the quality, internal consistency, and reliability of the constructs, this study conducted validity and reliability tests following the recommendations of Hair et al. (2019). Cronbach’s alpha (α) and composite reliability (CR) were used to assess internal consistency and reliability. The results showed that both α and CR values exceeded the 0.70 threshold, indicating strong internal consistency and construct reliability. Convergent validity was supported by Average Variance Extracted (AVE), as all AVE values were above 0.50, confirming an adequate level of convergent validity (Chen et al., 2024). The detailed results, including α, CR, AVE, and factor loadings, are presented in Table 2.

Discriminant validity was evaluated using two methods. First, the Heterotrait-Monotrait ratio (HTMT) results in Table 3 were all below the 0.90 cutoff (Henseler et al., 2015). Second, the Fornell-Larcker criterion, presented in Table 4, showed that the square root of each construct’s AVE was greater than its correlations with other constructs (Fornell and Larcker, 1981). Together, these results indicate that the constructs used in this study demonstrated adequate reliability, convergent validity, and discriminant validity.

In the second phase, the structural model was assessed following the recommendations of Hair et al. (2019). The Bootstrap method with 5,000 bootstrap samples was employed to conduct significance testing within a 95% confidence interval. The analysis focused on standardized path coefficients (β), p-values, and the coefficient of determination (R²) to evaluate the model’s explanatory power (Latif et al., 2024).

As shown in Table 5, most hypotheses were supported. Specifically, perceived autonomy (PA), perceived competence (PC), communication quality (CQ), task-technology fit (TTF), and school support (SS) all had significant positive effects on student engagement (SE), confirming H1, H2, H3, H5, and H12, respectively. However, H4 (ITF → SE) and H13 (lecturer support → SE) were not supported due to non-significant p-values. In addition, communication quality significantly influenced both perceived autonomy and competence (H6, H7), while both individual-technology fit (ITF) and task-technology fit (TTF) significantly influenced PA and PC (H8–H11). According to Hair et al. (2019), R² values of 0.25, 0.50, and 0.75 indicate weak, moderate, and substantial explanatory power, respectively. The R² values were 0.652 for SE, 0.329 for PA, and 0.437 for PC, indicating moderate to substantial explanatory power.

Mediation analysis (see Table 6) further revealed several significant indirect effects. PC partially mediated the relationships between CQ, ITF, and TTF with SE, while PA also showed some mediation effects, although weaker. Specifically, the path CQ → PC → SE had the strongest indirect effect (β = 0.058, p = 0.010), confirming the importance of perceived competence in facilitating AI teaching assistant-driven engagement.

Figure 4 illustrates the final structural model with path coefficients and R² values.

To complement the linear structural model and further explore complex nonlinear relationships, this study developed three ANN models to predict Perceived Autonomy (PA), Perceived Competence (PC), and Student Engagement (SE). As shown in Figure 5, all models used Communication Quality (CQ), Individual Technology Fit (ITF), and Task-Technology Fit (TTF) as key input variables, with additional variables (PA, PC, LS, SS) added in the SE model.

Model performance was evaluated using Root Mean Square Error (RMSE). The average RMSE values (see Table 7) were relatively low, indicating high consistency and accurate predictions (Leong et al., 2013; Raut et al., 2018).

The sensitivity analysis, as shown in Table 8, assesses the relative importance of independent variables in influencing the ANN model’s predicted values. Normalized importance was calculated to compare each factor’s contribution relative to the most significant predictor. In the PA model, TTF (normalized importance = 100%) emerged as the strongest predictor, followed by CQ (88.3%). In the PC model, TTF again showed the highest influence (100%), with CQ (74.6%) also playing a key role. For the SE model, the most influential variables were TTF (100%), SS (77.3%), and CQ (71.8%), highlighting the importance of both technological fit and school support in driving student engagement.

To complement the linear and nonlinear analyses, fsQCA was applied to identify multiple pathways that lead to high student engagement. This method accounts for causal complexity and asymmetry, allowing for the identification of equifinal configurations that can each achieve the outcome of interest (Ragin, 2009). Prior to fsQCA, all variables were calibrated into fuzzy sets using the direct method proposed by Moreno et al. (2016). For each construct, we defined three qualitative anchors to determine full membership (0.95), crossover point (0.5), and full non-membership (0.05). The analysis yielded five configurations (see Table 9) that explain high levels of engagement with AI teaching assistants in design theory education. Each configuration represents a unique and sufficient path to high engagement, suggesting that different groups of students may benefit from different combinations of factors.

To assess the quality of the model, consistency and coverage scores were used. All configurations and the overall solution exceed the recommended thresholds for consistency (>0.80) and coverage (>0.20), indicating that the model is reliable and explains a substantial proportion of the outcome (Rasoolimanesh et al., 2021). These results demonstrate that AI teaching assistants can enhance engagement through various mechanisms. The configurations will be discussed in detail in the discussion section.

To complement the quantitative findings, a thematic analysis was conducted on qualitative interview data collected from students who interacted with the AI teaching assistant in design theory courses. We identified three major themes:

The results from the PLS-SEM, ANN, and fsQCA analyses revealed that student engagement with AI teaching assistants in design theory education is shaped by a combination of psychological, technological, and contextual factors. Among these, perceived competence (PC), perceived autonomy (PA), communication quality (CQ), task-technology fit (TTF), and school support (SS) emerged as the most significant predictors.

Our psychological constructs drawn from Self-Determination Theory (SDT) showed strong explanatory power. Both PA (β = 0.127, p < 0.05) and PC (β = 0.196, p < 0.001) had significant direct effects on student engagement, aligning with SDT’s emphasis on intrinsic motivation through autonomy and competence satisfaction (Chiu et al., 2024; Ryan and Deci, 2020). These findings are in line with studies in digital education contexts, showing that autonomy-supportive environments foster deeper engagement and self-regulation (Fryer et al., 2019). In design education, where student-led exploration and iterative learning are central (Dorst, 2011), the ability to independently use the AI assistant to seek clarification or explore ideas supports these core pedagogical goals. ANN results reinforced these findings: perceived competence (PC) and perceived autonomy (PA) were among the most influential predictors of engagement, with PC (67.349%) and PA (45.973%). Furthermore, fsQCA analysis showed that PA and/or PC were present in three out of five engagement configurations, indicating their crucial role across diverse learner profiles. These results align with prior studies in digital education showing that autonomy-supportive environments foster deeper engagement and self-regulation.

In our technological constructs, the communication quality of the AI assistant had a substantial effect on student engagement (β = 0.177, p < 0.01) and strongly influenced SDT constructs (PA and PC). ANN sensitivity analysis further highlighted CQ as one of the top predictors (71.789%). In the fsQCA configurations, CQ appeared as a core condition in three out of five pathways (Solutions 1, 2, 4), reinforcing that students’ perception of the AI teaching assistant’s responsiveness, accuracy, and credibility is critical for their engagement. In line with previous research, this suggests that students’ perception of the AI teaching assistant as a competent, credible, and responsive communicator plays a critical role in facilitating their use of the tool (Chiu et al., 2024), supporting prior research on the significance of interaction quality in digital learning (Saifi et al., 2025). In design theory education, where abstract concepts and contextual knowledge require AI teaching assistants to have clear articulation, communication quality becomes even more critical to help students grasp difficult content.

Third, TTF showed a strong and significant influence on engagement (β = 0.207, p < 0.001) and was ranked highly in the ANN sensitivity analysis (100% normalized importance across models predicting PA, PC, and SE). Critically, TTF was the only condition present in all five fsQCA configurations, indicating that students are most engaged when they perceive the AI assistant as well-aligned with their learning tasks. This is consistent with prior studies suggesting that students are more engaged when tools are well integrated with learning activities (Al-Emran et al., 2025; Dahri et al., 2024). In the context of design theory, where students are often involved in timelines, terminology, and theory, a well-matched AI assistant should streamline their workflow, providing quick access to examples or definitions that support rather than disrupt their conceptual reasoning.

Interestingly, individual technology fit (ITF) did not significantly predict student engagement in the PLS-SEM model (β = 0.054, p > 0.05). This suggests that being comfortable with digital technologies does not automatically translate into active engagement with an AI teaching assistant in design theory courses. Many design students are already accustomed to using complex digital tools, such as Adobe Illustrator or Photoshop for graphic design or Rhino and Blender for 3D modeling and rendering. As a result, their general digital literacy is relatively high, but this does not guarantee that they will find value in using a text-based AI assistant. What they seem to care about more is whether the tool truly supports their learning goals. For example, helping them analyze key design movements, understand theoretical frameworks, or apply abstract principles to design critiques.

Contextual factors also played a key role. School support (SS) had a significant direct effect on engagement (β = 0.216, p < 0.01) and emerged as an important factor in ANN (77.318% normalized importance). In fsQCA, SS was present in four out of five configurations (Solutions 1, 3, 4, 5), underscoring its broad impact across different learner types. These results suggest that institutional endorsement and integration of AI tools can encourage student buy-in. This is aligned with previous research (Molefi et al., 2024), presenting that school support is crucial to drive students to use AI tools. However, Lecturer Support (LS) did not reach statistical significance, which may reflect the self-driven nature of design disciplines, where students often engage in “learning by doing.” Notably, while lecturer support did not show a significant effect in the PLS-SEM model, it appeared in several configurations identified through fsQCA. This apparent discrepancy reflects the different logics of the two methods: while SEM captures linear, net effects (Hair et al., 2019), fsQCA identifies condition combinations (Ragin, 2009) that are sufficient for high engagement. Lecturer support, though not impactful on its own, can act as a complementary facilitator when paired with strong school support, task-technology fit, or student competence. In theory-based design courses, learners tend to take ownership of concept exploration, particularly when assignments demand individual interpretation or application. In contrast, visible institutional support, such as technical training, resource access, or official endorsement, plays a more important role in legitimizing and encouraging the use of AI teaching assistants (Al-Emran et al., 2025).

Collectively, these results highlight that while individual motivational states and perceptions of technology quality are important, broader institutional integration and task alignment are equally vital. Multi-method triangulation (PLS-SEM, ANN, fsQCA) strengthens confidence in these findings by demonstrating convergence across statistical modeling, machine learning, and configurational analysis. Table 10 summarizes the cross-validation outcomes, revealing consistent predictors and demonstrating the complementary nature of psychological, technological, and contextual factors in shaping student engagement with AI teaching assistants.

While the general linear analysis above identifies individual effects, student engagement is a multifaceted outcome often arising from specific combinations of factors. We used fsQCA to capture this complexity to uncover multiple conjunctural paths leading to high engagement. This approach recognizes that there may be five distinct profiles of highly engaged learners, each with its own recipe of conditions. These configurations are summarized as follows:

The qualitative interviews provided a nuanced view of how students experienced the AI teaching assistant in a design theory course. Three main themes emerged: (1) the AI teaching assistant’s immediacy and usefulness, (2) its limited capacity for deep inquiry, and (3) evolving student trust and critical engagement.

First, students appreciated the AI teaching assistant’s 24/7 availability, which allowed them to access support anytime during independent study. This aligns with previous research. It presents the accessibility of the AI teacher and the availability of support for students outside traditional classroom hours as one of the key benefits (Saihi et al., 2025). This enables students to seek help and clarification at their convenience (Roca et al., 2024). Many students liken it to having a personal tutor, especially helpful during late-night work. In the design process, this aligns with the “Discover” phase of the Double Diamond model, where learners seek information widely (Grönman and Lindfors, 2021). However, students also reflected that while the AI teaching assistant is good for quick answers, it often lacks the depth to support meaningful learning. This highlights the need to integrate the AI teaching assistant into learning designs that foster reflection and deeper engagement, which is especially important in design education, where understanding historical and theoretical context is key.

Second, while the AI teaching assistant performs well in knowledge delivery, its responses are often “flat” or lacking in critical depth. Students feel it rarely asks follow-up questions or challenges in their thinking unless prompted. This is in line with the views articulated in previous articles, where there is a consensus that AI-generated textual content appears logically weak, is not accurate enough in terms of facts and veracity, is not critical enough in terms of data exposition, and is not novel enough (Dwivedi et al., 2023). Instructors in design education often stimulate learning by asking, “Why do you think this works?” or “Have you considered this angle?,” questions that drive design thinking (Zhang et al., 2024). Although some students attempt to elicit deeper responses by framing open-ended questions, they still find the interaction less rich than classroom dialogue. In the future, one potential solution is to embed scaffolded prompt templates or critical thinking cues within the AI teaching assistant interface to guide students toward deeper inquiry. Instructors can also reinforce this by integrating AI teaching assistant responses into classroom critique and encouraging students to compare AI-generated responses with their own interpretations. Although the AI may not yet replace the nuanced dialogue of a human educator, it can become a valuable complement when situated within a reflective pedagogical structure. Nevertheless, this process of formulating better prompts became a reflective exercise in itself, helping students structure their thinking, similar to Schön’s concept of “reflection-in-action” in design studios (Schön, 1987).

Third, students develop stronger information literacy and critical habits over time. Initially, many students trust the AI teaching assistant fully, but gradually adopt a verification mindset, like cross-checking its answers against textbooks or other sources. Aligning with this, previous research criticizes that GenAI provides false information or exhibits phenomena of confabulation or hallucination (Arkoudas, 2023). However, in our research, we see a positive shift toward AI literacy—critical thinking, a key component of digital competence in today’s learning environments (Zhang and Zhang, 2024). Educators could enhance this by incorporating guide verification tasks or teaching strategies for assessing AI-generated content, essential practices in ethical and informed design.

Importantly, several students describe the AI teaching assistant as a “safe space” to ask questions they might avoid in class. This aligns with previous research (Pesonen, 2021; Roehrer et al., 2024). This anonymity supports self-confidence and reduces the fear of judgment, helping students engage more actively in learning. This also aligns with a method called “thinking out loud” in design education (Calder, 2018). The role of an AI teaching assistant as a non-judgmental thinking partner can be particularly beneficial in design education, which often requires emotional resilience and persistence through ambiguity.

However, concerns about over-reliance and ethical use also emerged. Some students worry that excessive use might limit their independent thinking. This is in line with prior research, suggesting that AI poses a risk to students in that their ability to think independently and express themselves verbally may decline (Dwivedi et al., 2023). Besides, other students question how their data was handled or express fears of unintentionally crossing academic boundaries (Cotton et al., 2024). These concerns align with broader calls in design education for responsible technology use and ethical engagement with tools, which need to be addressed in more specific future studies.