Exploring How to Improve Programming Learning Experience with Generative AI Tools: An Integrated Means-End Chain and fsQCA Approach

Abstract

Recent advancements in generative artificial intelligence (AI) are poised to transform education, but doubts persist about their effectiveness in enhancing the learning experience. Previous studies were more “technology-driven” rather than “learner-driven” and largely neglected how different configurations (specific combinations of variables) affect generative AI-assisted learning experience. Based on the means-end chain theory, the research identifies functional and psychological factors that influence programming learning experiences with generative AI tools, including tool ease of use, interactivity, quality of the result, knowledge level, and self-efficacy. Through an analysis of the survey results of 221 university students and fuzzy-set qualitative comparative analysis (fsQCA), the findings reveal two types of generative AI-assisted programming learning enhancement patterns: (1) the dual-driven pattern, which combines functional and psychological consequences; (2) the psychological consequence-dominated pattern. In addition, although the five factors cannot alone constitute the necessary conditions for a learning experience, tool ease of use, interactivity, and knowledge level play an indispensable role in improving the generative AI-assisted learning experience. The findings offer valuable insights for programming educators, deepening their understanding of how generative AI tools influence students’ programming learning experiences and informing the development of more effective strategies for integrating them into computer science education.

Plain Language Summary

How Generative AI Tools Help Students Learn Programming: Exploring What Makes The Learning Experience Better

This study explores how generative AI help university students improve their programming learning experience. This research used both functional consequences and psychological consequences as a measure of generative AI-assisted learning experience based on the means-end chain theoretical framework, and extended the configurational understanding of the factors that can affect learning experience with generative AI tools. This research collected data from 221 university students through a questionnaire. Then, this study employed fsQCA to analyze the configuration of the coupling of factors to generate high and low level of knowledge acquisition effect and learning satisfaction effect. This study found that no single factor alone can guarantee a great programming learning experience. Factors like how easy the AI tool is to use, how interactive it is, how accurate the feedback is, and the learner’s knowledge level or confidence all play a role, but they are not enough by themselves. Instead, the study identified four different combinations of these factors that can lead to a better learning experience. The study also found two main patterns that improve the programming learning experience with AI tools. The first is a “dual-driven” pattern, where both functional and emotional outcomes matter. This pattern includes three key combinations: 1) an easy-to-use tool combined with a good knowledge level; 2) a tool that’s interactive and provides accurate feedback, along with a good knowledge level; and 3) a tool that’s easy to use, interactive, and gives good feedback, combined with the learner’s confidence. The second pattern focuses more on emotional outcomes, where the key combination for better satisfaction with the learning experience is having a tool that’s easy to use, interactive, and a good match with the learner’s knowledge.

Keywords

generative AI programming learning experience means-end chain fsQCA

Introduction

In a digitally driven world, mastery of computer programming skills rapidly transitions from a specialized skill to a vital necessity. Programming has been recognized as a core competency essential for university students (Nouri et al., 2020). Computer programming involves designing and building computer programs to perform specific tasks. Through programming learning, learners can develop systematic thinking and acquire the ability to solve complex problems, enabling them to better adapt to information circumstances where technology shapes our daily existence and future possibilities (Groothuijsen et al., 2024).

Despite the significant emphasis on programming education in university curricula, students studying programming languages such as Python face various challenges. Survey data indicate that dropout rates in introductory programming courses are generally high, reaching over 30% in some contexts (G. Fan et al., 2025). Moreover, studies across different countries and regions consistently report that students experience considerable difficulties, particularly related to the heavy cognitive load, the abstract nature of the content, and insufficient instructional support (Demir, 2022). For many learners who begin programming courses, factors such as a steep learning curve, high cognitive load, and the high level of abstraction in content can negatively influence their sustained learning motivation (Utreras & Pontelli, 2021). In resource-constrained contexts, such as developing countries, insufficient access to qualified instructors and digital infrastructure has compounded these difficulties (Bull & Kharrufa, 2024; G. Fan et al., 2025). Besides, traditional programming education primarily relies on direct instruction from teachers, providing insufficient support for learners’ independent exploration and understanding of complex problems, which limits the accessibility and depth of programming learning (X. Y. Hou et al., 2023).

Embracing the rapid pace of AI, AI-assisted learning has become a subject of interest for scholars and educators, as it has emerged as a significant driver of innovation in education. The application of generative AI tools (such as ChatGPT and Google’s Bard) can generate code through natural language and explain complex programming concepts, effectively lowering the threshold for learning programming (Ezeamuzie, 2023). Compared to traditional teaching methods, generative AI tools for programming learning have been highlighted for their capacity to provide personalized and adaptive learning experiences across different cultural contexts. Tlili et al. (2023) pointed out that generative AI offers learners real-time and customized feedback and guidance, thereby improving the efficiency and experience of programming learning. Many studies have validated the potential of generative AI in programming education. On the one hand, such tools can generate code examples in real-time and provide step-by-step explanations to help learners understand complex programming concepts, thereby enhancing the intuitiveness of the learning process (Ryan et al., 2021). On the other hand, they effectively reduce frustration for beginners by offering instant feedback (Tlili et al., 2023) and support personalized learning paths, promoting active learner engagement (Zawacki-Richter et al., 2019).

Despite the promising potential of generative AI tools in programming education, they also introduce some new challenges. On one hand, learners may face information overload when using AI-assisted tools, as the generated code and feedback can be too complex, making it difficult for learners to understand and apply them (Steyvers & Kumar, 2024). Moreover, while generative AI can provide instant feedback and guidance, whether such feedback truly addresses the individualized needs of learners—especially those with varying programming skill levels—remains a significant concern (Ryan et al., 2021). On the other hand, while AI’s automation can enhance the efficiency of programming learning, it may also lead to dependency among some learners, weakening their active learning and problem-solving abilities, which could negatively impact their overall learning experience (Miller, 2023). These concerns underscore the need for more inclusive evidence from regionally diverse educational contexts to better understand the opportunities and risks associated with AI integration in programming education. Thus, this study aimed to present a systematic analysis of complex configurations for achieving high levels of students’ programming learning experience with Generative AI tools.

Based on the objective, two specific research questions were posed: (1) What functional and psychological factors shape university students’ programming learning experiences with generative AI tools? (2) How do different configurations of these factors jointly influence students’ learning outcomes and satisfaction? The findings will enrich the existing literature on AI-assisted learning in programming education, which has traditionally focused on the net effects of instructional tools without considering their complex configurations.

Literature Review

Generative AI for Education

Generative AI refers to a category of artificial intelligence technology that utilizes generative models to create new content, such as text, images, and audio (Al-Emran et al., 2024; Liu et al., 2024). These tools are characterized by their adaptability, scalability, and capability to offer instant responses, which significantly extend the role of traditional educational technologies that mainly function as static repositories or assessment systems (Becker et al., 2022). In the context of programming education, AI tools demonstrate several key features. They can serve as intelligent tutors by delivering personalized feedback, as code-generation assistants by producing working examples or alternative solutions, and as learning companions by supporting iterative practice and exploration (Sajja et al., 2024). Such affordances expose learners to diverse problem-solving strategies and help cultivate better coding practices. Recent research has further explored AI-assisted programming environments. X. Hou et al. (2024) proposed the Code Tailor system, which fully utilizes large language models to provide personalized Parsons puzzles for students in need, aiming to improve each student’s engagement and cognitive level in learning programming while enhancing convenience. Kazemitabaar et al. (2023) developed Code Aid, an LLM-based programming assistant, which provides more personalized learning feedback through timely human-computer interaction during the learning process. This balances educators’ needs for innovation in teaching and significantly enhances students’ engagement when encountering complex concepts in the learning process. In addition, Kwak et al. (2023) examined how artificial intelligence tutoring systems offer personalized feedback and support tailored to students’ learning progress and understanding. It provides timely, dynamic information feedback to help students choose more efficient learning paths to master more difficult programming concepts. G. Fan et al. (2025) provided evidence from China showing that AI-assisted pair programming enhanced students’ intrinsic motivation, reduced programming anxiety, and improved programming performance to levels comparable to those of human–human pair programming. Beyond these developments, emerging evidence from non-Western contexts also shows that generative AI tools are increasingly integrated into programming-related learning. Archana et al. (2025) reported that Indian university students actively use AI tools to support code understanding and debugging. Al-Dokhny et al. (2024) further demonstrated that multimodal AI systems can enhance task performance among Asian students by improving code interpretation and problem-solving efficiency.

Factors Influencing AI-Assisted Learning Experiences

With the emergence of generative AI tools, such as ChatGPT, AI tools are gradually transforming traditional approaches to programming learning, serving as novel support tools. Existing research indicates that the application of AI tools significantly reduces learners’ cognitive load through effective interaction and code suggestions (Evans et al., 2024), thereby enhancing the efficiency of completing programming tasks (Hartley et al., 2024). In addition, the code generation capabilities of AI tools can help beginners overcome frustration caused by programming syntax and logic issues, thereby enhancing their self-efficacy (Bandura, 1997). In addition to aspects of cognitive support, the application of AI tools significantly influences learners’ emotional experiences through personalized support and conversational interaction (Vistorte et al., 2024). For instance, generative AI utilizes natural language understanding technologies to provide explanations that align more closely with learners’ needs, thereby reducing anxiety during the learning process (Yan et al., 2024). Although some researchers have identified factors such as cognitive load and interactivity that significantly influence the learning experience with generative AI tools for programming, other scholars present differing perspectives regarding these influencing factors. For example, Wang et al. (2022b) argued that under certain conditions, the interactivity between generative AI tools and learners, particularly its impact on the programming learning experience, is moderated by students’ individual learning styles and other personal characteristics. Similar findings have been reported in non-Western contexts, where learner dispositions and contextual norms further shape how students benefit from AI-assisted programming. For example, Archana et al. (2025) found that Indian learners’ self-efficacy and perceived support strongly influenced their engagement with AI tools, while Ali and Shaban (2025) showed that individual traits significantly condition the emotional and cognitive gains students derive from AI-driven learning systems.

This further illustrates that the learning experience with generative AI tools for programming results from the combined effect of multiple factors, with complex causal relationships inherent in its driving mechanisms. It underscores the limitations and partiality of models driven by a single factor, thereby necessitating an analysis of the driving mechanisms of generative AI tools in programming learning experiences from a configurational perspective.

Means-End Chain Theory and Fuzzy-Set Qualitative Comparative Analysis in Education Research

The means-end chain theory and the qualitative comparative analysis method provide new perspectives and methodological support for educational research by integrating theoretical models and analytical approaches. The means-end chain theory, proposed by Gutman (1982), is a theoretical model that analyzes the connection between consumer behavior and personal values through the “attribute-consequence-value” framework (Zhou et al., 2020). While this theory has been widely applied in fields such as tourism management (Yin et al., 2024) and consumer behavior analysis (Kumar et al., 2024), it has also undergone specific development and practical application in the field of educational research. Studies have shown that the use of the MEC framework offers valuable insights for designing educational tools by analyzing both the perceived value and functional needs of students regarding educational technologies (N. Wang et al., 2022a). However, applying generative AI tools based on the MEC framework in programming education remains a relatively unexplored research direction.

The fuzzy-set qualitative comparative analysis (fsQCA) method possesses distinct advantages. It demonstrates strengths in revealing causal relationships within complex educational contexts based on multi-factor configurational analysis (Greckhamer et al., 2018). It has been widely applied in fields such as healthcare management (Y. Wang et al., 2022b) and hotel management (Rey-Martí et al., 2017). In recent years, scholars have introduced the fsQCA method into educational research. For example, Sánchez-Mena et al. (2019), based on survey data, used the fsQCA method to analyze the combinations of factors influencing teachers’ use of educational video games. Wu and Wang (2022) applied the fsQCA method to analyze configurational paths influencing the efficiency of MOOC development in Chinese universities. Fu et al. (2023), based on evaluation data from 27,316 students, used the fsQCA method to explore the combinations of factors influencing MOOC learners’ satisfaction. The above literature primarily adopts the single fsQCA method to analyze complex causal relationships. These studies demonstrate that fsQCA can identify various factor combinations, revealing the nonlinear relationships that influence educational outcomes. This provides methodological support for exploring the complex interactions between educational tools and learner behaviors.

Although previous studies have typically applied MEC or fsQCA in isolation, recent research suggests combining the two can generate more comprehensive insights. MEC provides a structured framework to conceptualize the hierarchical linkages between attributes, consequences, and values, while fsQCA is particularly effective in capturing configurational causality across multiple conditions. For instance, Wang et al. (2022a) integrated MEC with fsQCA to examine how individual characteristics, social capital, and perceived value jointly shape users’ continuance intention toward innovative wearable products. Their findings highlight that MEC can define the attribute–consequence–value chain, whereas fsQCA can empirically identify multiple alternative configurations leading to the same outcome. This illustrates the complementary nature of the two approaches: MEC offers the theoretical lens for structuring conditions, and fsQCA reveals the complex, non-linear causal pathways among them. Building on this rationale, the present study employs an integrated MEC–fsQCA approach to investigate the multifaceted pathways through which functional and psychological factors impact learners’ experiences with AI-assisted programming.

Summary of Literature Review

In the context of learning with generative AI tools, two gaps exist in the investigation of learners’ experiences. First, the existing literature on Generative AI and introductory programming reveals a limited focus on student-centered perspectives (Amoozadeh et al., 2023). Second, research on generative AI tools often overlooks learning-related aspects, particularly students’ thought processes, motivations, and perceptions of these tools. This gap in student-centered research highlights a lack of understanding of “what students actually do” (T. Wang et al., 2023). Many studies primarily target the performance of AI tools (Zawacki-Richter et al., 2019). While there is a positive trend in the literature toward moving beyond merely evaluating the capabilities of large language models, a strong call remains for research that delves deeper into the complexities of students’ experiences, perceptions, and interactions in this field (Vistorte et al., 2024).

Another gap is that previous studies have predominantly employed quantitative analysis methods, such as regression analysis and structural equation modeling, to investigate the factors influencing the learning experience (Zawacki-Richter et al., 2019). However, these research methods primarily control for single-variable factors, overlooking the complex relationships and interactions among multiple dimensions that influence the learning experience. As a result, they fail to comprehensively reveal how generative AI tools optimize students’ overall learning experience through the combined effects of various factors across different dimensions (Yan et al., 2024). Because distinguishing different kinds of complex causality can better reflect authentic learning conditions, there is a need to investigate programming learning experiences with generative AI tools from a configurational perspective.

Therefore, this study is expected to fill these gaps through an integrated means-end chain and fsQCA approach. The means-end chain model is employed to identify the attribute-level and outcome-level factors that influence learning experiences with generative AI tools for programming. Then, the fsQCA method is applied to investigate the relations between these elements. This research contributes to the following areas: (1) expanding the learner-driven research perspective and establishing a research model that integrates functional and psychological consequences of the learning experience; (2) providing new insights into how the determinants of AI-assisted learning experiences jointly shape learning outcomes; (3) providing strategies to improve programming learning experiences in the AI-assisted learning environments.

Theoretical Framework

The means-end chain model is a theoretical framework used to analyze the impact of product attributes on consumers’ usage behavior and value orientation. The model analyzes consumers’ product usage behavior through the outcomes generated by the product attributes (the attribute-result chain). Essentially, AI-assisted programming learning is a process in which AI tools are applied to provide services to programming learners. The learning experience can be viewed as the result of the combined effects of the tool attributes and the individual attributes of the learners. Therefore, the attribute-result chain analysis framework, derived from the means-end chain model, is used to analyze the formation mechanism of the AI-assisted programming learning experience.

For the attribute-level factors, the tool attributes inherent to the application of AI are the fundamental reasons that attract learners to use AI tools in the programming learning process. These are primarily reflected in three variables: tool ease of use, interactivity, and quality of the results. Tool ease of use refers to the degree to which the tool is convenient for learners, including aspects such as learnability, comprehensibility, and ease of operation (Davis, 1989). Interactivity refers to the tool’s ability to engage in efficient communication and feedback with learners, including real-time responsiveness, personalized interaction, and dynamic regulation (Means, 1994). The quality of the results refers to the correctness and reliability of the content generated by the tool, including aspects such as accuracy in meeting task objectives, comprehensiveness of information, and logical coherence (Gulwani et al., 2017). Since AI-assisted programming learning essentially falls within the conceptual category of service, the service experience is influenced not only by the attributes of the service-oriented product but also by the characteristics of the users themselves. Therefore, in addition to the tool attributes inherent to the application of AI, the attribute-level factors should also include the individual characteristics of the learners. These personal attributes primarily consist of two variables: self-efficacy (Ithriah et al., 2020) and knowledge level (Zviel-Girshin, 2024).

According to the means-end chain theory, learners’ experience in AI-assisted programming learning results from the combined effects of the attributes of the generative AI tools and the individual characteristics of the learners. The learning experience is primarily reflected in two aspects: functional consequences and psychological consequences. Among these, functional consequences mainly reflect the knowledge acquisition brought to learners by AI-assisted programming learning (Sun et al., 2024). Therefore, functional consequences can be measured through the knowledge acquisition variable. Psychological consequences mainly reflect the learners’ satisfaction with the assistance provided by AI-assisted programming learning (Shahzad et al., 2024), and can be measured through the satisfaction variable.

Based on the means-end chain model, this research constructs an “attributes-consequences” analytical framework for the learning experience of AI-assisted programming, as shown in Figure 1.

Figure 1.

Research framework.

Methods

Procedure

The methodological research model is shown in Figure 2. This research begins by identifying the factors that influence learning experiences with generative AI tools for programming, based on the means-end chain theory. Then, an online survey was conducted, targeting university students who have used generative AI tools to assist with programming learning, and the survey data were cleaned. Finally, the fsQCA method was employed to identify configurations leading to functional and psychological experience outcomes, thereby determining the formation mechanism of learning experiences with generative AI tools for programming.

Figure 2.

Methodological framework for the research.

Fuzzy-Set Qualitative Comparative Analysis Method

To further investigate the complex combinations of factors influencing learning experiences with generative AI tools for programming, the fuzzy set qualitative comparative analysis method is employed to explore the causal relationships between learning experience, programmers’ personal characteristics, and the learning environment. Unlike traditional empirical methods, which typically focus on binary relationships between independent and dependent variables, fsQCA allows for the analysis of causal relationships that exhibit variation across multiple levels. Furthermore, the incorporation of the “fuzzy” technique enables the analysis of conditions that encompass variations between high and low levels, rather than being limited to crisp boundaries.

This study employs fsQCA for the following reasons: First, learning experiences with generative AI tools for programming result from the interaction of multiple factors and cannot be attributed to a single variable. The fsQCA method analyzes complex causal relationships from a set-theoretic perspective (Misangyi et al., 2017), thereby uncovering the pathways through which combinations of multidimensional factors influence the learning experience. Second, fsQCA addresses the asymmetrical causal relationships often seen in educational contexts, where factors may contribute differently to positive or negative outcomes (J. Fan & Tian, 2024). Given its ability to handle complex interdependencies and asymmetry, fsQCA allows for a more nuanced understanding of the factors that influence learning experiences. Third, the fsQCA approach is known for its low sample size requirements and is applicable even when the number of cases is fewer than 50 (Chenxi & Haijie, 2023; Shen et al., 2025). Moreover, empirical research typically recommends a minimum sample size of more than 10 times the number of measurement items (Chin, 1998).

Participants and Data Collection

To validate the proposed conceptual research model, a self-constructed survey questionnaire was developed for data collection and analysis. The questionnaire consisted of two sections: the first part recorded the basic information of the respondents, and the second part covered the variables involved in the research. The design of the questionnaire is based on a review of the literature; however, the content has been adjusted to suit the research context. The questionnaire items were adapted from the following prior studies: Tool ease of use (Davis, 1989), interactivity (Means, 1994), accuracy of results (Gulwani et al., 2017), self-efficacy (Bandura, 1997; Ithriah et al., 2020), knowledge level (Zviel-Girshin, 2024), knowledge acquisition (Sun et al., 2024), and learning satisfaction (Shahzad et al., 2024). Each dimension contains three unique questions, totaling 21 questions, which are measured using a 5-point Likert scale (1 = strongly disagree, 2 = disagree, 3 = neutral, 4 = agree, 5 = strongly agree). The Likert scale, as a core psychometric tool in social science research, is widely applied by presenting hypothetical statements and asking participants to indicate their level of agreement (Chyung et al., 2017). Detailed measurement indicators are provided in Appendix A.

After comprehensive validation of the questionnaire content, the survey was distributed for 20 days (from August 15 to September 4, 2024) through the professional online survey platform Wenjuanxing (https://www.wjx.cn), which is widely used by over 30,000 enterprises and 90% of universities in China. To facilitate access for respondents, the survey link was distributed through social media channels, including WeChat and QQ. Before completing the survey, participants were provided with clear instructions on how to fill out the questionnaire, emphasizing the principles of anonymity and confidentiality. Respondents were required to select an “agree” option to proceed with the survey. The questionnaire was designed with a mandatory completion mechanism, ensuring that respondents had to complete all sections before submitting their responses. To ensure data quality, we screened responses based on completion time. The mean completion time was 3.42 min (SD = 1.15), with the majority of respondents (approximately 92%) finishing within a reasonable range. Following established practice for online survey quality control, responses completed in less than 30 s or longer than 10 min—times considered too short to reflect careful reading or too long to reflect normal engagement—were excluded. A total of 17 cases were removed based on this time-screening criterion. After data cleaning, 221 valid questionnaires were retained for analysis, and the data were further processed using SPSS 27.0, IBM SPSS Statistics, Version 27.0 (IBM Corp., Armonk, NY, USA and fsQCA.

Results

Descriptive Analysis

According to the descriptive statistics of the survey sample, Table 1 shows a nearly equal gender distribution, with males accounting for 54.30% and females for 45.70%. This gender ratio demonstrates a good balance, consistent with the trend of increasing gender equality in modern programming education (Groothuijsen et al., 2024). In the use of AI-assisted programming tools, ChatGPT is the most used tool, with 53.39% of students utilizing it, followed by Copilot (including Copilot Chat), with a usage rate of 30.32%. This indicates that the application of generative AI tools is becoming increasingly prevalent in undergraduate programming learning, with ChatGPT emerging as the first-choice tool due to its ease of use and high interactivity.

Table 1.

Demographic Characteristics of the Respondents.

Item	Description	Frequency	Percentage
Gender	Male	120	54.30
Gender	Female	101	45.70
Age distribution	18–20	79	35.75
	21–23	114	51.58
	24–26	26	11.76
	27+	2	0.91
Education	Bachelors	103	46.61
	Masters	81	36.65
	PhD	37	16.74
AI programming software used most often	ChatGPT	118	53.39
	Copilot (Including Copilot Chat)	67	30.32
	Code Gee X	6	2.71
	Codium	9	4.07
	Other	21	9.51

Table 1 presents the detailed demographic characteristics of the respondents. Regarding age distribution, most respondents are between 18 and 23 years old, with 35.75% of students aged 18 to 20 and 51.58% aged 21 to 23. Students aged 24 and above are relatively fewer, with 11.76% aged 24 to 26 and only 2% aged 27 and above. In terms of educational background, most respondents are bachelor’s students (46.61%), followed by master’s students (36.65%) and PhD students (16.74%).

Reliability and Validity Evaluation

To ensure the reliability and validity of the questionnaire data, tests were conducted using SPSS 27.0, IBM SPSS Statistics, Version 27.0 (IBM Corp., Armonk, NY, USA. The evaluation results for each variable are shown in Table 2. The Cronbach’s α coefficients for all variables range between .7 and .9, and the composite reliability (CR) values exceed 0.7, indicating high reliability for all variables in the questionnaire. It is generally considered that when the factor loading and average variance extracted (AVE) of each variable are greater than 0.5, the variables demonstrate good convergent validity (Fornell & Larcker, 1981). The research found that the factor loading of all variables exceeded 0.7, and the average variance extracted (AVE) values ranged from 0.690 to 0.720. This indicates that the measurement items in the questionnaire demonstrate excellent convergent validity.

Table 2.

Reliability and Validity Test.

Results variable	Items	Factor loading	Cronbach’s alpha	CR	AVE
Tool ease of use (TEOU)	TEOU1	0.869	.766	0.874	0.699
	TEOU2	0.835
	TEOU3	0.802
Interactivity (IN)	IN1	0.858	.778	0.874	0.699
	IN2	0.822
	IN3	0.827
Accuracy of results (AOR)	AOR1	0.855	.799	0.882	0.714
	AOR2	0.882
	AOR3	0.796
Knowledge level (KL)	KL1	0.619	.789	0.882	0.720
	KL2	0.938
	KL3	0.947
Self-efficacy (SE)	SE1	0.879	.771	0.868	0.690
	SE2	0.913
	SE3	0.680
Knowledge acquisition (KA)	KA1	0.871	.785	0.878	0.706
	KA2	0.786
	KA3	0.861
Learning satisfaction (LS)	LS1	0.854	.794	0.886	0.718
	LS2	0.892
	LS3	0.818

To further verify the model’s convergent and discriminant validity, we compared the square root of the AVE values for each variable with the correlations between different variables. As shown in Table 3, the square root of the AVE value for each variable exceeds its correlation with any other variable. All latent constructs satisfy this criterion. Therefore, the measurement model demonstrates good reliability and validity.

Table 3.

Discriminant Validity.

Items	TEOU	IN	AOR	KL	SE	KA	LS
TEOU	.745
IN	.523	.741
AOR	.445	.407	.762
KL	.387	.308	.710	.813
SE	.444	.294	.444	.467	.873
KA	.625	.558	.493	.415	.502	.726
LS	.647	.677	.177	.109	.250	.554	.721

Note. Correlation matrix and the square root of AVE (in bold).

Calibration Process

Calibration is the process of assigning set membership scores to cases (Schneider & Rohlfing, 2016). Before conducting fsQCA, it is necessary to calibrate each conditional variable based on the actual survey data collected from the selected cases. Referring to the studies of scholars such as Fiss (2011) and Greckhamer (2016), the quartile method was used to calibrate the questionnaire data. Specifically, the survey data for the five attribute variables—“tool ease of use,”“interactivity,”“quality of the result,”“knowledge level,” and “self-efficacy”—and the two outcome variables—“knowledge acquisition” and “learning satisfaction”—were calibrated using the quartile method. The calibration points were set at 5% (lower quartile) as fully unaffiliated, 50% (median) as the crossover point, and 95% (upper quartile) as fully affiliated. This approach follows the standard practice outlined by Fiss (2011) and Greckhamer (2016), ensuring a balance between sensitivity and robustness in the fsQCA analysis. The calibration points and descriptive statistics for each variable are shown in Tables 4 and 5.

Table 4.

Descriptive Statistical Analysis Results of Research Variables.

Variable	Descriptive analysis
Variable	Mean	SD	Minimum	Maximum
TEOU	3.610	0.833	1.667	5.000
IN	3.650	0.756	2.000	5.000
AOR	3.967	0.730	2.000	5.000
KL	3.715	0.861	2.000	5.000
SE	3.667	0.806	2.000	5.000
KA	3.764	0.759	2.000	5.000
LS	3.911	0.624	2.333	5.000

Table 5.

Summary of Calibration.

Type of variable	Variable	Fuzzy calibration points
Type of variable	Variable	Fully affiliation	Intersection	Completely unaffiliated
Antecedent variable	TEOU	5.000	4.000	1.500
	IN	5.000	3.750	2.750
	AOR	5.000	4.000	3.000
	KL	5.000	4.000	2.500
	SE	5.000	4.000	2.333
Outcome variable	KA	5.000	3.667	2.333
Outcome variable	SE	5.000	4.000	3.000

Necessary Condition Test

Conditional necessity analysis is conducted to determine whether specific conditions must exist for the outcome to occur. The necessity analysis test is primarily based on the consistency index, with a consistency value greater than 0.9 being recognized as a necessary condition (Mattke et al., 2021). The necessity analysis was conducted using fsQCA 3.0, and the results are presented in Table 6. The consistency values for individual attribute conditions with respect to high or non-high learning experience were all below 0.9, indicating that none of the conditions constitute a necessary condition for determining the level of learners’ programming learning experience.

Table 6.

Necessary Conditions for Learning Experience.

Sets of conditions	Learning experience
	Knowledge acquisition effect				Learning satisfaction effect
	Consistency		Coverage		Consistency		Coverage
	High level of KA	Low level of KA	High level of KA	Low level of KA	High level of LE	Low level of LE	High level of LE	Low level of LE
TEOU	0.833	0.541	0.839	0.640	0.709	0.569	0.801	0.603
∼TEOU	0.642	0.864	0.543	0.859	0.649	0.813	0.616	0.723
IN	0.828	0.518	0.805	0.592	0.689	0.586	0.738	0.589
∼IN	0.580	0.830	0.506	0.850	0.617	0.740	0.614	0.690
AOR	0.799	0.646	0.679	0.646	0.806	0.574	0.770	0.514
∼AOR	0.583	0.679	0.583	0.799	0.491	0.743	0.552	0.782
KL	0.819	0.691	0.658	0.653	0.862	0.604	0.777	0.510
∼KL	0.567	0.638	0.610	0.806	0.457	0.736	0.551	0.833
SE	0.790	0.595	0.726	0.636	0.807	0.605	0.803	0.609
∼SE	0.604	0.744	0.562	0.805	0.607	0.805	0.603	0.808

Configurations

Configuration analysis, also known as the sufficiency analysis of antecedent condition configurations, can reveal the sufficiency of different attribute combinations in influencing learning experience outcomes. In fsQCA configuration analysis, the calibrated case data are first subjected to logical judgment using a truth table. Subsequently, the fsQCA software generates three types of solutions based on the logical remainders in the truth table: Complex Solution, Intermediate Solution, and Parsimonious Solution. Referring to the recommendations of scholars (Fiss, 2011; Zheng et al., 2021), the frequency threshold in the truth table was set to 1.0, the consistency threshold to 0.80, and the PRI (proportional reduction in inconsistency) threshold to 0.80. Samples that did not meet these conditions were marked as 0 (Zheng et al., 2021). The results in Tables 7 and 8 show that the consistency values of all configuration paths exceed the 0.8 consistency threshold, indicating that the results of the configuration analysis are reliable.

Table 7.

Configurations for Achieving High Levels of Programming Learning Experiences.

Antecedent condition	High level of learning experiences
	High level of KA				High level of LS
	H1a	H1b	H1c	H1d	H2a	H2b	H2c	H2d
TEOU	●		•	●	●	•	●
IN		●	●	⨂	●		●	●
AOR		•	•	⊗			•	•
KL	•	•		⊗	●	•		•
SE	⊗		•	•		⨂	•	⨂
Consistency	0.869	0.831	0.894	0.923	0.863	0.887	0.875	0.915
Raw coverage	0.435	0.663	0.628	0.332	0.381	0.663	0.652	0.329
Unique coverage	0.060	0.043	0.012	0.050	0.002	0.041	0.022	0.037
Overall solution consistency	0.827				0.871
Overall solution coverage	0.826				0.836

Note. ● indicates the presence of a core condition; • indicates the presence of a peripheral condition; ⊗ indicates the absence of a core condition; ⨂ indicates the absence of a peripheral condition.

Table 8.

Configurations for Achieving Low Levels of Programming Learning Experiences.

Antecedent condition	Low level of learning experiences
	Low level of KA			Low level of LS
	N1a	N1b	N1c	N2a	N2b	N2c
TEOU		⊗	⊗		⊗	⨂
IN	⊗	⊗		⊗	⨂	⊗
AOR	•	⊗	•	⊗	•	⊗
KL	⨂	⨂	•			⨂
SE		⊗		•		⊗
Consistency	0.934	0.908	0.953	0.906	0.935	0.965
Raw coverage	0.521	0.370	0.543	0.428	0.423	0.570
Unique coverage	0.023	0.004	0.174	0.062	0.079	0.089
Overall solution consistency	0.867			0.862
Overall solution coverage	0.755			0.833

Core conditions exhibit strong causal contributions to the outcome, as they appear consistently in both parsimonious (simplified) and intermediate solutions in fsQCA analysis. In contrast, peripheral conditions play auxiliary roles, appearing only in the intermediate solution and indicating weaker or context-dependent influences.

From the means—end chain perspective, AI-assisted programming enhances learners’ experiences by linking the use of AI tools to both functional and psychological consequences. Functionally, AI-assisted programming primarily enhances knowledge acquisition, whereas psychologically, it enhances learning satisfaction. Therefore, this study focuses on these two dimensions and, based on Table 7, identifies the two pathways that generate high programming learning experiences through high knowledge acquisition and high learning satisfaction.

Dual-driven-ease + knowledge (H1a and H2b): the original coverage of H1a is 0.435, which is a third path for improving knowledge acquisition. Among these paths, in driving programming learners to achieve high KA, TEOU of AI tools plays a core role as a core condition, while the KL serves as a supplementary role as a peripheral condition. The original coverage of H2b is 0.663, which is the maximum of the four configuration paths, indicating that this configuration is a significant path for the formation of high LS. For achieving high LS, both TEOU and KL play central roles, acting as core conditions. Meanwhile, in both H1a and H2b, SE is absent. This suggests that even for learners with low SE, a combination of high TEOU and KL can jointly drive the generation of both high KA and high LS. This configuration achieves a relatively high learning experience when using AI tools to assist in programming learning. Thus, this configuration represents a dual-driven type of configuration, driven by both functional consequences and psychological consequences.

Dual-driven-interactivity + accuracy + knowledge (H1b and H2d): the original coverage of H1b is 0.663, which is the maximum of the four configuration paths, indicating that this configuration serves as a primary pathway for the formation of high KA. The original coverage of H2d is 0.329, which is the minimum among the four configuration paths, and is a secondary path formed by high LS. In these two paths, it can also be observed that, both in terms of KA and LS, interactivity (IN) plays a central role, serving as a core condition driving high knowledge acquisition outcomes and high learning satisfaction outcomes. AOR and the level of relevant knowledge serve as auxiliary factors, acting as marginal conditions. Additionally, in H2d, SE is in a state of absence, which indicates that for programming learners with low SE, a combination of high IN, high AOR, and high relevant KL can still lead to a high programming learning experience outcome. Therefore, this configuration pathway can be considered a dual-driven type of configuration.

Dual-driven-ease + interactivity + accuracy + self-efficacy (H1c and H2c): in configuration H1c, interactivity of AI tools is a critical core condition, while TEOU, AOR, and SE serve as supplementary peripheral conditions. In contrast, for achieving high LS, both TEOU and IN play a pivotal role as core conditions, while AOR and KL act as peripheral conditions. In terms of coverage value, the coverage of H1c is 0.628, which is second in the configuration for generating high knowledge acquisition. Similarly, the coverage of H2c is 0.652, which is second in the configuration for generating high LS. Therefore, this pathway is a significant route to enhancing programming learning experiences. This suggests that when AI tools are highly interactive, easy to operate, and provide relatively accurate feedback, programming learners can achieve both high knowledge acquisition (KA) and high learning satisfaction (LS). Consequently, this configuration pathway also represents a dual-driven model, propelled by both functional and psychological outcomes.

Functional-only-ease + self-efficacy (H1d): the original coverage of this configuration is 0.381, which is the minimum among the five configuration paths. This suggests that this configuration serves as a secondary pathway for guiding learners to achieve high knowledge acquisition. In this path, TEOU and SE both play core roles. This indicates that when programming, learners possess a high level of SE, and AI tools demonstrate high usability. Even if learners’ knowledge level is relatively low and the interactivity of AI tools is limited, they can still drive the generation of high KA. However, as shown in Table 8, N2a reveals that the absence of IN and the absence of AOR lead to the generation of low LS. That is, although the configuration in H1d can lead to high KA as a functional consequence, it simultaneously results in low LS as a psychological consequence for learners. Therefore, this configuration cannot drive programming learners to achieve a high learning experience.

Psychological-dominant-ease + interactivity + knowledge (H2a): the original coverage of this configuration is 0.381, which is a third path for improving learning satisfaction. This path indicates that high TEOU and IN, as well as KL, can drive the generation of high LS. At the same time, by comparing the configuration pathways for generating low KA results in Table 8, it can be observed that this configuration does not drive the generation of low KA. That is, TEOU and IN and KL generate high LS without leading to low KA. Therefore, this configuration can enhance the learning experience of AI-assisted programming learning. At the same time, in this configuration, the high learning experience is primarily driven by high learning satisfaction. Thus, this configuration represents a type that is psychologically consequence-dominant.

In summary, the enhancement model of learning experiences with generative AI tools for programming can be classified into two types: the dual-driven type, characterized by functional consequences and psychological consequences, and the psychological consequence-dominant type. The dual-driven pattern, encompassing both functional and psychological consequences, consists of three primary configuration pathways: (1) Dual-driven-ease + knowledge (H1a, H2b); (2) dual-driven-interactivity + accuracy + knowledge (H1b, H2d); (3) dual-driven-ease + interactivity + accuracy + self-efficacy (H1c, H2c). The psychological consequence-dominant configuration is primarily represented by Psychological-dominant-ease + interactivity + knowledge (H2a). Notably, Functional-only-ease + self-efficacy (H1d) supports knowledge acquisition but not overall high experiences due to low satisfaction.

Robustness Test

Robustness testing is a critical aspect of fsQCA research. Common methods for testing robustness include raising the consistency threshold, adding or removing cases, and introducing additional conditions (Melendez-Torres et al., 2019). In this study, we apply the method of increasing the consistency threshold for cases, as proposed by Rihoux & Ragin (2009), to assess robustness. By expanding the case consistency threshold from 0.8 to 0.85, the robustness of all states in Tables 7 and 8 was examined, with case frequencies remaining unchanged. The results indicate that the configurations of the new model are consistent with those of the original model, and there is a clear subset relationship between the configurations of the two models (Schneider & Wagemann, 2012). Therefore, the results obtained in this study are robust.

Discussion

The research finds that the influencing mechanism of learning experiences with generative AI tools is a complex configurational process, involving multiple factor combinations under both the tool and learner dimensions (as shown in Table 7). Relying solely on AI tools or learners makes it challenging to enhance the programming learning experience, which aligns with the findings of Evans et al. (2024). The fsQCA results reveal two overarching patterns. One is a dual-driven pattern, where functional and psychological consequences work together to enhance both knowledge acquisition and learning satisfaction. The other is a psychological-dominant pattern, which boosts learning satisfaction without compromising knowledge acquisition. These findings confirm that effective AI-assisted programming learning requires a synergy between tool features and learner readiness, while also revealing multiple pathways that can achieve a high learning experience.

By integrating MEC with fsQCA, this study extends MEC beyond its typical linear “attribute–consequence–value” structure. It demonstrates that learners’ AI-assisted programming experiences are shaped by multiple parallel chains rather than a single dominant path. The identification of several sufficient but non-necessary configurations shows that MEC components interact configurationally (Wu & Wang, 2022). For example, the ease of use and interactivity of a tool may function as core attributes in some chains, while learner knowledge level or self-efficacy serve as critical links in others. Furthermore, the discovery of a psychological-dominant pathway indicates that psychological consequences can sometimes arise independently from strong functional outcomes, enriching MEC theory by highlighting asymmetry and non-linearity in how consequences form (Yan et al., 2024). This contributes to a more nuanced, context-sensitive understanding of MEC in educational technology environments, particularly when learners interact with adaptive, generative AI systems.

The findings of this study provide insights to guide educational practice. To ensure the sustainable development of AI-assisted programming instruction, educators should consider learners’ individual characteristics. Beyond the functional performance of AI tools, it is also essential to enhance the psychological value they provide. Therefore, for educators and developers, these results offer guidance on improving AI-assisted learning tools and services to meet the diverse needs of learners. Moreover, learners’ self-efficacy and knowledge levels strongly influence their learning experience (Ezeamuzie, 2023; Sweller et al., 2011). Educators should consider these characteristics when designing instructional activities and selecting AI tools to enhance their effectiveness. For instance, tools that emphasize interactivity and high-quality feedback are more suitable for learners with stronger prior knowledge. In contrast, tools with greater ease of use can help learners with moderate knowledge levels build confidence and improve their perceived self-efficacy. Personalized support and adaptive learning pathways can further help students overcome programming difficulties and enhance both learning outcomes and engagement (Demir, 2022).

Conclusion

This research employed both functional and psychological consequences as measures of learning experiences with generative AI tools, based on the means-end chain theoretical framework, and extended the configurational understanding of the factors influencing these experiences. This research collected data from 221 university students through a questionnaire. Then, this study employed fsQCA to analyze the configuration of factors that generate high and low levels of knowledge acquisition and learning satisfaction.

The following key conclusions were drawn: (1) A single factor is insufficient to constitute a necessary condition for a high-level programming learning experience. The ease of use, interactivity, feedback accuracy, relevant knowledge level, and self-efficacy, when acting independently, do not exhibit a decisive influence on a high-level learning experience. In addition, the research identified four sufficient configuration paths through fuzzy-set qualitative comparative analysis that can drive learning. (2) The enhancement patterns for AI-assisted programming learning experiences are mainly categorized into two types: dual-driven by functional consequences and psychological consequences, and psychologically dominant consequence patterns. The dual-driven pattern of functional and psychological consequences is primarily composed of three configuration paths: (1) high tool ease of use and knowledge level; (2) high interactivity, feedback accuracy and knowledge level; (3) high tool ease of use, interactivity, feedback accuracy and self-efficacy. The psychological consequence-dominated pattern is primarily characterized by a configuration consisting of high tool ease of use, high interactivity, and a high relevant knowledge level.

Limitations and Future Research

Although this research systematically explores the impact of applying generative AI tools on undergraduates’ programming learning experiences by integrating the means-end chain model and the fuzzy-set qualitative comparative analysis method, it still has several limitations that warrant future investigation. First, the sample has limited representativeness. Although 221 valid questionnaires were collected, the data primarily came from learners in a specific region, which may limit the generalizability of the findings and fail to represent learners from different cultural and educational backgrounds fully. Second, although this research explores the synergistic effects among multiple influencing factors, it does not provide a detailed analysis of the dynamic change process of these factors. Future studies could expand the scope by including cross-cultural comparisons to explore how AI-assisted learning tools perform across diverse educational and cultural settings. Finally, this study is mainly based on data from learners within a single cultural and linguistic background. Future research could further explore the applicability of generative AI tools in cross-cultural and multilingual programming environments.

Footnotes

Appendix A: Research Items

ORCID iD

Huijuan Fu

Ethical Considerations

This study adhered to ethical research principles and the APA Ethical Principles of Psychologists and Code of Conduct (section 8.05). As the research involved an anonymous online survey of adults and posed minimal risk, formal institutional ethics approval was not required. Before participating, respondents were informed of the study’ s purpose, the voluntary nature of their involvement, and the confidentiality of their responses. Informed consent was obtained electronically, and only those who agreed were able to proceed with the survey.

Author Contributions

Rui Wang: Methodology, Formal analysis, Writing—Original draft.

Jie Chen: Data curation, Data analysis, Writing—Original draft.

Jiahe Zhao: Validation, Resources, Writing—Review & editing.

Huijuan Fu: Conceptualization, Validation, Writing—Review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by the National Natural Science Foundation of China [grant number 72164015], the Humanities and Social Science Research Fund Project of Jiangxi Provincial Department of Education [grant number JY24110], and the Research Project on Undergraduate Education and Teaching Reform in Jiangxi Province [Grant No. BKJG-2026-7-38].

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Data sharing does not apply to this article, as no datasets were generated or analyzed during the current study.

References

Al-Dokhny

Alismaiel

Youssif

Nasr

Drwish

Samir

(2024). Can multimodal large language models enhance performance benefits among higher education students? An investigation based on the task–technology fit theory and the artificial intelligence device use acceptance model. Sustainability, 16(23), 10780. https://doi.org/10.3390/su162310780

Al-Emran

Abu-Hijleh

Alsewari

A. A.

(2024). Examining the impact of Generative AI on social sustainability by integrating the information system success model and technology-environmental, economic, and social sustainability theory. Education and Information Technologies, 30(7), 9405–9426. https://doi.org/10.1007/s10639-024-13201-0

Ali

S. I.

Shaban

(2025). The AI-personality nexus: How nursing students’ traits shape use, learning and adaptation in artificial intelligence-driven education in Saudi Arabia. Nurse Education in Practice, 87, 104498. https://doi.org/10.1016/j.nepr.2025.104498.

Amoozadeh

Daniels

Nam

Kumar

Chen

Hilton

Ragavan

S. S.

Alipour

M. A.

(2023). Trust in generative AI among students: An exploratory study. arXiv, 2023, arXiv.2310.04631. https://doi.org/10.48550/arXiv.2310.04631

Archana

S. N.

Renjith

V. R.

Padmakumar

P. K., C., S.

Aboobaker

(2025). AI assisted learning and research: an exploratory study among university students and scholars. Discover Education, 4(1), 1–20. https://doi.org/10.1007/s44217-025-00814-x

Bandura

(1997). Self-efficacy: The exercise of control. W. H. Freeman.

Becker

B. A.

Denny

Finnie-ansley

Luxton-reilly

Prather

Santos

E. A.

(2022). Programming is hard—Or at least it used to be: educational opportunities and challenges of AI code generation. arXiv, 7, arXiv.2212.01020. https://doi.org/10.48550/arXiv.2212.01020

Bull

Kharrufa

(2024). Generative artificial intelligence assistants in software development education: A vision for integrating generative artificial intelligence into educational practice, not instinctively defending against it. IEEE Software, 41(2), 52–59. https://doi.org/10.1109/ms.2023.3300574

Chenxi

Haijie

(2023). Configuration of network embedding and financing capability: research on the growth path of small and medium-sized sports enterprises under the heterogeneity of the business environment. SAGE Open, 13(3), 21582440231187876. https://doi.org/10.1177/21582440231187876

10.

Chin

W. W.

(1998). The partial least squares approach to structural equation modeling. In Marcoulides

G. A.

(Ed.), Modern methods for business research (pp. 295–336): Lawrence Erlbaum Associates .

11.

Chyung

S. Y.

Roberts

Swanson

Hankinson

(2017). Evidence-based survey design: The use of a midpoint on the Likert scale. Performance Improvement, 56(10), 15–23. https://doi.org/10.1002/pfi.21727

12.

Davis

F. D.

(1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. Management Information Systems Quarterly, 319–340. https://doi.org/10.2307/249008

13.

Demir

(2022). The effect of different usage of the educational programming language in programming education on the programming anxiety and achievement. Education and Information Technologies, 27(3), 4171–4194. https://doi.org/10.1007/s10639-021-10750-6

14.

Evans

Vansteenkiste

Parker

Kingsford-Smith

Zhou

(2024). Cognitive load theory and its relationships with motivation: a self-determination theory perspective. Educational Psychology Review, 36(1), 7. https://doi.org/10.1007/s10648-023-09841-2

15.

Ezeamuzie

N. O.

(2023). Project-first approach to programming in K–12: Tracking the development of novice programmers in technology-deprived environments. Education and Information Technologies, 28(1), 1–31. https://doi.org/10.1007/s10639-022-11180-8

16.

Fan

Liu

Zhang

Pan

(2025). The impact of AI-assisted pair programming on student motivation, programming anxiety, collaborative learning, and programming performance: A comparative study with traditional pair programming and individual approaches. International Journal of STEM Education, 12(1), 1–17. https://doi.org/10.1186/s40594-025-00537-3

17.

Fan

Tian

(2024). Satisfaction with online Chinese learning among international students in China: A study based on the fsQCA method. Sustainability, 16(3), 1086. https://doi.org/10.3390/su16031086

18.

Fiss

P. C.

(2011). Building better causal theories: A fuzzy set approach to typologies in organization research. Academy of Management Journal, 54(2), 393–420. https://doi.org/10.5465/AMJ.2011.60263120

19.

Fornell

Larcker

D. F.

(1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), 39–50. https://doi.org/10.1177/002224378101800104

20.

H. J.

Xiao

Y. C.

Mensah

I. K.

Wang

(2023). Exploring the configurations of learner satisfaction with MOOCs designed for computer science courses based on integrated LDA-QCA method. Education and Information Technologies, 29(8), 9883–9905. https://doi.org/10.1007/s10639-023-12185-7

21.

Greckhamer

(2016). CEO compensation in relation to worker compensation across countries: The configurational impact of country-level institutions. Strategic Management Journal, 37(4), 793–815. https://doi.org/10.1002/smj.2370

22.

Greckhamer

Furnari

Fiss

P. C.

Aguilera

R. V.

(2018). Studying configurations with qualitative comparative analysis: Best practices in strategy and organization research. Strategic Organization, 16(4), 482–495. https://doi.org/10.1177/1476127018786487

23.

Groothuijsen

van den Beemt

Remmers

J. C.

van Meeuwen

L. W.

(2024). AI chatbots in programming education: Students’ use in a scientific computing course and consequences for learning. Computers and Education: Artificial Intelligence, 7, 100290. https://doi.org/10.1016/j.caeai.2024.100290

24.

Gulwani

Polozov

Singh

(2017). Program synthesis. Foundations and Trends^®in Programming Languages, 4(1–2), 1–119. https://doi.org/10.1561/2500000010

25.

Gutman

(1982). A means-end chain model based on consumer categorization processes. Journal of Marketing, 46(2), 60–72. https://doi.org/10.1177/002224298204600207

26.

Hartley

Hayak

U. H.

(2024). Artificial intelligence supporting independent student learning: An evaluative case study of ChatGPT and learning to code. Education Sciences, 14(2), 120. https://doi.org/10.3390/educsci14020120

27.

Hou

Ying, Wu

Wang

Ericson

B. J.

(2024). Codetailor: Llm-powered personalized parsons puzzles for engaging support while learning programming [Paper presentation] Proceedings of the Eleventh ACM Conference on Learning@ Scale, Atlanta, GA, USA, Georgia Tech.

28.

Hou

X. Y.

Ericson

B. J.

Wang

(2023). Understanding the effects of using Parsons problems to scaffold code writing for students with varying CS self-efficacy levels. arXiv, 2023, arXiv.2311.18115. https://doi.org/10.48550/arXiv.2311.18115

29.

Ithriah

Ridwandono

Suryanto

(2020). Online learning self-efficacy: The role in e-learning success. Journal of Physics: Conference Series, 1569, 022053. https://doi.org/10.1088/1742-6596/1569/2/022053

30.

Kazemitabaar

Chow

C. K. T.

Ericson

B. J.

Weintrop

Grossman

(2023). Studying the effect of AI code generators on supporting novice learners in introductory programming [Paper presentation]. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, Hamburg, Germany.

31.

Kumar

Rauschnabel

P. A.

Agarwal

M. N.

Singh

R. K.

Srivastava

(2024). Towards a theoretical framework for augmented reality marketing: A means-end chain perspective on retailing. Information & Management, 61(2), 103910. http://doi.org/10.1016/j.im.2023.103910

32.

Kwak

Jenkins

Kim

(2023). Adaptive programming language learning system based on generative AI. Issues in Information Systems, 24(3), 222–231. https://doi.org/10.48009/3_iis_2023_119

33.

Liu

Zhang

L. J.

Biebricher

(2024). Investigating students’ cognitive processes in generative AI-assisted digital multimodal composing and traditional writing. Computers & Education, 211(C), 104977. https://doi.org/10.1016/j.compedu.2023.104977.

34.

Mattke

Maier

Weitzel

Thatcher

J. B.

(2021). Qualitative comparative analysis in the information systems discipline: A literature review and methodological recommendations. Internet Research, 31(5), 1493–1517. https://doi.org/10.1108/intr-09-2020-0529

35.

Means

(1994). Technology and education reform: The reality behind the promise: Jossey-Bass.

36.

Melendez-Torres

Sutcliffe

Burchett

H. E.

Rees

Thomas

(2019). Developing and testing intervention theory by incorporating a views synthesis into a qualitative comparative analysis of intervention effectiveness. Research Synthesis Methods, 10(3), 389–397. https://doi.org/10.1002/jrsm.1341

37.

Miller

A. L.

(2023). AI Assisted learning: A tool or a threat?. Journal of e-Learning Research, 2(2), 52–65. https://doi.org/10.33422/jelr.v2i2.510.

38.

Misangyi

V. F.

Greckhamer

Furnari

Fiss

P. C.

Crilly

Aguilera

(2017). Embracing causal complexity: The emergence of a neo-configurational perspective. Journal of Management, 43(1), 255–282. https://doi.org/10.1177/0149206316679252

39.

Nouri

Zhang

Mannila

Norén

(2020). Development of computational thinking, digital competence and 21st century skills when learning programming in K-9. Education Inquiry, 11, 1–17. https://doi.org/10.1080/20004508.2019.1627844

40.

Rey-Martí

Felício

J. A.

Rodrigues

(2017). Entrepreneurial attributes for success in the small hotel sector: a fuzzy-set QCA approach. Quality & Quantity, 51, 2085–2100. https://doi.org/10.1007/s11135-016-0466-3

41.

Rihoux

Ragin

C. C.

(2009). Configurational comparative methods: Qualitative comparative analysis (QCA) and related techniques (Vol. 51). SAGE Publications Inc.

42.

Ryan

French

Kennedy

(2021). Beyond the Iron Triangle: Improving the quality of teaching and learning at scale. Studies in Higher Education, 46(7), 1383–1394. https://doi.org/10.1080/03075079.2019

43.

Sajja

Sermet

Cikmaz

Cwiertny

Demir

(2024). Artificial intelligence-enabled intelligent assistant for personalized and adaptive learning in higher education. Information (Switzerland), 15(10), 596. https://doi.org/10.3390/info15100596

44.

Sánchez-Mena

Martí-Parreño

Miquel-Romero

M. J.

(2019). Higher education instructors’ intention to use educational video games: An fsQCA approach. Educational Technology Research & Development, 67(6), 1455–1478. https://doi.org/10.1007/s11423-019-09656-5

45.

Schneider

C. Q.

Rohlfing

(2016). Case studies nested in fuzzy-set QCA on sufficiency: Formalizing case selection and causal inference. Sociological Methods & Research, 45(3), 526–568. https://doi.org/10.1177/0049124114532446

46.

Schneider

C. Q.

Wagemann

(2012). Set-theoretic methods for the social sciences: A guide to qualitative comparative analysis: Cambridge University Press.

47.

Shahzad

M. F.

Zahid

(2024). Exploring the impact of generative AI-based technologies on learning performance through self-efficacy, fairness & ethics, creativity, and trust in higher education. Education and Information Technologies, 30(3), 3691–3716. https://doi.org/10.1007/s10639-024-12949-9

48.

Shen

Zhang

Zheng

Ruan

(2025). How do self-identity and motivational foundation influence Chinese prosocial behavior intention? Findings from NCA and fsQCA. SAGE Open, 15(1), 21582440251326350. https://doi.org/10.1177/21582440251326350

49.

Steyvers

Kumar

(2024). Three challenges for AI-assisted decision-making. Perspectives on Psychological Science, 19(5), 722–734. https://doi.org/10.1177/17456916231181102

50.

Sun

Boudouaia

Zhu

(2024). Would ChatGPT-facilitated programming mode impact college students’ programming behaviors, performances, and perceptions? An empirical study. International Journal of Educational Technology in Higher Education, 21(1), 1–22. https://doi.org/10.1186/s41239-024-00446-5

51.

Sweller

Ayres

Kalyuga

(2011). Measuring cognitive load. In Cognitive load theory (Explorations in the Learning Sciences, Instructional Systems and Performance Technologies, Vol. 1, pp. 71–85). Springer. https://doi.org/10.1007/978-1-4419-8126-4_6

52.

Tlili

Shehata

Adarkwah

M. A.

Bozkurt

Hickey

D. T.

Huang

Agyemang

(2023). What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education. Smart Learning Environments, 10(1), 15. https://doi.org/10.1186/s40561-023-00237-x

53.

Utreras

Pontelli

(2021). Introductory programming and young learners with visual disabilities: A review. Universal Access in the Information Society, 22(7), 169–184. https://doi.org/10.1007/s10209-021-00830-2.

54.

Vistorte

A. O. R.

Deroncele-Acosta

Ayala

J. L. M.

Barrasa

López-Granero

Martí-González

(2024). Integrating artificial intelligence to assess emotions in learning environments: a systematic literature review. Frontiers in Psychology, 15, 1387089. https://doi.org/10.3389/fpsyg.2024.1387089

55.

Wang

Xie

Ali

Brem

Wang

(2022a). How do individual characteristics and social capital shape users’ continuance intentions of smart wearable products? Technology in Society, 68(Suppl C), 101818. https://doi.org/10.1016/j.techsoc.2021.101818

56.

Wang

Cao

Gong

Wang

(2022b). Interaction and learning engagement in online learning: The mediating roles of online learning self-efficacy and academic emotions. Learning & Individual Differences, 94, 102128. https://doi.org/10.1016/j.lindif.2022.102128

57.

Wang

Díaz

D. V.

Brown

Chen

(2023). Exploring the role of AI assistants in computer science education: Methods, implications, and instructor perspectives. arXiv, 2023, arXiv.2306.03289. https://doi.org/10.48550/arXiv.2306.03289.

58.

Wang

(2022). Formation mechanism of popular courses on MOOC platforms: A configurational approach. Computers & Education, 191, 104629. https://doi.org/10.1016/j.compedu.2022.104629

59.

Yan

Nakajima

Sawada

(2024). Benefits and challenges of collaboration between students and conversational generative artificial intelligence in programming learning: An empirical case study. Education Sciences, 14(4), 433. https://doi.org/10.3390/educsci14040433

60.

Yin

Z.-Y.

Huang

A.-M.

Huang

Z.-Y.

(2024). Virtual tourism attributes in cultural heritage: Benefits and values. Tourism Management Perspectives, 53, 101275. http://doi.org/10.1016/j.tmp.2024

61.

Zawacki-Richter

Marín

V. I.

Bond

Gouverneur

(2019). Systematic review of research on artificial intelligence applications in higher education–where are the educators?. International Journal of Educational Technology in Higher Education, 16(1), 1–27. https://doi.org/10.1186/s41239-019-0171-0

62.

Zheng

Ulrich

Sendra-García

(2021). Qualitative comparative analysis: Configurational paths to innovation performance. Journal of Business Research, 128, 83–93. https://doi.org/10.1016/j.jbusres.2021.01.044

63.

Zhou

Chen Funk

D. C.

Kunkel

(2020). Solving the athleisure myth: A means-end chain analysis of female activewear consumption. Journal of Sport Management, 35(1), 81–93. https://doi.org/10.1123/jsm.2019-0358

64.

Zviel-Girshin

(2024). The good and bad of AI tools in novice programming education. Education Sciences, 14(10), 1089. https://doi.org/10.3390/educsci14101089