To Use but Not to Depend: Pedagogical Novelty and the Cognitive Brake of Ethical Awareness in Computer Science Students’ Adoption of Generative AI

Abstract

The integration of Generative Artificial Intelligence (GenAI) into higher education represents a paradigm shift from static skill acquisition to dynamic, human–AI collaboration. However, the psychological mechanisms governing students’ adoption—specifically the interplay between pedagogical novelty, ethical awareness, and habit formation—remain underexplored. To address this, this study develops and implements a dynamic practical curriculum incorporating AI and ethical awareness, aiming to foster responsible behavioral patterns in computer programming education. Employing a quasi-experimental design, we implemented a 16-week dual-track instructional intervention (incorporating AI-integrated pedagogy and ethical scaffolding) for 148 computer science students. Structural Equation Modeling (SEM) was applied to test an extended UTAUT2 framework. The findings reveal three critical theoretical insights that redefine GenAI adoption: (1) The eclipse of utility: contrary to established models, traditional utilitarian drivers of performance expectancy (β = 0.076, p = 0.39) and effort expectancy (β = 0.125, p = 0.13) yielded non-significant effects on behavioral intention. This suggests that for digital natives, algorithmic efficiency has devolved into a baseline hygiene factor, losing its motivational power. (2) The dominance of pedagogical novelty: hedonic motivation emerged as the paramount predictor of both habit (β = 0.457, p < 0.001) and behavioral intention (β = 0.336, p = 0.001). This confirms that adoption is driven by the situational interest and interactional novelty inherent in the human–AI partnership. (3) The cognitive brake mechanism: ethical awareness exhibited a divergent regulatory role. While it significantly legitimized conscious behavioral intention (β = 0.166, p = 0.011), it showed a non-significant, negative association with habit (β = −0.032, p = 0.653). This demonstrates that ethical reasoning functions as a cognitive brake (system 2) and actively disrupts the formation of mindless, automated dependency (system 1). These results provide empirical evidence for a dual regulation model of AI adoption and suggest that sustainable education requires leveraging pedagogical novelty to drive engagement while utilizing ethical awareness to prevent blind habituation.

1. Introduction

The swift progress of artificial intelligence (AI) technologies has strongly influenced diverse industries. Education, especially in computer science (CS), is no exception. The integration of Generative Artificial Intelligence (GenAI) into CS pedagogy has precipitated a fundamental transformation in CS pedagogy (Zawacki-Richter et al., 2019; Shaukat et al., 2020; Prather et al., 2023). Unlike conventional instructional resources, typified by static textbooks and pre-recorded lectures, GenAI tools act as dynamic agents and are capable of providing real-time code explanation, debugging assistance, and personalized feedback (Becker et al., 2023; Long et al., 2025). For CS students, who have previously grappled with the high cognitive load and abstract nature of programming syntax through passive learning materials, the emergence of AI-integrated curricula represents a significant shift in the educational medium (Kabudi et al., 2021; Barbosa et al., 2024; Sari et al., 2024). While the potential of these tools to enhance learning is evident, the psychological mechanisms driving their adoption remain complex and underexplored, particularly regarding how students balance the appeal of technological novelty with the constraints of academic ethics.

Despite the rapid proliferation of GenAI, existing research utilizing the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) has predominantly focused on utilitarian drivers, such as performance expectancy and effort expectancy (Dwivedi et al., 2019; Tamilmani et al., 2021). However, recent scholarship suggests that for digital natives, the utility of digital tools is progressively taken for granted, representing a baseline expectation rather than a motivator (Kocielnik et al., 2019; Dwivedi et al., 2023). Instead, the primary driver may be shifting toward hedonic motivation, sparked by the pedagogical novelty of the AI medium itself. Compared to the static nature of traditional textbooks, AI-integrated materials offer a form of situational interest through interactive novelty and potentially override traditional efficiency metrics as the central determinant of habit formation and adoption behavior (Huang et al., 2022; Petrescu et al., 2023; Artemova, 2024; Robertson & Georgeon, 2025).

Furthermore, the introduction of AI into the curriculum brings ethical considerations to the forefront. In the context of programming, ethical training is often framed solely as a barrier designed to deter plagiarism (McIntire et al., 2024; Prashar et al., 2024). Nevertheless, ethical awareness may play a dual, regulatory role within the adoption process. On one hand, it legitimizes the intention to use AI by clarifying permissible boundaries. On the other hand, it may act as a cognitive brake that requires deliberative processing to inhibit the formation of mindless, automated habits (Oravec, 2023; Cotton et al., 2024; M. Li et al., 2024). Current literature has yet to empirically disentangle how ethical awareness differentially impacts conscious versus automated behavioral pathways. To address these theoretical gaps, this study adapts the UTAUT2 framework to investigate the adoption of AI-integrated programming materials with ethical awareness. Specifically, this research aims to answer the following questions:

• RQ1: Does the pedagogical novelty of GenAI-integrated resources supersede traditional utilitarian performance expectancy and effort expectancy in shaping CS students’ adoption behaviors?
• RQ2: How does ethical awareness differentially regulate the conscious intention to use AI versus the formation of automated habits, acting as a cognitive mechanism to prevent mindless dependency?
• RQ3: Does the intrinsic enjoyment derived from AI interaction act as the primary catalyst for habit formation, mediating the relationship between pedagogical novelty and sustained usage intention?

By answering these questions, this study provides critical insights for educators seeking to leverage the engaging power of AI while cultivating responsible and reflective coding practices.

2. Literature Review

2.1. Generative AI as a Pedagogical Medium Beyond Static Resources

The landscape of CS education is undergoing a fundamental transition from static instructional resources to dynamic, generative agents. Traditionally, novices have relied on textbooks and online forums, resources that impose a high extrinsic cognitive load by requiring learners to independently bridge the gap between abstract syntax and code execution (Becker et al., 2023; Qureshi, 2023). In contrast, GenAI tools function as conversational partners, offering real-time scaffolding, code synthesis, and personalized explanations. This shift represents a change in the pedagogical medium itself (X. Chen et al., 2021; Chiu, 2021; Chiu et al., 2024; Dzogovic et al., 2024). Unlike passive tools, GenAI agents introduce a layer of interactional novelty, transforming the solitary act of coding into a two-way dialog (Cano & Nunez, 2024; Grange et al., 2026). Understanding this shift is essential for analyzing students’ adoption drivers.

2.2. Theoretical Lens: UTAUT2 and the Hygiene Factor Phenomenon

UTAUT2 serves as the prevailing framework for understanding consumer technology adoption, theorizing that performance expectancy (efficiency) and effort expectancy (ease of use) are central determinants of behavior (Venkatesh et al., 2012). However, the applicability of these utilitarian constructs is increasingly debated in the context of digital natives (Tamilmani et al., 2021; Chang & Chang, 2023). For current university students, high adoption rates of digital tools have rendered efficiency and usability as baseline expectations or hygiene factors rather than distinguishing motivators. When a technology’s high performance is taken for granted, utilitarian drivers may reach a ceiling effect and diminish their predictive power relative to experiential or hedonic factors. This theoretical nuance is critical when examining GenAI, as it is a technology where efficiency is an inherent and assumed attribute rather than a variable feature.

2.3. Reframing Hedonic Motivation: Situational Interest and Pedagogical Novelty

Within the UTAUT2 framework, hedonic motivation is traditionally defined as the enjoyment derived from using a technology. While often conflated with gamification mechanics, recent educational psychology research suggests a broader interpretation rooted in situational interest. The novelty effect of interacting with a generative agent, one that can understand natural language and produce creative output, can cause a state of intrinsic cognitive engagement (Min & Schwarz, 2022; Asher & Harackiewicz, 2025; Guo & Fryer, 2025). In this context, hedonic motivation reflects the intellectual curiosity and satisfaction derived from the pedagogical novelty of the AI medium. Unlike the extrinsic loops of gamification, this form of enjoyment arises from the fluidity of the human–AI interaction itself, potentially acting as a dominant psychological driver in learning environments where traditional methods are perceived as tedious or static.

2.4. The Dual Nature of Ethical Awareness: Legitimacy vs. Automaticity

The integration of GenAI introduces complex ethical dimensions regarding academic integrity and authorship. Existing literature predominantly frames ethical awareness as an inhibitor, positing that greater ethical concerns reduce the likelihood of adoption (Al-Mughairi & Bhaskar, 2024). However, a cognitive regulation perspective may suggest a dual role. First, ethical awareness may provide cognitive legitimacy. When students possess a clear understanding of ethical boundaries, the ambiguity and perceived risk of using AI decrease, potentially empowering them to adopt the tool responsibly (Mikalef et al., 2022; Lu et al., 2024). Second, and conversely, ethical awareness involves deliberative cognitive processing, which requires conscious reflection and judgment. This stands in theoretical contrast to habit, which is characterized by automaticity and a lack of conscious thought. Consequently, a high level of ethical awareness may compel students to engage in reflective decision-making for each use case, theoretically counteracting the formation of mindless, automated usage loops. This tension between conscious ethical regulation and automated habit formation remains an underexplored area in technology acceptance research.

2.5. Systematic Instructional Design (SID) and Cognitive Load Theory (CLT)

The integration of GenAI into higher computer programming education necessitates a strategic realignment of Systematic Instructional Design (SID) with Cognitive Load Theory (CLT). Traditional programming instruction imposes a high intrinsic cognitive load because learners must simultaneously master complex syntax and abstract logic (Sweller, 2011). GenAI tools effectively reduce this intrinsic load by automating code generation. However, recent empirical evidence indicates that this efficiency frequently leads to a deficit in germane cognitive load, where learners engage in cognitive offloading rather than constructing necessary mental schemas. In addition, without specific pedagogical intervention, unguided AI usage negatively correlates with critical thinking development (Tian & Zhang, 2025). To mitigate these risks, SID frameworks must transition from content generation to competency validation. Instructional strategies should progress beyond substitution, where AI replaces learner effort, toward redefinition, where AI outputs serve as objects of analysis (Romero, 2025). By mandating that students verify and rectify AI-generated solutions, educators ensure that the cognitive capacity liberated by automation is redirected toward higher-order problem decomposition and logic verification (Manorat et al., 2025). This structural approach ensures that AI integration augments the cognitive processing essential for deep computational mastery.

In summary, the unique characteristics of GenAI and its pedagogical novelty, the baseline expectations of digital natives, and the complex cognitive demands of ethical reasoning, all together challenge the traditional assumptions of the UTAUT2 model. To empirically investigate these dynamics, we propose an extended research model that integrates ethical awareness and re-evaluates the roles of utilitarian versus hedonic drivers. The specific relationships and hypotheses are detailed in the following section.

3. Research Design

To empirically investigate the proposed research model, this study employed a quasi-experimental design aimed at examining the determinants of GenAI adoption among CS students in Python (version 3.80) programming education. This study systematically investigated these factors based on the UTAUT2 model, which was proposed in 2012 as an extended version of the UTAUT model (Venkatesh et al., 2012). This model provides a comprehensive and in-depth analytical framework for technology acceptance research (Dwivedi et al., 2020; Tamilmani et al., 2021).

3.1. Research Model and Hypotheses Development

Based on the literature review and the contextual shift toward pedagogical novelty and ethical cognitive regulation, we propose the research model depicted in Figure 1. The model posits that while traditional utilitarian drivers remain present, the hedonic motivation derived from AI interaction and the ethical awareness cultivated through training are the decisive predictors of habit and behavioral intention. Consistent with the research model, gender, major, age, and prior experience were included as control variables to account for potential confounding effects on habit and intention.

3.1.1. The Dominance of Hedonic Motivation (The Novelty Effect)

In the context of GenAI, hedonic motivation can be reconceptualized as the situational interest triggered by the pedagogical novelty of the AI medium (Huang et al., 2022; Petrescu et al., 2023; Artemova, 2024; Robertson & Georgeon, 2025). We conceive that the immediate feedback and conversational nature of AI tools reduce the cognitive work of coding, thereby reinforcing usage habits through positive reinforcement. Acknowledging the digital native context, we retain traditional UTAUT2 paths, but anticipate their influence may be overshadowed by hedonic motivation due to the ceiling effect of modern technology (Tamilmani et al., 2021).

H1. 

Hedonic motivation positively influences habit.

H2. 

Hedonic motivation positively influences behavioral intention.

3.1.2. The Cognitive Brake of Ethical Awareness

Ethical awareness is not a standard component of the original UTAUT 2 model. However, researchers have explored the integration of ethical considerations into technology acceptance models, including UTAUT2. While not explicitly part of UTAUT2, ethical awareness can be considered as an extension or additional construct that influences user behavior and intentions when it comes to technology adoption (Venkatesh et al., 2012). We hypothesize a divergent path for ethical awareness. According to dual process theory, intention is a conscious commitment (system 2), while habit implies automaticity (system 1). Ethical awareness serves as a cognitive brake, legitimizing the intention to use the tool by clarifying boundaries, while simultaneously requiring deliberative processing that constrains the formation of mindless habits. The following hypotheses are formulated from the perspective of ethical awareness.

H3. 

Ethical awareness positively influences behavioral intention.

H4. 

Ethical awareness will have a negative or non-significant influence on habit. (the cognitive brake hypothesis).

3.1.3. The Hygiene Factors and Rational Constraints

Behavioral intention of the UTAUT2 model is regarded as a vigorous construct and integrates diverse factors shaping technology adoption. It is formed by performance expectancy, effort expectancy, and social influence, and moderated by habit and hedonic motivation. Behavioral intention is a conservative pivotal factor of user behavior. The stronger the intention to use technology, the more likely it is to be used. It summarizes the motivational factors that impact the possibility of using technology (Kim et al., 2007). Additionally, price value represents the rational economic constraint. The following hypotheses are outlined from the perspective of behavioral intention.

H5. 

Performance expectancy positively influences behavioral intention.

H6. 

Effort expectancy positively influences behavioral intention.

H7. 

Social influence positively influences behavioral intention.

H8. 

Price value positively influences behavioral intention.

3.1.4. The Habituation Pathway

Habit is an important construct in the UTAUT2 model (Dwivedi et al., 2020). It reveals unconscious behavior triggered by environmental prompts rather than thoughtful decision-making (Ambalov, 2021). Habit has a robust predictive power over technology use and acceptance. Once a habit is formed, it can significantly influence future technology use, sometimes even beyond the influence of intention (C.-F. Chen & Chao, 2011). Allied with the model in Figure 1 following hypotheses are framed from the perspective of habit.

H9. 

Habit positively influences behavioral intention.

3.2. Instructional Intervention Design

To ensure the ecological validity of the study and to empirically test the pedagogical novelty and cognitive brake hypotheses, this research implemented a comprehensive dual-track instructional framework. As illustrated in Figure 2, the diagram illustrates the parallel progression of the AI-integrated pedagogy track (track A) and the ethical scaffolding track (track B) over the 16-week semester, highlighting key integration points. The intervention moves beyond simple tool adoption, structuring the semester-long course (16 weeks) into two parallel but interconnected tracks. The AI-integrated pedagogy track aimed at stimulating hedonic motivation via novelty, and the ethical scaffolding track aimed at regulating habit via cognitive friction.

3.2.1. Track A: The AI-Integrated Curriculum (Pedagogical Novelty)

The curriculum design shifts the pedagogical focus from syntax memorization to logic architecture, leveraging the generative capabilities of Large Language Models (LLMs) to reduce extrinsic cognitive load and foster situational interest. The AI agent applied in this study is either DeepSeek-V3 or DOLA (v9.6), as decided by the participants. The curriculum is divided into four progressive phases (see Figure 3 for the detailed workflow).

The flowchart demonstrates the student’s cognitive progression from the initial novelty trigger (Phase 1) to logic-based prompting (Phase 2), critical auditing (Phase 3), and finally to collaborative co-creation (Phase 4), where AI serves as a partner in creative production. In Phase 1 of the comparative awakening (weeks 1–3), to trigger immediate pedagogical novelty, students engage in a manual vs. AI contrast experiment. Students first randomly pick from the Python algorithmic problem list (as shown in Appendix A) and solve a classic algorithmic problem manually, documenting time spent and frustration levels. Subsequently, they utilize an AI agent (either DeepSeek or Dola) to generate the same solution. This stark contrast in efficiency and the conversational interface of the AI serve as a novelty trigger and activate the hedonic motivation pathway (Huang et al., 2022). The second phase involves prompt engineering and logic decomposition (weeks 4–8). This phase transitions students from code writers to prompt engineers. The instructional material focuses on decomposition-based prompting, where students need to break down complex problems into logical pseudocode before interacting with the AI. This ensures that the AI is used as a cognitive scaffold rather than a solution dispenser, maintaining high cognitive engagement (Chiu, 2021). In the following phase 3, about the AI auditor and debugging (weeks 9–13), to counter the hygiene factor effect (where efficiency becomes taken for granted), this phase introduces human auditing. Instructors provide pre-generated AI code containing subtle logical errors or security vulnerabilities. Students assume the role of code auditors, required to identify, explain, and fix these AI-induced errors. This activity reinforces technical competency while demystifying the infallibility of the tool (Moorhouse et al., 2023). In phase 4 regarding collaborative co-creation (weeks 14–16), the capstone project requires students to develop a functional application using AI for up to 80% of the codebase. The assessment focuses on the students’ ability to integrate modules and refine the user experience, symbolizing the shift to high-level creative problem-solving.

3.2.2. Ethical Awareness Training Design (Track B: The Ethical Scaffolding (The Cognitive Brake))

To test the cognitive brake hypothesis (H4), a dedicated ethical awareness training module was embedded. Parallel to the technical track, the ethical intervention is designed to introduce cognitive friction and force students to switch from system 1 (automatic habit) to system 2 (deliberative reasoning), thus preventing mindless dependency (Kahneman, 2011; Cotton et al., 2024). The specific mechanisms are detailed in Figure 4. The flowchart details how the intervention disrupts system 1 (automatic copy-pasting) and enforces system 2 (deliberative verification), thereby regulating habit formation.

In the first mechanism of the red team sensitization (weeks 1–3), students are tasked with red teaming exercises (as shown in Appendix B), where they intentionally probe the AI for hallucinations, biases, or incorrect code. In the meantime, the teaching of ethical principles (as shown in Appendix C) has been formally embedded to address AI’s ethical encounters (Ayinla et al., 2024; Olorunfemi et al., 2024; Zou et al., 2025). By exposing the flaws of the black box, this activity cultivates moral sensitivity and skepticism, laying the groundwork for critical usage (Dwivedi et al., 2023). As for Mechanism 2, about the pause and validation protocol (weeks 4–13), a mandatory AI disclosure checklist (as shown in Appendix D) is embedded into every assignment submission process. Students must categorize their code as AI-generated, human-modified, or human-written and provide a brief justification for using AI. This procedural requirement acts as a cognitive brake, disrupting the automaticity of copy-pasting and forcing a conscious ethical reflection before submission (Tlili et al., 2023, 2025). In mechanism 3 of the co-constructed code of conduct (weeks 14–16), instead of imposing top-down rules, students participate in drafting the classroom AI regulation. This participatory approach enhances the cognitive legitimacy of the ethical guidelines, ensuring that students internalize the norms rather than viewing them as external constraints.

3.3. Data Collection Procedure

The data collection strategy was designed to ensure ecological validity while adhering to ethical standards. As illustrated in Figure 5, the procedure was executed in four sequential phases over the course of a 16-week semester. In Phase 1, prior to the semester, the target population comprised CS students enrolled in the programming course. Recruitment was conducted during the first lecture. Students were explicitly informed that their survey responses would remain anonymous and would not influence their academic grades, a procedural remedy designed to reduce evaluation apprehension and social desirability bias (Podsakoff et al., 2003). In Phase 2 of intervention and observation (weeks 1–15), participants engaged in the 16-week dual-track instructional framework (as illustrated in Figure 2). During this period, check-in data were collected non-intrusively to verify engagement levels. This served as a manipulation check to ensure that the ethical scaffolding was actively experienced, not just theoretically presented.

In the following phase of survey administration (Week 16), post-intervention data were collected. The questionnaires (as shown in Appendix E) were distributed immediately after the submission of the project to capture the students’ cumulative experience. To ensure data quality, four trap questions of reversing the expressions were randomly inserted. The order of question blocks was randomized to mitigate order effects (Hair et al., 2019). Finally, Phase 4 of data screening and quality control could begin.

Table 1. Overall phases of the research with theoretical construct and mechanism.

Phase	Instructional Activity/Material	Theoretical Construct and Mechanism	Key References
Week 1–3	Manual vs. AI challenge	Pedagogical novelty	(Huang et al., 2022)
	Students compare manual coding time vs. AI generation	Triggering Situational Interest via contrast; Activating Hedonic Motivation.	(C.-H. Chen & Chang, 2024)
	Red team exercise	Moral sensitivity	(Dwivedi et al., 2023)
	Probing AI for hallucinations and bias	Exposing risks to build skepticism.	(Tlili et al., 2023)
Week 4–8	Decomposition-Based Prompting	Cognitive scaffolding	(Becker et al., 2023)
	Writing logic before code	Reducing extrinsic load while maintaining germane load.	(Prather et al., 2023)
Week 4–13	AI Disclosure Checklist	Cognitive brake (System 2)	(Kahneman, 2011)
	Mandatory form for every submission	Disrupting automaticity (Habit) to enforce deliberation.	(Cotton et al., 2024)
Week 9–13	The AI Auditor Task	Human-in-the-Loop	(Moorhouse et al., 2023)
	Fixing buggy AI code	Countering the hygiene factor perception; Building competence.	(Mosqueira-Rey et al., 2023)
Week 14–16	Capstone Co-Creation	Creative empowerment	(Chiu, 2021)
	80% AI code allowed	Reinforcing hedonic motivation through creative output.	(Fui-Hoon Nah et al., 2023)
	Code of Conduct Co-Design	Cognitive legitimacy	(Dwivedi et al., 2023)
	Class voting on rules	Internalizing norms to support behavioral intention.

presents the overall phases of the research. A total of 148 responses were initially received. The dataset underwent a rigorous screening process. Five responses with >10% missing data were removed. Three responses exhibiting unvarying patterns. Four responses failing the trap questions were excluded. This resulted in a final valid sample of n = 136. A post hoc power analysis using G*Power (version 3.1.9.7) indicated that this sample size provides sufficient power (>0.80) to detect medium effect sizes (f² = 0.15) for the proposed SEM model (Cohen, 2013).

4. Results

4.1. Characteristics of the Samples

As shown in Figure 6, the sample includes 136 participants from a university in Guangdong, China. All of them are computer science majors. This group is primarily male (118 participants, 86.76%) with 18 females (13.24%), which is common due to the characteristics of the CS major and has meaningful reference for related study (Shrestha & Das, 2022). Ages range from 19 to 24, with most being 20 years old (38.97%).

4.2. Reliability and Validity of Constructs

The data collected is analyzed by Mplus editor 8 (Version 1.6 (1)) and IBM SPSS R27.0. The model displays good reliability and validity for most of the constructs. As shown in

Table 2. Construct validity and reliability.

Construct	Item	Factor Loading	Cronbach’s Alpha	AVE	Composite Reliability
Performance expectancy (PE)	PE1	0.777	0.861	0.514	0.863
	PE2	0.752
	PE3	0.743
	PE4	0.681
	PE5	0.699
	PE6	0.639
Effort expectancy (EE)	EE1	0.77	0.872	0.541	0.875
	EE2	0.767
	EE3	0.833
	EE4	0.604
	EE5	0.677
	EE6	0.739
Social influence (SI)	SI1	0.709	0.843	0.475	0.844
	SI2	0.702
	SI3	0.638
	SI4	0.752
	SI5	0.659
	SI6	0.668
Hedonic motivation (HM)	HM1	0.81	0.881	0.556	0.882
	HM2	0.842
	HM3	0.754
	HM4	0.646
	HM5	0.743
	HM6	0.658
Price value (PV)	PV1	0.736	0.89	0.547	0.89
	PV2	0.755
	PV3	0.77
	PV4	0.749
	PV5	0.77
	PV6	0.766
Habit (H)	HT1	0.785	0.891	0.583	0.893
	HT2	0.834
	HT3	0.75
	HT4	0.782
	HT5	0.765
	HT6	0.654
Behavioral intention (BI)	BI1	0.544	0.789	0.386	0.787
	BI2	0.65
	BI3	0.472
	BI4	0.65
	BI5	0.69
	BI6	0.692
Ethical awareness (EA)	EA1	0.606	0.794	0.403	0.797
	EA2	0.661
	EA3	0.393
	EA4	0.721
	EA5	0.715
	EA6	0.655

, habit is a significant factor with high factor loadings (0.758–0.834) and strong reliability (Cronbach’s alpha = 0.891, CR = 0.893), and an AVE of 0.583. Meanwhile, hedonic motivation has a strong reliability (Cronbach’s alpha = 0.881) and validity (AVE = 0.557). Performance expectancy has comparatively good construct validity with the factor loadings from 0.639 to 0.777 and a high composite reliability of 0.863. Effort expectancy shows excellent internal consistency with a Cronbach’s Alpha of 0.872 and AVE of 0.541. Hedonic motivation shows excellent reliability and convergent validity with an AVE of 0.556, and it aligns well with previous research on habit in technology acceptance (Venkatesh et al., 2012). PV maintains high reliability with a composite reliability of 0.890 and an AVE of 0.547, supporting its role in technology acceptance (Davis, 1989). Social influence also shows a strong impact with reliability (Cronbach’s alpha = 0.843) and an AVE (0.475) slightly below the threshold (0.5) (Takona, 2024). At the same time, both behavioral intention and ethical awareness have relatively low AVEs of 0.386 and 0.403 but high CRs of 0.787 and 0.797, where higher composite reliability compensates for the lower AVE, maintaining model validity (Fornell & Larcker, 1981).

4.3. Discriminant Validity

The discriminant validity is assessed and presented in

Table 3. Discriminant validity between constructs.

$\sqrt[2]{\mathbf{A}\mathbf{V}\mathbf{E}}$	PE	EE	SI	HM	PV	HT	EA	BI
PE	0.717
EE	0.573 **	0.736
SI	0.583 **	0.531 **	0.689
HM	0.660 **	0.596 **	0.671 **	0.746
PV	0.507 **	0.541 **	0.441 **	0.579 **	0.740
HT	0.434 **	0.443 **	0.591 **	0.578 **	0.478 **	0.764
EA	0.235 **	0.209 *	0.144	0.208 *	0.263 **	0.121	0.634
BI	0.530 **	0.522 **	0.543 **	0.648 **	0.632 **	0.517 **	0.312 **	0.621

** Correlation is significant at the 0.01 level (2-tailed). * Correlation is significant at the 0.05 level (2-tailed).

. The square root of the AVE ($\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$) for each construct is compared to the correlations between each construct. The discriminant validity is found when $\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ is higher than its correlations with the rest of the constructs (Takona, 2024). Regarding performance expectancy, the $\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ (0.717) is higher than the correlations with constructs of effort expectancy (0.573), social influence (0.583), hedonic motivation (0.660), price value (0.507), habit (0.434), ethical awareness (0.235), and behavioral intention (0.530). This indicates robust discriminant validity for performance expectancy. Regarding effort expectancy, $\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ (0.736) exceeds its correlations with other constructs, including social influence (0.531), hedonic motivation (0.596), price value (0.541), habit (0.443), ethical awareness (0.209), and behavioral intention (0.522). This supports the unique nature of effort expectancy within the model. The statistical characterization is similar to social influence, hedonic motivation, price value, habit, ethical awareness, and behavioral intention. Significantly, habit ($\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}=0.764$) exceeds correlations with ethical awareness (0.235) and behavioral intention (0.517). Ethical awareness ($\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ = 0.634) is distinctly different from its correlations with behavioral intention (0.312), maintaining its unique position in the model, but has one lower correlation with habit (0.121). Behavioral intention ($\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ = 0.621) is higher than most of its correlations with other constructs. However, the correlation between behavioral intention and hedonic motivation, as well as price value, is slightly higher than the $\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}} \mathrm{o}\mathrm{f}$ behavioral intention. It is theoretically acceptable for the $\sqrt[2]{\mathrm{A}\mathrm{V}\mathrm{E}}$ to be slightly lower than the correlation with a certain factor and other factors (Henseler et al., 2015).

These results validate that each construct is conceptually distinct, promoting confidence in the reliability and applicability of the model. Ethical awareness and behavioral intention provide insights and promote the formulation of specialized strategies and interventions designed to independently improve these areas. These perceptions contribute valuable knowledge to the field, supporting future research and practical applications in understanding and addressing specific constructs in varied contexts.

4.4. Regression Analyses

The regression analysis determines the predictive relationship through the modeling of habit, a positive habitual automated behavior in the continuous use of technology. Figure 7 and Figure 8 illustrate the regression model.

A scatterplot of regression standardized predicted values versus standardized residuals was used to assess model fit and validate assumptions. The regression model appears to fit the data quite well, and the results show the data representation points being widely dispersed, providing no visible patterns. The well-distributed points tell us about homoscedasticity, meaning constant variance of residuals across predicted values. This is to assist the reliability of the models’ predictions at each level of independent variates. Additionally, the randomness and lack of patterns in the residuals also confirm the model’s ability to capture the relationship between predictors and habit. Such an approach to diagnosis is adhering developed best practices for technology acceptance research (Venkatesh et al., 2012). Figure 8 shows a scatterplot of the regression standardized predicted values versus the regression standardized residuals for the dependent variable behavioral intention. These qualities of the scatterplot have guided us towards a suitable and legitimate regression model for behavioral intention, which will help enforce finding robustness and inference reliability in the results presented in this area of study. Previous research highlights the importance of these types of diagnostics in contexts such as technology acceptance and behavioral intention (Davis, 1989; Venkatesh et al., 2012). Such a distribution indicates that the linear relationship assumed by the regression model is sound (Henseler et al., 2015), and it recommends that the effect of an AI-infused curriculum on learning outcomes is gradual and predictable.

5. Discussion

This study set out to unravel the complex psychological mechanisms driving CS students’ adoption of GenAI within a higher education programming curriculum. By integrating ethical awareness into the UTAUT2 framework and implementing a dual-track pedagogical intervention, we aimed to move beyond the traditional efficiency-centric view of technology acceptance. The SEM results (as shown in Figure 9 and

Table 4. Model fitness test metrics.

χ2/df	RMSEA	GFI	AGFI	CFI	IFI	TLI
1.52	0.062	0.691	0.655	0.845	0.849	0.835

and

Table 5. Hypothesis test results.

Hypotheses	Independent Constructs	Dependent Constructs	r2	Path Coeff	t	p-Value	Results
H1	Hedonic motivation	Habit	0.366	0.457	5.356	0.000	Supported
H2	Hedonic motivation	Behavioral intention	0.507	0.336	3.377	0.001	Supported
H3	Ethical Awareness	Behavioral intention	0.507	0.166	2.594	0.011	Supported
H4	Ethical Awareness	Habit	0.366	−0.032	−0.45	0.653	Not supported
H5	Performance expectancy	Behavioral intention	0.507	0.076	0.863	0.39	Not supported
H6	Effort expectancy	Behavioral intention	0.507	0.125	1.523	0.13	Not supported
H7	Social Influence	Behavioral intention	0.507	0.086	0.939	0.349	Not supported
H8	Price value	Habit	0.366	0.222	2.565	0.011	supported
H9	Habit	Behavioral intention	0.507	0.162	2.012	0.046	Supported

) provide strong empirical support for a paradigm shift from utilitarian tool adoption to an experience-driven, ethically regulated partnership. As shown in Table 4, the χ²/df ratio (1.52) was well below the threshold (3), indicating a sound balance of model complexity and fit (Kline, 2016). The RMSEA (0.062) is also below the 0.08 cutoff, approaching the ideal 0.06 value and showing minimal population-level approximation error (Hu & Bentler, 1999). Relative fit indices (CFI = 0.845, IFI = 0.849, TLI = 0.835) all exceed the 0.8 threshold, consistent with small-sample SEM standards in the social sciences (Marsh et al., 2004). Traditional indices AGFI (0.655) and GFI (0.691) are lower, but this reflects their sensitivity to the small size of samples (n = 136) and model complexity, with modern indices now preferred (Byrne, 2016).

5.1. The Eclipse of Utility: Efficiency as a Hygiene Factor

A striking finding from our model is the non-significance of performance expectancy (β = 0.076, p = 0.39) and effort expectancy (β = 0.125, p = 0.13). This defies the traditional logic of technology acceptance, where utility is typically the primary driver (Venkatesh et al., 2012). We interpret this through the lens of the hygiene factor hypothesis (Tamilmani et al., 2021). For digital native students, the ability of AI to generate code quickly and easily is no longer a differentiator, and it is a baseline expectation. Its presence does not motivate, but its absence would dissatisfy. This suggests that in the era of LLMs, utilitarian efficiency may have hit a ceiling effect, thereby losing its predictive power for behavioral intention.

5.2. The Primacy of Pedagogical Novelty: Hedonic Motivation as the Engine

One of the most significant findings is the dominance of hedonic motivation as the strongest predictor of both habit (β = 0.457, p < 0.001) and behavioral intention (β = 0.336, p = 0.001). Contrary to early technology acceptance studies, where utilitarian drivers (performance expectancy) were paramount (Venkatesh et al., 2012), our results indicate that for digital native students, the efficiency of AI is taken for granted and effectively devolved into a hygiene factor (Tamilmani et al., 2021). Consequently, performance expectancy and effort expectancy showed non-significant effects on intention. This shift suggests that the driver of adoption has moved from tool utility to pedagogical novelty. The interactive, conversational nature of GenAI introduces a layer of situational interest that static textbooks lack (Huang et al., 2022). This novelty effect transforms the coding process from a solitary struggle with syntax into a collaborative dialog, generating intrinsic enjoyment (Wu et al., 2024; J. Li et al., 2025). This finding aligns with the flow theory in programming education and suggests that AI reduces the frustration barrier. This allows students to enter a state of cognitive absorption that reinforces habit formation.

5.3. The Cognitive Brake Mechanism: System 1 vs. System 2

The study contributes a novel theoretical insight regarding the divergent role of ethical awareness. While it significantly positively influences behavioral intention (β = 0.166, p = 0.011), it has a non-significant, slightly negative relationship with habit (β = −0.032, p = 0.653). This apparent contradiction can be explained through dual process theory (Kahneman, 2011). Ethical awareness and intention can be regarded as the cognitive legitimacy. The positive link to intention suggests that ethical training empowers students. By clarifying the gray areas of academic integrity, ethical awareness reduces perceived risk and provides the cognitive legitimacy to use AI responsibly (Dwivedi et al., 2020, 2023). Students intend to use AI because they know how to use it correctly. Ethical awareness and habit act as the cognitive brake. Habit, by definition, implies automaticity (system 1). The lack of a link to habit confirms our cognitive brake hypothesis. Habit formation relies on system 1 thinking (automaticity, speed, low cognitive load). However, ethical reasoning involves system 2 thinking (deliberative, reflective, high cognitive load). High ethical awareness compels students to pause and evaluate the propriety of AI assistance for each specific task. This deliberative process acts as a cognitive brake, actively disrupting the formation of mindless, automated usage loops. Thus, the lack of a significant path to habit is not a failure of the intervention but a pedagogical success. It indicates that ethically aware students are resisting dependency and maintaining a reflective distance from the technology, and proves that the ethical scaffolding intervention successfully prevented students from sleepwalking into mindless dependency.

5.4. Rational Constraints in an Experiential Model

Despite the dominance of hedonic drivers, the significance of price value (β = 0.222, p = 0.011) underscores that students remain rational economic actors. In the higher education context, where students may be price-sensitive, the perceived cost-effectiveness of premium AI tools acts as a practical constraint. This suggests that while pedagogical novelty sparks interest, sustainable adoption requires a perception of fair value, aligning with recent findings on the freemium models of EdTech (Strzelecki, 2024). Social influence was non-significant (β = 0.086, p = 0.349). This indicates that for higher education programming tasks, adoption is a highly personal, task-oriented decision driven by internal experience (hedonic motivation) rather than external peer pressure, highlighting the individualistic nature of human–AI collaboration.

5.5. Theoretical Implications

This study extends the UTAUT2 framework in two key dimensions. First, it validates the hygiene factor hypothesis in the context of GenAI, proposing that utilitarian constructs may lose predictive power when technology reaches a saturation point of efficiency. Second, it introduces the dual regulation model of ethics, demonstrating that ethical awareness is not merely an antecedent to behavior but a distinct regulatory mechanism that promotes conscious intention while inhibiting automatic habituation.

5.6. Practical Implications for Educators and Policymakers

Educators should design curricula that exploit the pedagogical novelty of AI instead of just teaching syntax. Assignments should focus on AI-human collaboration and prompt engineering rather than rote memorization and maintain the situational interest that drives engagement. Ethical training should move beyond punitive warnings to competency building, which enables intention. The goal should be to foster intention without mindless habit. Educators need to introduce friction in the workflow to engage system 2 thinking and prevent the atrophy of critical thinking skills.

5.7. Long-Term Sustainability of the Novelty Effect

Sustaining AI’s pedagogical novelty requires translating its transient situational interest into enduring, cognitively engaged learning, rather than relying on its initial appeal alone. As learners grow familiar with AI tools, novelty naturally fades, increasing the risk of mindless reliance and eroded engagement if curricula focus only on surface-level use. To avoid this, it is necessary to design iterative, friction-rich workflows that keep system 2 critical thinking active and to replace punitive ethical warnings with competency-building to foster intentional judgment. This design cultivates intentional AI use over mindless habit, ensuring AI’s pedagogical value persists long after initial novelty diminishes and preventing critical thinking atrophy, sustaining meaningful engagement for long-term learning.

6. Limitations and Future Research

While this study offers robust insights, several limitations warrant future investigation. First, the participants were recruited from regions at different levels of economic development in China and had varying experience with computer use, and males accounted for a large proportion of the sample. Future research may adjust the study design to address these factors. Second, the novelty effect is temporally sensitive, and future longitudinal studies should examine whether the dominance of hedonic motivation persists as AI tools become ubiquitous. Third, the measurement of habit relied on self-reports, and future research could employ system logs to capture objective usage frequency and behavioral patterns. Fourth, the cognitive brake hypothesis may warrant experimental validation by using neurophysiological measures to observe the real-time cognitive load during ethical decision-making in coding tasks. Finally, an important direction for future research is to explicitly examine the cross-cultural and longitudinal extensions of the model, thus enhancing its generalizability and practical value in varied settings over time.

7. Conclusions

As GenAI is reshaping the landscape of CS education, understanding the psychological drivers of students’ adoption is significant. This study, grounded in a quasi-experimental design within a Python programming curriculum, reveals that the adoption of AI tools is no longer driven by mere efficiency. Instead, it is fueled by the pedagogical novelty and intrinsic enjoyment of the human–AI partnership. Crucially, we identified a cognitive brake mechanism where ethical awareness legitimizes conscious usage intention while preventing the formation of blind dependency habits. For educators and policymakers, in terms of design for novelty, curriculum design should concentrate on co-creating with AI instead of using commands and leverage the tool’s interactive nature to sustain student interest. In addition, ethical training should not just be a lecture but a procedural requirement that triggers system 2 thinking, preventing the formation of inactive habits. Meanwhile, given the significance of price value, institutions should consider subsidizing premium AI tools to ensure equitable access to these generative partners. Ultimately, this research suggests that the integration of AI in education is not a binary choice between prohibition and unrestrained use. By balancing the engaging power of pedagogical novelty with the reflective constraints of ethical awareness, educators can transform AI from a crutch for efficiency into a catalyst for critical, creative, and responsible learning.