Graduate Education in China Meets AI: Key Factors for Adopting AI-Generated Content Tools

3.1 AIGC in Graduate Education

Previous research involving educators and students underscores the pivotal role of AIGC tools as these tools can automate content generation across domains such as text writing, image design, video production, and music creation, meeting the intensified demand for depth, breadth, and innovation in graduate studies (Chan and Hu 2023; Cooper 2023). They significantly boost efficiency and stimulate innovative thinking by offering preliminary ideas (Bockting et al. 2023) and provide personalized support tailored to individual learning needs, generating instructional content aligned with students’ progress and comprehension levels, thereby offering personalized feedback (Mizumoto and Eguchi 2023). These advantages suggest that the educational benefits of AIGC tools outweigh the risks (Cotton, Cotton, and Shipway 2024; Foroughi et al. 2023).

Integrating AIGC tools into graduate education necessitates acknowledging potential adverse consequences. Over-reliance on these tools might undermine students’ independent thinking and innovation (Vuong et al. 2023; Wang et al. 2023). Concerns also arise regarding the quality and accuracy of the content generated which may perpetuate societal and cultural biases (Lund et al. 2023; Foroughi et al. 2023). Inaccurate identification of such biases by graduate students could significantly impact academic research and teaching. Additionally, the risks of academic dishonesty (Houston and Corrado 2023; Sun and Hoelscher 2023), privacy and data security concerns (Yan et al. 2024), potential copyright infringement (Cooper 2023), and disparities in technological access raising educational equity issues (Knox 2023; Yan et al. 2024) must be considered.

3.2 Factors Affecting Intention to Use AIGC Tools

Adopting AIGC tools is influenced by a complex interplay of systemic aspects, individual differences, societal impacts, and environmental factors, as detailed in Table 1.

Research shows that factors like compatibility and perceived ease of use positively influence the intention to use AIGC tools, while complexity has a negative effect (Kasilingam 2020). Information accuracy mediates user intentions (Foroughi et al. 2023) and individual differences such as cognitive attitudes and effort expectancy also play a role, despite effort expectancy negatively impacting mental attitudes (Hooda et al. 2022; Ma and Huo 2023). Social influence has a limited impact (Agrawal 2023). Existing studies focus on specific AIGC tools and theoretical perspectives, often overlooking the experiences of graduate students and underestimating privacy concerns which can lead to personal information leaks (Liu and Ma 2024). The effects of privacy protection measures and the relationship between privacy concerns, influencing factors, and individual trust are inadequately considered.

4.1 Performance Expectancy, Effort Expectancy, Social Influence, and Facilitating Conditions

The four constructs derived from the unified theory of acceptance and use of technology (UTAUT) model (Venkatesh et al. 2003) have been widely employed in various contexts to examine user intentions and behaviors (Martins et al. 2014; Foroughi et al. 2023).

In an educational context this relates to graduate students’ expectations of AIGC tools’ efficacy in addressing academic challenges. Specifically, it reflects students’ belief in these tools’ ability to improve tasks such as paper writing, research, and multimedia creation, thereby streamlining the learning process and enhancing academic efficiency and outcomes (Anthony et al. 2023; Foroughi et al. 2023). Thus, the hypotheses as follows:

H1. Positive performance expectancy positively affects the intention to use AIGC tools.

Effort expectancy refers to the anticipated level of effort required to learn and use technology, or the perceived ease of technology adoption (Foroughi et al. 2023). For graduate students, this reflects their perception of the effort needed to learn and use AIGC tools. The ease of use of new technologies significantly influences their adoption (Al-Emran et al. 2023; Liu et al. 2023). If AIGC tools offer more superficial, precise, and convenient functionalities and foster easier information interaction, students are likely to adopt these tools. Consequently, we proposed that:

H2. Positive effort expectancy positively affects the intention to use AIGC tools.

Individuals often face uncertainty when adopting a new technology or system (Mu and Lee 2017). Thus, in educational and professional settings, particularly at the graduate level, they seek guidance from their immediate social networks (Hursen 2019; Tasso et al. 2021). They consult peers, advisors, and other significant individuals to decide whether to adopt a particular technology tool and gauge the extent of social influence on their willingness to embrace it. Accordingly, the study put forth the following hypotheses:

H3. Positive social influence positively affects the intention to use AIGC tools.

In education, facilitating conditions imply the availability of guidance, technical assistance, and compatibility between new technologies and the other tools of the user (Farazouli et al. 2024; Al-Emran et al. 2023). This study measures students’ confidence in their access to the resources, knowledge, and support necessary to use AIGC tools for learning. Thus, the following hypothesis was formulated:

H4. Positive facilitating conditions positively affect the intention to use AIGC tools.

4.2 Hedonic Motivation, Learning Value, and Habit

Expanding upon the UTAUT model, Venkatesh et al. (2012) introduced three new variables into the UTAUT2 model to better predict technology adoption intentions. The revised model has effectively explained student behaviors concerning various technologies, including e-learning (El-Masri and Tarhini 2017; Mehta et al. 2019).

As for graduate students using AIGC tools, hedonic motivation may stem from the enjoyment of innovation, satisfaction from artistic creation, a sense of accomplishment from problem-solving, and the growth experience of learning new technologies (Foroughi et al. 2023). Exploring uncharted territories with these tools, overcoming academic research bottlenecks, and generating a pleasant, fulfilling user experience may enhance their willingness to use generative AI tools. Prior research has confirmed that intrinsic factors such as enjoyment and pleasure significantly influence students’ intentions to adopt new technologies for education (Alalwan et al. 2017; Oluwajana et al. 2019). The following hypothesis is formulated based on discussions:

H5. Positive hedonic motivation positively affects the intention to use AIGC tools.

Learning value reflects the user’s perception of a system’s practical utility (Sitar-Taut and Mican 2021; Foroughi et al. 2023) and measures the meaningfulness and value attributed to using generative AI tools in terms of time-saving, enhancing the learning process, and achieving personal development objectives. This construct replaces the price value due to the current prevalence of freely available AIGC tools (e.g., ChatGPT 3.5, Stable Diffusion) which renders the price value construct less relevant (Ain et al. 2016; Hong et al. 2022). Learning value is a crucial predictor of behavioral intention in the educational process (Dajani and Abu Hegleh 2019; Zacharis and Nikolopoulou 2022). If learners do not perceive these tools as valuable for expanding knowledge, saving time, and achieving learning goals, their confidence and interest in the learning process may diminish (Faqih and Jaradat 2021; Mehta et al. 2019). We postulate as follows:

H6. Positive learning value positively affects the intention to use AIGC tools.

When individuals establish a regular habit, they subconsciously execute tasks which can be quantified in terms of behavioral frequency, consistency, and degree of automation (Baudier et al. 2020; Chen et al. 2022). Habit assesses the extent of habituation among students in using AIGC tools. A well-established habit can significantly enhance future usage patterns (Fan and Suh 2014; Yang et al. 2022), leading to a better understanding of the technology and increasing the likelihood of its continued use (Jeyaraj 2022; Sebastián et al. 2023). Therefore, the following hypothesis is proposed:

H7. Positive habit positively affects the intention to use AIGC tools.

4.3 Content Accuracy and Privacy Concerns

Two pivotal indicators, content accuracy and privacy concerns, are incorporated in this study. Content accuracy pertains to the efficacy of these tools in meeting the academic needs of graduate students and providing accurate outputs which influence their reliance and willingness to use these tools (Yucel and Usluel 2016; Foroughi et al. 2023). Privacy concerns reflect the significance of students’ apprehensions about personal data security which markedly influences the acceptance and utilization of new technologies (Liu et al. 2023; Yan et al. 2024).

Research consistently reveals that people trust technology more when they perceive its content as accurate and dependable, leading to usage (Blanco-González et al. 2023; Aparicio et al. 2017). This trust is particularly crucial in the educational context where learners are likelier to adopt new systems for achieving their learning goals when high content accuracy is high (Al-Adwan et al. 2022; Long et al. 2019). Teachers assert that the accuracy of the information generated by the AIGC tool impacts students’ intention to use it and their behavior (Chan 2023). Accordingly, we proposed that:

H8. Positive information accuracy positively affects the intention to use AIGC tools.

Aligned with studies that use privacy concerns as a proxy for privacy (Belanger and Crossler 2011; Zhu and Grover 2022), this concept is defined as the degree to which individuals pay attention to privacy information when using IT services (Liu et al. 2023). The use of AIGC tools in education raises significant privacy concerns, particularly regarding the use of students’ data for training large language models (Tsai and Gasevic 2017; Brown et al. 2022), risking exposure of private information. If graduate students fear for their personal data’s safety they may hesitate to adopt these tools (Korir et al. 2023; Mutimukwe et al. 2022), reducing their willingness to embrace IT (Fox and Connolly 2018). This leads to the following hypothesis:

H9. Positive privacy concerns negatively affect the intention to use AIGC tools.

4.4 Trust

Trust in human-machine relationships is defined as the extent to which a user believes that an automated system will perform as expected (Papenmeier et al. 2022). To instill trust in students, AIGC tools must consistently deliver high-quality and reliable services efficiently and securely. Conversely, a lack of trust can increase resistance to adopting these tools (Almaiah et al. 2019). Trust significantly correlates with the acceptance and use of educational services, as shown by a positive association between trust and the intention to use (Parhamnia 2022). Therefore, we postulate the following:

H10. Positive trust positively affects the intention to use AIGC tools.

The association between trust and the intention to use AIGC tools is multifaceted and influenced by several critical factors. One such determinant is information accuracy wherein delivering high-quality and reliable content is pivotal for establishing trust and fostering usage intentions toward information technology (Papenmeier et al. 2022). Effective delivery of accurate content instills confidence in the performance of AIGC tools among graduate students, thereby cultivating trust (Foroughi et al. 2023). Privacy concerns, integral to trust formation (Hanif and Lallie 2021; Zeng et al. 2023), may diminish confidence in AIGC tools due to user apprehensions about personal information security, potentially leading to negative consequences. In high-risk scenarios, users might bypass the technology and rely on their capabilities (Jøsang and Presti 2004). Social influence also profoundly impacts trust formation (Wei et al. 2019; Liu et al. 2023), with positive societal evaluations and recommendations from others likely strengthening users’ trust in AIGC tools and subsequently increasing their intention to use them. Finally, effort expectancy reflects users’ expectations regarding the effort required for technology use (Lian and Li 2021; Risman and Budiarti 2023). User-friendly tools and needless cognitive burden contribute to users’ sense of control over the technology, enhancing trust levels and ultimately fostering the formation of usage intentions. These factors collectively constitute a complex trust network, leading to the formulation of the following hypotheses:

H11. Positive information accuracy positively affects the trust in AIGC tools.

H12. Positive privacy concerns negatively affect the trust in AIGC tools.

H13. Positive social influence positively affects the trust in AIGC tools.

H14. Positive effort expectancy positively affects the trust in AIGC tools.

Figure 1 illustrates the definitions of all influencing factors and their interrelationships.

Figure 1:

The conceptual AIGCT-SI model.

5.1 Questionnaire Design and Measurement

The questionnaire began with an overview of AIGC and collected demographic information. Measurement scales for the 11 constructs were adapted from established literature to fit the AIGC context. Performance expectancy and information accuracy were based on Foroughi et al. (2023), effort expectancy and hedonic motivation on Nikolopoulou et al. (2021), and social influence, facilitating conditions, and intention to use on Venkatesh et al. (2012). Habit was adapted from Farooq et al. (2017), privacy concerns from Liu et al. (2023), and trust from Almaiah et al. (2019).

The questionnaire was pre-tested by three experts to ensure face and content validity. Revisions were made based on their feedback, addressing ambiguous expressions and adding options for learning duration. A pilot test with 25 students showed Cronbach’s alpha values for all constructs exceeded 0.7, indicating reliable measurement results. The final survey, consisting of 39 items, is presented in Appendix A. All items were rated on a five-point Likert scale from 1 (strongly disagree) to 5 (strongly agree).

5.2 Data Collection and Verification

The study focuses on postgraduate students currently enrolled in Chinese universities. Data was collected from March 7 to April 7, 2024, utilizing the Wenjuanxing (www.wjx.cn) platform with authorized re-distribution. A total of 494 responses were received, with 41 excluded due to incomplete or contradictory responses, resulting in 453 usable questionnaire entries for analysis. The effective response rate of 91.7 % met the criteria for sample stability. The demographic details of respondents are presented in Table 2.

Various strategies were implemented to minimize non-response errors before, during, and after data collection. Respondents were assured of anonymity and confidentiality, emphasizing that the data would be used solely for scientific research. This approach alleviated privacy and trust concerns. The survey design reduced response time and effort to minimize survey fatigue. Non-response bias tests, comparing early and late responses using a t-test, showed no significant differences, indicating minimal non-response error.

To address potential common method bias (CMB) from self-reported surveys (Podsakoff et al. 2003), Harman’s single-factor test (Harman 1976), and the full collinearity test (Kock and Lynn 2012) were employed. Exploratory factor analysis revealed ten factors, with the initial factor accounting for only 28.6 % of the variance, below the 40 % threshold. The full collinearity test showed all construct variance inflation factor (VIF) values in the inner model were below 3.3 (Kock 2015), indicating no multicollinearity. Thus, CMB was not a significant concern in this study (see Appendix B for details).

5.3 Data Analysis

The main objective of this exploratory study is to identify factors influencing the intention to use AIGC tools, using a hybrid SEM-fsQCA approach (Kim et al. 2024; Elshaer et al. 2024; Foroughi et al. 2023). Partial least squares structural equation modeling (PLS-SEM) is utilized to measure the predictability of each component, while fuzzy-set qualitative comparative analysis (fsQCA) is employed to understand the non-linear and complex relationships among factors, validating the proposed hypotheses. PLS-SEM analysis uses IBM SPSS Amos Graphic and SPSS Statistics 29.0 while fsQCA analysis is conducted with fsQCA 4.1 software.

6.1 PLS-SEM Results

6.1.2 Structural Model Assessment

Applying the maximum likelihood estimation method to conduct a structural equation modeling analysis on the AIGCT-SI model revealed significant effects between its structures, with all paths validated for significance using the analysis of variance (ANOVA) and t-tests, as depicted in Table 5.

Table 5:

Results of hypotheses testing.

Hypotheses	Relationships	Path Coefficients(β)	T-Values	p-Values	Decision
H1	PE → IU	0.211	4.039	***	Supported
H2	EE → IU	0.135	2.613	0.009**	Supported
H3	SI → IU	0.032	0.672	0.502	Rejected
H4	FC → IU	0.100	2.057	0.040*	Supported
H5	HM → IU	0.004	0.071	0.943	Rejected
H6	LV → IU	0.167	3.116	0.002**	Supported
H7	HB → IU	0.137	2.158	0.031*	Supported
H8	IA → IU	0.040	0.504	0.614	Rejected
H9	PC → IU	−0.060	−1.347	0.178	Rejected
H10	TR → IU	0.140	2.279	0.023*	Supported
H11	IA → TR	0.556	9.513	***	Supported
H12	PC → TR	−0.099	−2.103	0.035*	Supported
H13	SI → TR	0.130	2.669	0.008**	Supported
H14	EE → TR	0.149	2.998	0.003**	Supported

1. Notes: *p < 0.05; **p < 0.01; ***p < 0.001.

Performance expectancy (β = 0.211, p < 0.001), effort expectancy (β = 0.135, p < 0.01), facilitating conditions (β = 0.1, p < 0.05), learning value (β = 0.167, p < 0.01), and habit (β = 0.137, p < 0.05) all positively and directly influenced users’ intention to use. Notably, performance expectancy exhibited the most substantial impact on usage intention. However, contrary to theoretical predictions, hedonic motivation did not significantly influence graduate students’ intention to use AIGC tools. Figure 2 shows the results of the structural model testing.

Figure 2:

Structural model evaluation. Notes: Unstandardized β, *p < 0.05; **p < 0.01; ***p < 0.001, solid line: support, dashed line: not significant.

While information accuracy, privacy concerns, and social influence did not directly impact the intention to use AIGC tools they significantly influenced trust. Information accuracy (β = 0.556, p < 0.001) had the strongest positive impact on trust while privacy concerns (β = −0.099, p < 0.05) negatively affected trust. Social influence (β = 0.13, p < 0.01) and effort expectancy (β = 0.149, p < 0.01) also significantly influenced trust. Additionally, trust (β = 0.14, p < 0.05) positively impacted usage intention. Thus, these factors indirectly affect the intention to use AIGC tools by influencing trust as a mediator.

The mediating effect of trust was examined using bootstrapping with 5000 samples (see Table 6). Mediation was established if the 95 % confidence interval (CI) of the indirect effect did not include zero. A significant direct effect indicated partial mediation while an insignificant direct effect indicated full mediation (Preacher and Hayes 2004, 2008). The results show that trust fully mediated the impact of privacy concerns, information accuracy, and social influence on graduate students’ intention to use AIGC tools, with these factors primarily influencing usage intention through trust. Effort expectancy partially mediated through trust, affecting usage intention both directly and indirectly.

Table 6:

Results of mediating effect tests.

Independent variable	Dependent variable	Effect	Point estimate	Bootstrap 5000, 95 % CI
				Lower	Upper
Privacy concerns	Intention to use	Indirect effect	−0.015	−0.046	−0.001
		Direct effect	−0.065	−0.177	0.046
		Total effect	−0.08	−0.192	0.033
Information accuracy	Intention to use	Indirect effect	0.069	0.010	0.143
		Direct effect	0.035	−0.106	0.184
		Total effect	0.104	−0.014	0.227
Effort expectancy	Intention to use	Indirect effect	0.018	0.002	0.045
		Direct effect	0.115	0.029	0.204
		Total effect	0.132	0.047	0.222
Social influence	Intention to use	Indirect effect	0.018	0.003	0.047
		Direct effect	0.031	−0.070	0.128
		Total effect	0.048	−0.055	0.147

6.2 FsQCA Results

Applying fsQCA to supplement the assessment of joint effects among predictor variables, this method is commonly acknowledged as a complement to the PLS approach, offering insights into distinct configurations or components achieving significance (Pappas and Woodside 2021). To ascertain the existence of a unique key predictor variable independently correlated with the intention to use AIGC tools, a necessary conditions analysis was performed (Mikalef and Pateli 2017). Determining the necessary and sufficient conditions for a given outcome involves evaluating critical criteria of consistency and coverage. The sufficiency of the formulation is ensured by confirming that its coverage and consistency values are not less than 0.2 and 0.8, respectively. Identifying the necessary antecedents leading to the focal outcome depends on coverage and consistency values exceeding 0.9 (Dul 2016).

The seven configurations introduced in Table 7 highlight their potential to stimulate AIGC usage. With an overall solution coverage of 0.613 it is inferred that these configurations collectively contribute significantly to the intention to use.

The diverse factors influencing intention to use, consistently exceeding 0.9 for each configuration, indicate a high level of coherence in their combined impact. Coverage analysis underscores the importance of each configuration in explaining variations in graduate students’ intention to use AIGC tools.

Factors such as performance expectancy, effort expectancy, and facilitating conditions consistently appear across all configurations, highlighting their broad impact on graduate students’ intention to use AIGC tools. Privacy concerns feature in four configurations, with three indicating a positive effect on the intention to use and two showing a negative impact. This implies that graduate students’ intention to use AIGC tools may be influenced by complex factors related to privacy concerns, potentially displaying bi-directionality.

Furthermore, the necessary conditions analysis was conducted to uncover whether the proposed ten constructs are essential drivers for AIGC tools. The results of the necessary conditions analysis, as showcased in Table 8, indicate that achieving a high level of intention to use AIGC tools is possible without any of the predicted factors.

7.1 Findings from PLS-SEM

PLS-SEM results show that performance expectancy, effort expectancy, facilitating conditions, and habit significantly influence graduate students’ intention to use AIGC tools, consistent with prior research (Foroughi et al. 2023). Learning value also plays a significant role, validating the model for this specific group. The strong effect of performance expectancy can be attributed to the efficiency gains and improved learning outcomes provided by AIGC tools. For example, AI tools like automated literature review systems can reduce time spent on manual data collection, directly addressing academic needs and encouraging adoption.

Contrary to expectations, hedonic motivation did not significantly affect usage intention. This may be due to the task-oriented nature of graduate studies where efficiency and academic results take precedence over enjoyment (Lavelle and Bushrow 2007). Additionally, the complex nature of AIGC tools may diminish the enjoyment factor as students prioritize utility over pleasure in high-pressure academic environments (Cohen and McConnell 2019).

Bootstrapping was employed to estimate the stability of mediation effects, confirming that trust fully mediates the relationship between privacy concerns and adoption intention, as explained by social cognitive theory and information processing theory. Social cognitive theory, which emphasizes observational learning and environmental feedback in decision-making (Bandura 2001; Venkatesh et al. 2003), suggests that social influence can enhance trust in AIGC tools, thereby influencing the intention to adopt them. Meanwhile, information processing theory (Miller 1956) highlights how individuals collect and interpret information, indicating that information accuracy and privacy concerns shape trust. Although these factors do not directly affect adoption intention, they strengthen trust which in turn influences intention. Notably, effort expectancy also has a partial mediation effect through trust, underscoring its dual role in both directly influencing trust and shaping usage intention.

7.2 Findings from FsQCA

The fsQCA analysis reveals seven distinct configurations influencing graduate students’ adoption of AIGC tools, reflecting diverse needs and preferences. Performance expectancy, effort expectancy, and facilitating conditions consistently emerge as key factors, suggesting their collective influence on adoption.

Performance expectancy serves as the initial motivator, driving students to consider AIGC tools if they believe these tools will enhance academic performance. However, adoption may be hindered if the tools are perceived as too complex (effort expectancy) or if the necessary support (facilitating conditions) is lacking. Even with strong performance benefits the absence of ease of use or adequate support can limit adoption.

Privacy concerns exhibit a bidirectional relationship with trust. Students who trust the providers of AIGC tools are more likely to accept privacy risks, meaning that as trust increases privacy concerns may decrease. In certain configurations (e.g., M6 and M7) factors like information accuracy, trust, and privacy concerns are not decisive, suggesting that students may adopt AIGC tools despite privacy concerns if other factors, such as performance expectancy or facilitating conditions, are compelling.

8.1 Integrated Analysis of PLS-SEM and FsQCA Findings

The integration of PLS-SEM and fsQCA offers a more comprehensive understanding of the factors influencing graduate students’ intention to use AIGC tools (see Table 9). PLS-SEM identifies direct and indirect effects while fsQCA highlights how combinations of factors affect adoption.

For example, PLS-SEM reveals the strong direct influence of performance expectancy while fsQCA shows that performance expectancy when combined with effort expectancy and facilitating conditions has a greater impact on adoption. This suggests that while some factors are powerful on their own, their combined effects are even more significant for driving usage intention.

The combined insights from PLS-SEM and fsQCA highlight the importance of both performance benefits and ease of use (effort expectancy and facilitating conditions) in driving AIGC tool adoption. PLS-SEM focuses on the strength and direction of relationships while fsQCA shows how specific factor combinations lead to adoption. Together, these methods reveal the complexity of adoption behavior and the necessity of addressing both individual factors and their interactions.

8.2 Conclusion

In an era increasingly shaped by AI, this study uses PLS-SEM and fsQCA to explore the factors influencing the adoption of AIGC tools by Chinese graduate students. The developed AIGCT-SI model integrates elements from the UTAUT and adds factors such as information accuracy, trust, and privacy concerns.

Results from PLS-SEM reveal complex interactions among several factors affecting students’ intentions to use AIGC tools. Performance expectancy, effort expectancy, social influence, facilitating conditions, learning value, habit, and trust all play direct roles in shaping these intentions. Notably, trust acts as a full mediator for privacy concerns, information accuracy, and social influence, indicating that while these factors do not directly influence intentions they strengthen trust’s impact on adoption. Additionally, effort expectancy partially mediates through trust, suggesting that increased trust in AIGC tools enhances adoption among students in a competitive academic environment.

The fsQCA analysis uncovers seven distinct configurations reflecting the diverse needs and preferences of Chinese graduate students, highlighting the nuanced roles of privacy concerns and trust across these configurations. The combination of PLS-SEM and fsQCA results underscores the importance of performance expectancy, ease of use (effort expectancy and facilitating conditions), and trust in driving AIGC tool adoption. Adopting multiple analytical methods proves essential for capturing the complexity of adoption factors and encourages further investigation into potential inconsistencies between findings.

Insights from this research can inform the design and implementation of AIGC tools tailored to better meet the evolving needs of graduate students in an AI-driven landscape. Practical recommendations for universities and policymakers include simplifying user interfaces and enhancing data protection measures to foster trust among students. Developers, in particular, should implement strong data protection protocols alongside clear and transparent privacy policies to further build trust among users.