Factors Influencing University Students’ Continuance Intentions towards Self-Directed Learning Using Artificial Intelligence Tools: Insights from Structural Equation Modeling and Fuzzy-Set Qualitative Comparative Analysis

Abstract

This study investigates the intricate causal mechanisms of university students’ sustained use of artificial intelligence (AI) tools for self-directed learning (SDL) within the theoretical framework of self-determination theory (SDT). Employing a convenience sampling strategy, 387 university students from China were included in the study. Methodologically, we employed structural equation modeling (SEM) for the measurement and causal analysis, and we employed fuzzy-set qualitative comparative analysis (fsQCA) for the configurational analysis. The research results emphasize several important insights. Perceived usefulness, intrinsic motivation, and satisfaction play important roles in encouraging university students’ continuance intentions. Satisfaction emerges as a pivotal mediator, bridging the connection between perceived usefulness, intrinsic motivation, and continuance intention. The system quality, the information quality, and social interaction have significant positive influences on perceived usefulness. Perceived autonomy and perceived competence display strong correlations with both intrinsic motivation and satisfaction. Moreover, the results from the fsQCA show five configurations, in which the key factors collectively shape students’ continuance intentions through complex interactions through various configurations. The findings reveal diverse configurations by which university students form continuance intentions towards using AI tools for SDL, providing detailed insights into the profound and indirect impacts on forming continuance intention behaviors. This enriches and advances the current theoretical understanding.

1. Introduction

The disruption and development of artificial intelligence (AI) technology have rapidly transformed the educational system. AI is expected to transform education and revolutionize the teaching and learning process [1]. AI-driven tools such as chatbots and language models, including ChatGPT, are poised to enhance the educational experience [2]. The application of AI in higher education, particularly in promoting students’ self-directed learning (SDL), has demonstrated considerable promise [3]. However, this increased adoption also poses challenges, particularly in maintaining students’ motivation to continue effectively using these tools [4]. Although AI tools provide significant benefits for SDL, it has been observed that students often have difficulty in maintaining their continuance intentions in the long run. This suggests that while the initial acceptance of AI tools for learning is a critical step toward success, the continued use of technology and the continued intention are the real determinants of final success.

In recent years, research on SDL has garnered extensive attention. Autonomy refers to the capability of managing one’s own learning process [5]. Wei [6] posits that factors influencing SDL encompass metacognition, self-efficacy, learning strategies, goal setting, learning environments, and teachers. Renjie [7] argues that learners’ autonomy is hindered by a lack of learning strategies, attitudes, motivation, and access to resources. Regarding the relationship between SDL and AI, Mahendra et al. [3] aimed to delve into students’ perceptions of how AI-driven applications, as learning aids, facilitate their autonomy in learning. Xiao et al. [8] examined how teacher support based on self-determination theory (SDT), along with student attributes such as gender and achievement levels, impact AI-based learning in secondary school settings. Furthermore, Qiu et al. [9], utilizing SDT as a theoretical framework, explored how teacher support moderates the influence of students’ domain-specific knowledge on their need satisfaction and intrinsic motivation through information technology. While scholars have extensively studied the relationship between SDL and AI, encompassing specific factors like self-efficacy, satisfaction, and intrinsic motivation within SDL, the majority of the existing research focuses primarily on the direct role of AI in areas akin to learning assistance. There remains a dearth of comprehensive analyses on the full spectrum of factors influencing sustained intention and their configuration within this context.

Firstly, previous studies have often centered on the prerequisites influencing students’ initial acceptance of AI tools in SDL environments [10,11]. However, the adoption of AI tools encompasses not merely an acceptance process but, more importantly, a process of sustained usage by students [12]. Consequently, this study aimed to delve into the factors driving students’ willingness to continuously engage in SDL using AI tools, with a focus on the continuance intention. Secondly, there is a persistent lack of attention paid to intrinsic motivation when examining the intention for continued usage. For instance, Chang argues [13] that the Technology Acceptance Model (TAM) proposed by Davis overly emphasizes the influence of external motivation, neglecting the intrinsic motivations that shape technology acceptance. Furthermore, due to this oversight of intrinsic motivation, Roca and Gagné [14] found that the Expectation Confirmation Model (ECM), originating from research on continued usage behavior in workplace information systems, struggles to explain the continued usage behavior of e-learning users. Therefore, it is imperative to further explore, within the framework of SDT, the intricate interplay between intrinsic and extrinsic motivational factors and their impact on the intention to continuously use AI tools.

In the existing literature, SDT suggests that an individual’s behavior is directly related to their motivations, and it proposes two primary forms of motivation: extrinsic motivation and intrinsic motivation [15]. Ryan [16] discusses the relationship between intrinsic and extrinsic motivation and the basic human needs for autonomy, competence, and relatedness. Roca et al. [14] states that SDT facilitates the understanding of the user’s acceptance of an e-learning system and the influence of extrinsic and intrinsic motivation on the continued intention to use the system. The existing research also shows that the motivation for sustained participation by learners is considered to be a combination of internal factors, such as curiosity and personal interest, and external factors, such as the impact on job ability development and the university reputation, as exemplified in the study on MOOCs [17,18,19,20]. In [21], Yildirim et al. consider motivation as a factor to benefit students through the use of AI for SDL. However, they use a relative theoretical approach to explain this viewpoint [21]. It is noteworthy that the existing works lack combined studies of quantitative and qualitative comparative methods on college students’ sustained use of AI tools for SDL. In particular, what kind of cross combinations of variables can serve as sufficient conditions for students’ continuance intentions? Which factors can be considered the necessary conditions for the continuance intention? These questions cannot be fully addressed by structural equation models and require in-depth exploration through a mixed-method approach combined with configurational analysis.

In the context of the current research, this study aims to address these challenges. Using an integrated extrinsic and intrinsic motivation perspective, this paper applies SEM to explore the various factors of university students’ continued SDL with AI tools. Furthermore, by utilizing fsQCA, we examine the interplay of key variables and the factors influencing their choices. This paper seeks to deepen the understanding of students’ intentions to continue using these tools, providing insights for educators and AI tool developers to optimize learning outcomes.

2. Hypothesis Development

2.1. AI Tool Design Factors

According to DeLone and McLean [22], the system quality affects the system usage and satisfaction [23,24,25]. When users realize that AI tools provide a complete learning system or a convenient service, their willingness to continue using the AI platform will be positively affected. Therefore, DeLone and McLean’s assertion can be confirmed [23]. Wang et al. [26] investigated the impact of the e-learning system quality on the usage of and satisfaction with its implementation. Lin and Lu [27] verified the positive impact of the system quality on learners’ intentions to use the system. According to Seddon’s IS (information systems) success model, user satisfaction relies on six variables, which encompass the system quality, information quality, and perceived usefulness [28]. When users realize that AI tools provide a convenient service, their willingness to continue using the AI tools will be positively affected. Therefore, the following hypothesis is proposed:

H1. 

The system quality is positively related to the perceived usefulness.

The information quality can be ascertained by considering both the source and the content of the information [29], and it plays a key role in its use [30]. It has been demonstrated that the information quality exerts a significant impact on users’ utilization of information systems, particularly within the context of e-learning systems [31]. Saeed et al. [32] argued that the information quality positively impacts users’ online behavior. Li et al. [33] found that the information quality positively influenced the behavioral intention to reuse the e-learning system. Therefore, the following hypothesis is proposed:

H2. 

The information quality is positively related to the perceived usefulness.

Socialized interaction refers to the interactive behaviors of users [34]. Due to the separation of time and space and SDL, users expect to communicate and interact with other learners and receive support to improve their perception of the usefulness of AI tools. Hammond and Wiriyapinit [35] argued that socialized interaction in an MBA online distance learning course fostered a positive learning experience. Balaji and Chakrabarti [36] concluded that online discussion sites positively influence student interaction and facilitate the learning process. Consistent with the existing research, it is argued that social interaction positively impacts knowledge sharing and learning performance. Therefore, the following hypothesis is proposed:

H3. 

Socialized interaction is positively related to perceived usefulness.

2.2. Self-Determination Theory

Perceived autonomy refers to the user’s intrinsic psychological need to be able to learn autonomously [16]. Hence, SDT suggests that intrinsic motivation is enhanced if the activity has perceived autonomy. Users can learn autonomously based on their interests and hobbies, which is embodied in the form of intrinsic motivation [37]. This means that intrinsic motivation and satisfaction are likely to exhibit a positive association with the extent of learner autonomy when utilizing AI tools. Some examples include Sorebo et al. [37], who assume that the perceived level of autonomy has a positive effect on intrinsic motivation and subsequently confirm this hypothesis, and Hosseini et al. [38], who argue that perceived autonomy has a positive influence on satisfaction, which has been empirically verified. Therefore, the following hypothesis is proposed:

H4. 

Perceived autonomy is positively related to intrinsic motivation.

Perceived competence denotes the user’s aspiration to feel capable or in control [16]. SDT suggests that the ability to satisfy perceptions affects the level of motivation [16]. Sorebo [37] found that perceived competence has a significant positive effect on users’ satisfaction. SDT suggests that the ability to satisfy perceptions affects the level of motivation [39]. Therefore, the following hypothesis is proposed:

H5. 

Perceived competence is positively related to intrinsic motivation.

Intrinsic motivation refers to the engagement in an activity that is solely driven by one’s personal interest or beliefs regarding the activity [39]. Bhattacherjee [40] defined the IS continuance intention as the intention of an individual to persist in using an information system. Users who are interested in AI tools enjoy the process and have the desire to persist in using the AI tools. Previous research has confirmed that intrinsic motivation is an important antecedent of a user’s intention to continue using a technology [37]. Sorebo et al. [37] argue that intrinsic motivation has a positive influence on satisfaction, and this has been empirically verified. Therefore, the following hypotheses are proposed:

H6. 

Intrinsic motivation is positively related to perceived usefulness.

H7. 

Intrinsic motivation is positively related to satisfaction.

Perceived usefulness refers to the extent to which an individual holds the belief that utilizing a specific system will enhance their job performance [40]. Perceived usefulness is a crucial motivator for the user satisfaction and continuance intention towards information systems [40]. Perceived usefulness is a significant and direct factor influencing the continuance intention [41]. It also has a positive impact on satisfaction [40,42]. Perceived usefulness is a key motivator of the information system user satisfaction and continuance intention [40,41,42]. The perceived usefulness of AI tools depends on how much the learner’s use of the AI tools leads to an improvement in their learning performance. If the user’s performance decreases after using the AI tools, this will generate “negative feelings” and affect their willingness to continue using it. Therefore, the following hypothesis is proposed:

H8. 

Perceived usefulness is positively related to satisfaction.

2.3. Satisfaction and Continuance Intention

Satisfaction and the continuance intention are two important elements in expectation confirmation (ECM) theory, which posits that the continuance intention is primarily governed by the satisfaction derived from the usage of an information system (IS) [37]. Satisfaction is defined as the user’s reaction after using tools or systems [43]. Perceived user satisfaction is one of the success measures in IS [44]. Previous studies have confirmed the relationship between users’ satisfaction and their intentions to persist with a specific information system (IS) [45]. Thus, satisfaction has an impact on the AI tool continuance intention. The decision to keep using a system is affected by the user satisfaction and its perceived usefulness [46]. Furthermore, earlier research has identified this evaluation as a key predictor of the continued use of e-learning [47]. Combing the discussion of SDT, the following hypotheses are proposed:

H9. 

Perceived autonomy is positively related to satisfaction.

H10. 

Perceived competence is positively related to satisfaction.

H11. 

Intrinsic motivation is positively related to the continuance intention.

H12. 

Perceived usefulness is positively related to the continuance intention.

H13. 

Satisfaction is positively related to the continuance intention.

We can take a closer look at the model in Figure 1.

3. Research Methods

3.1. Measurement

3.1.1. Research Instrument

This study used questionnaire survey methods to measure our theoretical model. To ensure the reliability and validity of the questionnaire, all measurement items were based on well-established scales. We structured the questionnaire into two distinct sections. The first section aims to gather information on the participants’ age, gender, educational background, as well as their usage of AI tools, such as the specific AI tools utilized, the duration of continuous usage, and the frequency of utilization (Appendix A). The second section encompasses 35 items, adapted from scholars such as the authors of [13,37,40], and specifically designed to address issues pertinent to AI tools. All items were assessed utilizing a five-point Likert scale, where the response options spanned from 1 (strongly disagree) to 5 (strongly agree). The initial draft of the questionnaire was sent to three experts and scholars within our institution for feedback. Revisions were made based on their suggestions to ensure the content validity of the scale (Appendix B).

3.1.2. Pilot Test, Sample Selection, and Data Collection

To avoid potential issues such as an incomplete questionnaire and respondents misunderstanding the questions, a pilot test was conducted with 50 students. The participants were given clear instructions on how to complete the questionnaire, and any doubts they had were addressed. After all participants completed the questionnaire, the responses were collected. After that, a reliability analysis was conducted on the variables from the pilot test, primarily using the widely adopted Cronbach’s α reliability coefficient. The Cronbach’s α coefficients for all variables were found to be above 0.8, indicating that the measurement indicators have a high level of internal consistency reliability, and the survey data are relatively reliable.

From a regional analysis perspective in China, cities such as Zhejiang, Guangdong, Beijing, and Shanghai show higher levels of engagement in discussions about ChatGPT. Notably, extensive discussions have already taken place in southern and coastal regions of China [48]. Consequently, the sample for this study was primarily drawn from selected universities in Zhejiang Province, China. Convenience sampling was conducted based on the total population of 15,000 students. Participants were informed about the study’s purpose and confidentiality before they consented to participate. The survey was primarily conducted through the Wenjuanxing platform from late September to the end of November 2023. To ensure the completeness of the collected data, the time-tracking feature of the Wenjuanxing platform was activated during the survey distribution. Overall, 485 responses were collected from university students across various universities in Zhejiang Province. Strict measures were implemented to ensure the quality of the samples, consistent with the methodological recommendations outlined by Hair et al. [49], as described below. First, based on insights from the pilot test, it typically took 2–5 min to complete the questionnaire. Therefore, participants who completed the survey in an unusually short amount of time (100 s) were considered to have approached the task irresponsibly, leading to their data being classified as invalid. Secondly, a reverse item was included in the questionnaire. Data from participants who failed to accurately answer this item were considered invalid.

After rigorous screening, 98 invalid questionnaires were excluded. Among these, 32 questionnaires were discarded due to an insufficient response time, and 66 were excluded for incorrect responses to the reverse item. As a result, 387 valid questionnaires were retained for formal data analysis, representing a valid response rate of 79.8%. In terms of demographic composition, 44.5% of respondents were female, and 55.5% were male, indicating a relatively even gender distribution.

3.2. Data Analysis

Structural equation modeling (SEM) was employed to test the theoretical model and hypotheses, allowing for an understanding of the statistical relationships between the variables [50]. Several studies in the field of education have utilized SEM. For example, Kong et al. [51] developed a structural equation model to examine the effectiveness of programming learning on the development of computational thinking. Luan et al. [52] attempted to develop a model illustrating the relationships between students’ perceptions of social support and their engagement in online English learning. Karakose et al. [53] aimed to investigate the causal relationships among these variables using SEM with a sample of prospective mathematics teachers. Jung and Lee [54] conducted a study with 306 learners enrolled in MOOCs in South Korea to investigate the effectiveness of programming learning on the development of computational thinking. Moreover, Hsu [55] conducted a longitudinal study to assess the changes in technology integration proficiency among Taiwanese teachers from grades 1 to 9. Lin et al. [56] utilized structural equation modeling to uncover that the application of online learning strategies was predictive of students’ online learning outcomes. Goggins and Xing [57] investigated the correlation between participation behavior and learning, considering factors such as collective efficacy, social ability, and temporal dimensions. Therefore, this study adopted the SEM approach for several key reasons: First, the SEM method supports the estimation of latent variables in the external or measurement model and examines the interrelationships between the latent variables in the internal or structural model. Second, this approach facilitates a more comprehensive analysis of the influencing factors. Methodologically, SEM has become one of the most important statistical analysis paradigms in empirical research within the social sciences. We used SPSS 28.0 software and Amos26.0 software to conduct reliability and validity analyses and path analysis.

The fundamental idea behind the fuzzy-set qualitative comparative analysis (fsQCA) method is organizational configurational thinking, which involves generating various configurations to express similar or different social functions and impacts [58]. For example, Yu et al. [59] combined grounded theory and fsQCA to study the factors influencing the continuance intention in blended learning. In particular, fsQCA presents a holistic view by uncovering the combinations of conditions that foster the continuance intention in blended learning, granting a deeper understanding of the fundamental mechanisms and pathways involved. Li et al. [12] revealed various mechanisms of college students’ sustained engagement in online gamified learning based on the dual analysis of MPLUS and fsQCA. Additionally, research combining the PLS-SEM and fsQCA methods can be found in [10]. In general, the adoption of fsQCA in this study was motivated by several reasons: First, considering the limitations of symmetric statistical tests [58,60] and the existence of multiple realities, the same data set was also analyzed using fsQCA to explore the set relations of the conditions. Second, the study aimed to investigate the necessary conditions of the aforementioned variables for students’ continuance intentions towards using AI tools for SDL. Third, when the outcome is equivalent, the combination of antecedent conditions is not unique. This allows for a further exploration of the configurational effects and sufficient pathways that influence students’ continuance intentions. We applied fsQCA 4.1 software to analyze the configurations of the variables.

4. Results

4.1. Reliability and Validity Analyses

Utilizing SPSS 28.0, this study employed Cronbach’s alpha to assess the internal consistency of the scale. Generally, a Cronbach’s alpha above 0.7 signifies strong consistency within the scale [61]. The data analysis results in

Table 1. Results of confidence analysis and validity analysis.

Construct	Cronbach’s Alpha	Standardized Cronbach’s Alpha	KMO	Bartlett Sphericity Test			Structural Validity
				Chi-Square	Degree of Freedom	Significant Level
SQ	0.892	0.892	0.840	874.089	6	0.000	Favorable
IQ	0.907	0.907	0.889	1181.136	10	0.000	Favorable
SI	0.877	0.877	0.826	776.124	6	0.000	Favorable
PA	0.878	0.878	0.838	774.598	6	0.000	Favorable
PC	0.847	0.847	0.731	483.200	3	0.000	Favorable
IM	0.826	0.826	0.716	428.650	3	0.000	Favorable
PU	0.888	0.888	0.837	851.590	6	0.000	Favorable
SA	0.887	0.887	0.831	843.397	6	0.000	Favorable
CI	0.878	0.878	0.835	783.863	6	0.000	Favorable

reveal that the Cronbach’s alpha values for each construct—system quality, information quality, socialized interaction, perceived autonomy, perceived competence, intrinsic motivation, perceived usefulness, satisfaction, and continuance intention—as well as for the overall questionnaire, are 0.892, 0.907, 0.877, 0.878, 0.847, 0.826, 0.888, 0.887, 0.878, and 0.988, respectively. Therefore, with each construct and the entire questionnaire exceeding a Cronbach’s alpha of 0.8, it can be inferred that the scale utilized in this research exhibits high reliability.

Before performing factor analysis, it is essential to evaluate its suitability using the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s test of sphericity. Typically, a KMO value greater than 0.9 indicates a high suitability for factor analysis; values between 0.8 and 0.9 suggest that factor analysis is appropriate; values between 0.7 and 0.8 indicate moderate suitability [62]. Additionally, Bartlett’s test of sphericity assesses the validity of the factor analysis. A significance value (Sig.) less than 0.05 indicates correlations among the variables and thus supports the factor analysis. In this study, after conducting the factor analysis (Table 1), it was found that the KMO values demonstrate good structural validity, confirming the appropriateness of factor analysis for the variables. Furthermore, the questionnaire’s passage of the validity test indicates correlations among the variables, as shown by Bartlett’s test of sphericity with a significance value below 0.05. In conclusion, based on the factor analysis results, the questionnaire successfully passed the validity test.

We further assessed the reliability of the questionnaire using Amos 26.0 software, focusing on two indicators: the Composite Reliability (CR) and Average Variance Extracted (AVE). Generally, a CR value exceeding 0.7 and an AVE value above 0.5 indicate an acceptable consistency among the measurement variables [61]. Following the data analysis, it was observed that both the CR and AVE values surpass the recommended thresholds of 0.7 and 0.5, respectively (see

Table 2. Convergent validity test table.

Construct	Item Tag	Significance Estimate				Std.	SMC	AVE	C.R.
		UnStd.	S.E.	z-Value	p
SQ	P4	1				0.819	0.671	0.674	0.892
	P3	1.008	0.047	21.229	***	0.820	0.672
	P2	1.005	0.047	21.418	***	0.826	0.682
	P1	0.931	0.044	21.147	***	0.818	0.669
IQ	S5	1				0.825	0.681	0.669	0.910
	S4	0.973	0.047	20.791	***	0.827	0.684
	S3	0.913	0.046	19.933	***	0.805	0.648
	S2	0.952	0.047	20.373	***	0.816	0.665
	S1	0.975	0.048	20.377	***	0.817	0.667
SI	J4	1				0.786	0.618	0.644	0.878
	J3	1.078	0.055	19.428	***	0.818	0.669
	J2	1.09	0.056	19.313	***	0.814	0.663
	J1	0.989	0.053	18.591	***	0.791	0.626
PA	G4	1				0.792	0.627	0.651	0.882
	G3	0.971	0.05	19.35	***	0.803	0.645
	G2	1.054	0.053	19.954	***	0.822	0.676
	G1	1.044	0.053	19.576	***	0.810	0.656
PC	N3	1				0.798	0.637	0.657	0.852
	N2	1.039	0.052	19.957	***	0.811	0.658
	N1	1.059	0.052	20.335	***	0.823	0.677
IM	R3	1				0.832	0.692	0.714	0.909
	R2	0.988	0.047	20.849	***	0.830	0.689
	R1	1.004	0.047	21.470	***	0.845	0.714
PU	Y4	1.051	0.046	22.653	***	0.872	0.760	0.663	0.887
	Y3	1				0.831	0.691
	Y2	0.964	0.046	20.793	***	0.806	0.650
	Y1	0.934	0.044	21.277	***	0.818	0.669
SA	M4	0.923	0.045	20.645	***	0.802	0.643	0.621	0.831
	M3	1				0.792	0.627
	M2	0.979	0.05	19.486	***	0.799	0.638
	M1	0.976	0.052	18.674	***	0.772	0.596
CI	Z4	1				0.806	0.650	0.643	0.878
	Z3	0.934	0.045	20.573	***	0.801	0.642
	Z2	1.028	0.049	21.177	***	0.820	0.672
	Z1	0.965	0.048	19.963	***	0.781	0.610

Note: p cannot be equal to 0.00, but rather p < 0.01. “***” indicates that the number is less than 0.001. “SMC” represents Squared Multiple Correlation. “S.E.” represents Standard Error of Estimate. “UnStd.” represents Unstandardized. “Std.” represents Standardized. “AVE” represents Average Variance Extraction.

). This finding suggests that the internal consistency of the questionnaire items is high, thereby demonstrating acceptable reliability.

4.2. Structural Model Assessment

In this study, AMOS 26.0 software was employed to test the model and hypotheses previously discussed. The results, presented in

Table 3. Goodness-of-Fit indices.

Norm	Reference Indicators	Actual Results
CMIN/DF	1–3 is excellent, 3–5 is good	1.905
RMSEA	<0.05 is excellent, <0.08 is good	0.048
TLI	>0.9 is excellent, >0.8 is good	0.967
CFI		0.970
GFI		0.873
AGFI		0.851
CMIN	\	1023.092
DF	\	537

, show a chi-square value of 1023.09 with 537 degrees of freedom. The ratio of the chi-square value to the degrees of freedom is 1.91, which falls within the acceptable range of 1–3, indicating a good model fit. The Tucker–Lewis Index (TLI) and Comparative Fit Index (CFI) are both 0.97, exceeding the desired threshold of 0.9 and suggesting an excellent fit. The Goodness-of-Fit Index (GFI) and Adjusted Goodness-of-Fit Index (AGFI) are 0.87 and 0.85, respectively, both above 0.8 and thus within an acceptable range. Additionally, the Root-Mean-Square Error of Approximation (RMSEA) is 0.05, below the ideal value of 0.08. For more details about these indicators, one can refer to articles related to SEM like [63]. Overall, these statistics confirm that the model fits the data well.

4.3. Path Analysis and Hypothesis Test

Using standardized path coefficients to reflect the magnitude of the influence between the variables,

Table 4. Hypothesis testing results.

Hypothesis	Estimate	S.E.	Standard Error	C.R.	p	Outcome
H1: SQ→PU	0.087	0.089	0.035	2.55	0.011	Supported
H2: IQ→PU	0.581	0.573	0.061	9.435	***	Supported
H3: SI→PU	0.348	0.384	0.051	7.504	***	Supported
H4: PA→IM	0.603	0.592	0.046	12.793	***	Supported
H5: PC→IM	0.425	0.428	0.042	10.091	***	Supported
H6: IM→PU	−0.031	−0.033	0.042	−0.784	0.433	Not supported
H7: IM→SA	0.093	0.102	0.051	2.001	0.045	Supported
H8: PC→SA	0.216	0.239	0.072	3.325	***	Supported
H9: PU→SA	0.483	0.494	0.109	4.539	***	Supported
H10: PA→SA	0.229	0.246	0.083	2.963	0.003	Supported
H11: IM→CI	0.164	0.166	0.033	5.004	***	Supported
H12: PU→CI	0.824	0.779	0.05	15.435	***	Supported
H13: SA→CI	0.06	0.056	0.027	2.07	0.038	Supported

Note: In the table, p cannot be equal to 0.00, but rather p < 0.01. “***” indicates that the number is less than 0.001, and “→” indicates that the variable has an effect on the variable pointed to by the arrow. “S.E.” represents Standard Errors.

and Figure 2 present the structural equation path coefficients and hypothesis testing results. The relationships between the variables are all significant at the 99% confidence level (p < 0.05); therefore, the hypotheses H1–H5 and H7–H13 are supported. However, the initial hypothesis, H6, is not supported, as it did not yield significant results in this study.

5. Fuzzy-Set Qualitative Comparative Analysis

5.1. Calibration

Based on the insights from the existing empirical studies and structural equation modeling analyses, this research demonstrates that multiple variables impact the sustained autonomous learning of university students, and interactive effects may exist among these variables. Consequently, these variables were selected as antecedents to explore the mechanisms of their combined effects. Adhering to the recommendations of prior researchers [64,65],

Table 5. Calibration values.

Conception	Complete Affiliation	Strong Affiliation	Partial Affiliation	Partial Disaffiliation	Strong Disaffiliation	Absolute Disaffiliation
SQ	5	4.75	4.25	3.75	3	0
IQ	5	4.6	4.2	3.8	2.8	0
SI	5	4.6	4.25	3.75	3	0
PA	5	4.75	4.25	3.75	2.75	0
PC	5	4.67	4.33	3.67	3	0
IM	5	4.67	4.33	3.67	3	0
PU	5	4.75	4.25	3.85	3	0
SA	5	4.75	4.25	3.75	3	0
CI	5	4.75	4	3.75	3	0

adopts a six-value fuzzy-set assignment scheme for variable coding, where the fuzzy sets are assigned values of 1, 0.8, 0.6, 0.4, 0.2, and 0. Specifically, 1 indicates the presence of a condition, 0 denotes its absence, and the intermediate values signify degrees of presence falling between these extremes. This six-value assignment approach enables a more nuanced discussion across cases [66].

5.2. Analysis of Necessary Conditions

Before constructing the truth table, it is crucial to evaluate the significance of each antecedent condition. Following Ragin’s guidelines [67], the necessary conditions for the outcome should be a consistency score exceeding 0.90 and a coverage score surpassing 0.50. Furthermore, fsQCA emphasizes the notion of causal asymmetry, indicating variations in the formation process between the continuance intention in autonomous learning and its absence (e.g., low continuance intention in autonomous learning).

Table 6. Analysis of necessary conditions for the conditions and outcomes.

Conception	Result Variables
	CI
	Consistency	Coverage
SQ	0.703	0.853
~SQ	0.722	0.113
IQ	0.790	0.707
~IQ	0.663	0.109
SI	0.698	0.731
~SI	0.677	0.108
PA	0.695	0.792
~PA	0.714	0.112
PC	0.645	0.708
~PC	0.705	0.112
IM	0.703	0.712
~IM	0.722	0.116
PU	0.744	0.807
~PU	0.739	0.117
SA	0.732	0.760
~SA	0.705	0.113

Note: “~” denotes negation.

shows that all the antecedent conditions had agreement scores below 0.90, suggesting that none of these conditions was deemed necessary for either the presence or absence of the continuance intention in autonomous learning. By analyzing the variable and by using fsQCA 4.1 software, we investigated the necessary conditions for high intention, as depicted in Table 6. The analysis revealed that the consistency of all the influencing conditions was less than 0.9, suggesting the absence of any of the necessary conditions among them that influence learners’ intentions to use them. Therefore, it is essential to examine the configuration by integrating multiple antecedent conditions and evaluating the combined synergistic effects of various causal factors on the continuance intention in autonomous learning.

5.3. Selection Criteria for fsQCA Indicators

In our study, the selection criteria for the fsQCA indicators involved building a truth table to represent all possible combinations of the causal conditions [64]. Following the calibration and analysis of the necessary conditions, the truth table was constructed using frequency values and consistency thresholds. Following Ragin’s suggestions, a consistency threshold of 0.8 was strictly set in this study [67]. Consistency values greater than this threshold indicate a high level of agreement within each solution. Furthermore, a frequency threshold of two or more cases was employed for each combination, signifying a significant configuration requirement (

Table 7. Configurations that lead to high continuance intention towards autonomous learning.

Conception	Path
	P1	P2	P3	P4	P5
SQ	●		⊗	●	●
IQ	●	●	●	⊗	●
SI		•	⊗	⊗	•
PA	•	•	⊗	⊗	•
PC	●	●	●	⊗	●
IM	•	•	⊗	•	•
PU	•	•	•	•	⊗
SA	•	•	⊗	⊗	⊗
Consistency	0.950	0.957	0.928	0.981	0.976
Raw coverage	0.477	0.472	0.212	0.221	0.200
Unique coverage	0.006	0.010	0.037	0.058	0.011
Overall solution consistency	0.807
Overall solution coverage	0.747

Note: “●” represents the presence of a core casual condition; “•” represents the presence of a peripheral casual condition; “⊗” represents the absence of a peripheral casual condition; blank spaces indicate that a condition may be either present or absent.

). The consistency of each solution was reported, similar to significance testing, indicating the level of configuration required to produce the results. Additionally, coverage was provided for each solution, illustrating the degree to which the combined configuration of the five antecedents determined the continuance intention in autonomous learning.

5.4. Robustness Analysis

Two robustness tests were conducted in this study to validate the results. The first test involved adjusting the case frequency threshold, increasing it from 2 to 3 [68]. Additionally, the consistency threshold was modified from 0.80 to 0.90 while keeping the other conditions unchanged [66,69]. Following the analysis and a comparison of the results, it was noted that the adjusted grouping outcomes remained largely consistent, highlighting the robustness of the findings.

5.5. Results of fsQCA

The key results for high levels of the continuance intention in autonomous learning among university students are as follows.

According to Ragin’s recommendations, certain thresholds were established for constructing the truth tables in this study. These thresholds included a case frequency greater than 2, a consistency exceeding 0.8, and a PRI consistency higher than 0.75. The core condition chosen for this study was the simultaneous occurrence of intermediate and parsimonious solutions, and the configurations were presented based on the fits between these solutions. As a result, four configurations indicating a strong intention and four configurations representing a weak intention towards continuance in autonomous learning were identified (refer to Table 7). The consistency values for these configurations were above 0.9, indicating their high reliability. This suggests that Configurations P1 to P5 are sufficient conditions for demonstrating a strong intention towards continuance in autonomous learning.

Configuration P1 (SQ*IQ*PA*PC*IM*PU*SA): Quality Competence Synergy Configuration. The analysis of Configuration H1 indicates that the SQ, the IQ, and PC are the core conditions and play an important role in Configuration P1. Furthermore, in this study, PA, IM, PU, and SA were used as the peripheral conditions to synergize with the core conditions to further enhance the strength of the continuation intention. It is noteworthy that while these peripheral conditions are not universally required across all configurations, their presence undeniably provides additional support and momentum for the formation of the continuance intention.

Configuration P2 (IQ*SI*PA*PC*IM*PU*SA): Competence-Driven Engagement Configuration. The results of Configuration P2 highlight that the IQ and PC are the dominant core conditions within the configuration. This highlights the importance of high-quality information resources and students’ positive self-perceptions of their own learning abilities in contributing to the sustainability of SDL. Further analysis indicates that SI, PA, IM, PU, and SA, acting as peripheral conditions, variously amplify the impact of the core conditions on the continuance intention.

Configuration P3 (~SQ*IQ*~SI*~PA*PC*~IM*PU*~SA): Competence-Driven Value Configuration. The IQ and PC were the core conditions leading to the continuance intention. At the same time, we found that PU played a significant reinforcement role in the configuration, although it was considered a peripheral condition. Note that the SQ, SI, PA, IM, and SA are absent peripheral conditions in most configurations, indicating that while they are not prerequisites for the continuance intention, their absence in certain cases may weaken the strength of this intention. This highlights the need for a holistic approach when optimizing SDL environments, ensuring that no aspect is overlooked lest the overall effectiveness be compromised.

Configuration P4 (SQ*~IQ*~SI*~PA*~PC*IM*PU*~SA): System-Centric Motivation Configuration. The results observed that the SQ emerged as a pivotal condition across multiple high-consistency configurations. High-quality system support is an important factor in stimulating and sustaining university students’ intentions towards continued SDL. In addition, IM and PU as peripheral conditions show a reinforcing effect on the persistence intent of different configurations. Note that the IQ, SI, PA, PC and SA have potential effects. However, in the high-consistency configuration of this study, they are not necessary conditions but missing peripheral conditions.

Configuration P5 (SQ*IQ*PC*SI*PA*IM*PU*SA): Integrated Quality and Autonomy Configuration. The results show that the SQ, the IQ, and PC occupy key positions in this configuration and lead to the persistence intent towards SDL. In addition, SI, PA, and IM act as peripheral conditions that collaborate with the core conditions in a particular configuration to form continuous intent. Note that PU and SA are missing peripheral conditions. While they do not appear directly in the configuration P5 leading to sustained intent, their influence has not been completely negated. This finding also implies that even if learners do not fully perceive the system as useful or achieve complete satisfaction, the combined effects of the core and peripheral conditions can still uphold their willingness to continue learning.

6. Discussion

It is undeniable that AI systems and their applications are increasingly being integrated into classrooms and education in various forms [70]. In this paper, combing the AI tool design factors and SDT, we develop a structural equation model that expands the existing theoretical understanding of the factors influencing the continuous intention towards SDL supported by AI tools. This model serves as a foundation for future research endeavors in this domain. Furthermore, fsQCA determined the configurations of factors that result in either high or low continuance intentions, providing a fresh perspective on the complex relationships among these factors. The integration of the SEM and fsQCA methodologies provided comprehensive insights into both the individual internal factors and external influencing factors impacting students’ intentions to continue using AI tools.

The SQ is a core condition in Configurations P1, P4, and P5, playing an important role in these configurations. Hence, high-quality system support is a key factor in stimulating and sustaining university students’ continuous intentions towards SDL. As mentioned in the path analysis, it was found that the SQ significantly enhances the intentions of university students to continue their use of these tools. High-quality system support is a key factor in sustaining university students’ sustained SDL behaviors utilizing AI tools. This finding echoes other studies, such as one emphasizing the influence of the system quality on the perceived benefits of using ChatGPT, offering crucial insights into user and ChatGPT interactions and enhancing the comprehension of factors that bolster its effectiveness and utility [71]. Additionally, performances comparable to those enrolled in traditional economics courses were demonstrated by students utilizing the Smithtown Economics Intelligent Tutorial System (ITS for economics), yet only half the time was required to achieve these outcomes [72].

The importance of the IQ can be referred to as a core condition in Configurations P1, P2, P3, and P5. The significance of high-quality information resources, as well as students’ favorable self-evaluations of their learning capabilities, are crucial for fostering and maintaining autonomous learning practices among university students. As the path analysis reveals, the IQ positively affects perceived usefulness. This implies that university students place greater importance on the quality of information. The availability of quality, relevant, and comprehensive information ensures that the needs of students are met, thereby guaranteeing user satisfaction [73]. This is consistent with the previous conclusion that extrinsic motivation affects the continuous intention. The IQ highlights the importance of high-quality information resources in promoting the continuity of SDL. Study [74] considers that the high-quality outputs that ChatGPT generates have a positive on how students realize its usefulness, fostering their interest in and satisfaction with using it. These discoveries align with previous research, which shows that the output quality contributes to enhancing the perceived usefulness of technology [75,76]. Both the information and system quality are considered critical to customers because they simultaneously affect customer satisfaction [77].

PC serves as a core condition in Configurations P1, P2, P3, and P5, significantly contributing to their effectiveness. A learner’s positive self-perception of their abilities plays a pivotal role in sustaining motivation for SDL. The results from the path analysis also imply that PC has a positive influence on IM and SA. Therefore, attention should be paid to the perceived competence of college students, rather than to solely improving their learning efficiency. Providers focusing solely on improving learning efficiency may result in the misuse of AI tools and promote student laziness. Human laziness is significantly impacted by AI [78]. Furthermore, human psychological abilities, such as intuitive analysis, critical thinking, and creative problem solving, are being distanced from decision making [79]. However, Zhang and Aslan [80] argue that artificial intelligence in education has the potential to foster various types of interactions, boost learner engagement, create adaptive learning materials, provide metacognitive prompts, offer enriched learning environments, and ultimately enhance learning outcomes. This also confirms the importance of intrinsic motivation mentioned in the Introduction. Additionally, PC highlights the significance of learners’ positive self-perceptions of their abilities in sustaining motivation for SDL. Study [81] shows that perceived competence influences employees’ acceptance of collaborative robots, which provides a mechanism for understanding how such perceptions shape behavior.

Furthermore, SI, PA, IM, PU, and SA are not core conditions in any configuration. However, they act as peripheral conditions in certain configurations in concert with core conditions to further enhance the strength of the persistent intent. Note that these peripheral conditions are not required for all configurations, but they provide additional support and momentum for the formation of the continuance intention. This can also be reflected in the result of the SEM path analysis: PU, SA, and IM positively affect the continuance intention, thereby supporting hypotheses H11, H12, and H13; SA positively influences perceived usefulness, supporting hypotheses H3; PA positively impacts intrinsic motivation, supporting hypotheses H4. Specifically, active SI fosters learning support and social validation, PA imparts a sense of control over the learning process, IM sparks an inherent passion for learning, PU highlights the practical value of autonomous learning resources, and SA encapsulates an overall evaluation of the learning experience.

In conclusion, the analysis of the configurations showed that the SQ, the IQ, and PC are the most important factors. Specifically, a high-quality system platform, abundant and accurate information resources, and students’ positive perceptions of their own learning abilities together constitute the key drivers for students’ continued engagement with SDL. This finding suggests that the interaction between these key factors collectively affects students’ persistent intent behaviors. This reflects the combined effect of AI tool design factors, SDT, and persistent intent theory in this context.

6.1. Theoretical Contributions and Practical Implications

6.1.1. Theoretical Contributions

SDT as a paramount motivational theory in the social sciences has garnered widespread recognition within academia [8,82,83]. This study enhances its research model by including intrinsic motivation as a relevant variable. By applying this theoretical framework, our study effectively indicates the factors influencing university students’ continuance intentions towards SDL supported by AI tools. Its primary theoretical value lies in expanding the specific context and boundaries of SDT, thereby broadening its applicability and reinforcing its flexibility and adaptability.

The theoretical importance of this study stems from its exhaustive exploration of SDT, which incorporates extrinsic and intrinsic motivation. It introduces new perspectives and approaches to understanding the behavioral trends and patterns associated with the use of AI tools in the social sciences, thereby deepening our understanding of the relevant theories. Furthermore, this study offers valuable references for the exploration and implementation of AI tools in education. For example, generative AI tools such as ChatGPT and Sora, which possess the ability to produce human-like text and high-quality video content, present distinct cognitive and ethical considerations. This study addresses the unique challenges and opportunities associated with this technology by focusing on current students’ usage of popular AI-generated tools. Furthermore, we offer particularly novel and relevant insights pertinent to the educational domain.

Methodologically, our research pays attention to the combination of concurrently employing SEM and fsQCA in educational research. By taking this hybrid approach, we have gained a complete understanding of the factors that influence students’ continued intentions towards SDL supported by AI tools. This research therefore provides researchers and practitioners with more powerful and actionable insights. Our study further advocates the use of a combined SEM-fsQCA approach. This method can enable scholars to find the optimal solution through SEM and causal configuration, and to further apply fsQCA to obtain the ideal result, so as to improve the comprehensiveness of the research results. Finally, we propose a theoretical framework and research method that can be accepted by future research. In addition, it can also promote the progress and integration of AI technology in the field of education and the modernization of education.

6.1.2. Practical Implications

First, to provide high-quality system support, functions such as virtual reality, gaming, and simulations can be introduced by providers to offer practical or experiential learning opportunities. Second, providers are supposed to establish collaborations with governments, large companies, and universities to increase the knowledge base of AI tools and expand the database. At the same time, attention should be paid to filtering practical information quickly and effectively, and to providing feedback to college students on time. This aims to ensure high-quality and comprehensive information. Then, AI tools can be employed in student affairs to offer personalized degree planning, supplementary counseling, and advisory functions, among others [84]. During the learning process, solutions to issues encountered by university students can be offered by these tools, prioritizing the enhancement of their perceived competence. Furthermore, providers increase the emphasis on socialized interaction and genuine feedback. Surveys and other services can be conducted by providers to obtain timely and valuable suggestions. Comment sections or learning and communication groups can also be established to foster an environment where users can share ideas and engage in meaningful exchanges. AI techniques can simulate one-on-one human tutoring by offering personalized learning activities that address a learner’s cognitive needs and by delivering focused, timely feedback, all without requiring the physical presence of a human teacher [85], thereby satisfying the autonomy demand of university students.

7. Conclusions and Limitations

In this study, the design factors and intrinsic motivation factors of AI tools were included in the analysis to explore the factors that affect college students’ continuance intentions. In addition, the complex relationships between these multiple factors were rigorously examined using fsQCA, providing detailed insights into their interactions. The research model was validated, and we obtained the following conclusions. First, perceived usefulness is identified as the core factor affecting university students’ continuance intentions towards AI tools. Factors such as the system quality, the information quality, and socialized interaction play significant roles in shaping the perceived usefulness, with the information quality playing the most important role. Intrinsic motivation factors have a substantial positive impact on users’ intentions to continue using AI tools. The formation of intrinsic motivation is mainly influenced by users’ perceived autonomy and perceived abilities, of which perceived autonomy plays the most key role. Satisfaction significantly enhances users’ intentions to continue using AI tools. Perceived competence, perceived autonomy, perceived usefulness, and socialization interaction all contribute to the formation of satisfaction, with perceived competence exerting a more pronounced influence. In addition, the fsQCA analysis highlights the importance of high-quality system platforms, rich and accurate information resources, and students’ positive perceptions of their own learning abilities in motivating them to continue using AI tools for SDL. These conclusions provide theoretical support for the dissemination and innovative application of AI tools. In order to maximize the potential of AI tools, it is critical to consider both their functionality and utility, as well as students’ motivation for SDL. In addition, a mechanism should be considered to improve AI tools and take advantage of their ability to streamline the learning process. Moreover, it is essential to transfer the advantages of AI tools to users and enhance their abilities.

This study has certain limitations: Firstly, it mainly focused on the university student demographic, with relatively little exploration of other groups that use AI tools, such as corporate employees, primary and secondary school students, etc. Furthermore, the study focused only on students from China. This could limit the generalizability of the study’s conclusions. Future research should include a broader range of user groups to more comprehensively understand the needs and usage patterns of different AI tool users. Secondly, this study mainly focused on factors that have a positive impact on the willingness to continue using and may have overlooked the obstacles, negative experiences, cultural backgrounds, and technology acceptance that might hinder continued use. Cultural differences in educational practices and attitudes towards technology can affect the applicability of the results to other contexts. Therefore, future research could provide a more balanced perspective and offer insights into how to overcome barriers to continued use. Furthermore, a longitudinal approach would allow for a better understanding of how these factors influence students’ behaviors and how their importance changes with longer exposure to AI tools.