The seminal work of Marton and Säljö (1976) established the distinction between Deep and Surface Approaches to learning, a framework widely used to explain how learners cognitively engage with academic tasks. These approaches reflect both the intention and strategies that students employ.

Deep Approach (DA) is driven by intrinsic motivation and involves seeking underlying meaning, engaging actively with content, constructing personal interpretations, and integrating new knowledge with prior understanding (Marton and Säljö, 1976). Students employing DA typically demonstrate critical analysis, conceptual synthesis, long-term retention, and the ability to transfer knowledge across contexts (Gillaspy and Vasilica, 2021). Surface Approach (SA) is associated with extrinsic motivation, such as preparing for examinations, and focuses on memorizing disconnected facts or completing tasks with minimal cognitive effort. This approach often results in cognitive overload and dependence on rote strategies (Marton and Säljö, 1976).

Structural parameters such as rigid curricula, excessive workloads, and assessment strategies that prioritize factual reproduction have been empirically associated with students’ adoption of a SA, even in the presence of initially high motivation (Johnson et al., 2021; Kember, 2004). In contrast, learning designs that emphasize authentic inquiry, contextualized scaffolding, and reflective judgment are foundational for the cultivation of a DA, particularly by supporting learners’ epistemic engagement and meta-cognitive regulation (Inouye et al., 2023; Chin and Brown, 2000).

Understanding the parameters that influence these attentional states is critical for designing effective deep learning environments. See Table 1 for details. Cognitive load stands as a primary determinant; research indicates that poorly designed digital interfaces that impose high extraneous cognitive load can exhaust the limited capacity of DA (Skulmowski and Xu, 2022; Paas et al., 2003). Conversely, interactivity and engagement features, such as AI-driven adaptive feedback and personalized learning paths, can enhance intrinsic motivation. By keeping learners actively involved, these features support the maintenance of SA (Serrano et al., 2019). However, the digital environment is rife with potential disruptors. The presence of distractions, whether from social media notifications or the temptation to multitask, significantly compromises SA (Soyoof et al., 2024; Unsworth and McMillan, 2014). Furthermore, while the novelty of AI-generated content can initially spike DA, maintaining the SA required for deep conceptual processing necessitates that the content be perceived as relevant and valuable to the learner’s long-term goals (Chen et al., 2022).

Heutagogy conceptualized in its modern form by Tsakeni et al. (2025), extends beyond the principles of andragogy by placing full responsibility for learning decisions on the learner. Whereas andragogy emphasizes structured self-directed learning toward predefined competencies, heutagogy foregrounds the development of capability: the learner’s capacity to adapt, reflect, and act effectively in unfamiliar or changing contexts (Blaschke, 2012). The differences between the two are detailed in Table 2. This orientation aligns with Sen’s Capability Approach, which conceptualizes human development as the expansion of agency, freedom, and valued functioning, rather than the mere accumulation of discrete skills (Eder, 2025; Walker, 2005). From this perspective, the objective of higher education is to enhance learners’ freedom to achieve by cultivating self-efficacy and meta-cognitive awareness, particularly the capacity to know how to learn. Therefore, a heutagogical approach contributes to “capability” by transforming students from passive recipients of instruction into active agents who can leverage resources, including powerful tools like Generative AI, to realize their potential and lead flourishing lives (Holdsworth and Thomas, 2021).

Central to heutagogy is Double-Loop Learning, a concept introduced by Argyris (2002). Double-loop learning requires learners not only to correct errors (single-loop learning) but also to examine and modify the underlying assumptions, values, and strategies that produce those errors (Auqui-Caceres and Furlan, 2023). This process represents the core mechanism of deep cognitive transformation within heutagogical environments. Empirical qualitative research demonstrates that heutagogical designs can facilitate meta-cognition, reflective judgment, and non-linear learning, key components of capability development (Gillaspy and Vasilica, 2021; Merriam, 2001).

The shift toward DA within heutagogy is underpinned by motivational and cognitive processes that sustain deep engagement.

Self-Determination Theory (SDT) explains how intrinsic motivation arises from the satisfaction of three basic psychological needs: autonomy, competence, and relatedness (Deci and Ryan, 2000). Heutagogical environments inherently support autonomy by allowing learners to define their goals and pathways, thereby strengthening intrinsic motivation essential for sustained deep learning (Blaschke, 2012).

Meta-cognitive Theory comprises Knowledge of Cognition (KoC) and Regulation of Cognition (RoC), the latter referring to learners’ abilities to plan, monitor, and evaluate their strategies (Schraw and Dennison, 1994). RoC is strongly associated with adaptability and deep learning outcomes (Wan et al., 2025). Because heutagogy emphasizes continuous reflection and self-assessment (Gillaspy and Vasilica, 2021), it directly strengthens RoC, which is a powerful predictor of academic performance (Schraw and Dennison, 1994).

Together, SDT and meta-cognitive theory provide the motivational and cognitive foundations that explain why heutagogical environments can effectively promote DA.

Double-loop learning, which involves questioning underlying assumptions rather than merely correcting errors, is cognitively and emotionally demanding (Clark, 2021; Argyris and Schön, 1997). Self-Determination Theory helps explain why learners engage in this process: autonomy provides the internal endorsement necessary to challenge established norms, competence fosters the confidence to analyze and revise mental models (Chiu, 2021; Bandura, 1997), and relatedness ensures a psychologically safe environment for critical reflection (Ryan and Deci, 2020). In heutagogical and AI-mediated learning environments, these needs are supported through learner-defined goals, immediate feedback, and collaborative or conversational interfaces, positioning students to undertake the deep, transformative reflection characteristic of double-loop learning.

As GenAI becomes embedded in educational contexts, it introduces a new layer of complexity to learning engagement (Chan and Tsi, 2024; AlAli and Wardat, 2024). Current research highlights two dominant modes of student interaction with GenAI.

Students rely on AI primarily for task completion, answer retrieval, or reproducing existing solutions, especially when motivated by efficiency or minimal effort. This mode is negatively associated with autonomy, critical thinking, and applied knowledge (Pallant et al., 2025; Lee et al., 2024; Habib et al., 2025). In contrast, constructive use involves leveraging AI to inform deeper inquiry, extend existing knowledge, explore alternative perspectives, or generate preliminary drafts for critical evaluation. This mode correlates positively with deeper understanding and higher-order cognition (Pallant et al., 2025; Lee et al., 2024). Between the instrumental and constructive poles lies a continuum of AI engagement. Learners may begin with largely procedural use, such as drafting outlines or retrieving answers, but through prompts, scaffolding, or reflective tasks, they can progress toward more evaluative and constructive interactions. This spectrum has been conceptualized in levels, from minimal AI support to AI acting as a creative partner in knowledge construction (Lan et al., 2025). Recognizing this continuum is essential for educators, as the aim is not to prohibit instrumental use, but to guide students progressively toward deeper, reflective, and self-directed engagement.

These contrasting modes underscore the need for intentional pedagogical design. GenAI tools, such as coding assistants, inherently support non-linear and self-paced learning, qualities that align with heutagogical principles (Abdelhalim and Almaneea, 2025). However, without explicit guidance or structured reflection, the convenience of AI often reinforces SA rather than promoting DA (Habib et al., 2025). Therefore, pedagogical models must incorporate meta-cognitive scaffolding that directs learners toward constructive use and strengthens the RoC dimension of meta-cognition (Wan et al., 2025). Meta-cognitive scaffolding provides learners with support to plan, monitor, and evaluate their own learning (Zimmerman and Schunk, 2011). Traditionally, this role was fulfilled by human tutors; however, recent research highlights the potential of AI to extend this support. For instance, Akram et al. (2025) demonstrated that AI systems delivering real-time meta-cognitive prompts significantly enhanced students’ problem-solving performance and meta-cognitive awareness. Similarly, Azevedo et al. (2022) showed that intelligent tutoring systems can effectively trigger reflection on learning strategies. In the context of Generative AI, this potential is further amplified. Sok and Heng (2024) and Banjade et al. (2024) illustrate that Large Language Models (LLMs) can act as Socratic tutors rather than mere answer providers, prompting learners to justify their reasoning and engage in adaptive, personalized reflection. Collectively, this body of work indicates that AI can serve as a scalable tool for fostering the “learning to learn” competencies central to heutagogical practice.

Although theoretical literature strongly supports heutagogy’s potential, its empirical foundation remains limited (Ramas et al., 2023; Gillaspy and Vasilica, 2021). Several methodological challenges contribute to this gap. Distinguishing Heutagogy from Andragogy. Differentiating between structured self-directed learning and full self-determination is difficult in empirical settings (Gillaspy and Vasilica, 2021). Measuring Capability and Double-Loop Learning. Existing quantitative instruments do not adequately capture higher-order cognitive constructs such as capability, reflective judgment, or double-loop learning (Webber, 2012; Giannakos et al., 2025). GenAI-Specific Challenges. The rapid evolution of GenAI outpaces empirical research, making it difficult to isolate the effects of specific pedagogical interventions on learning outcomes (UNESCO, 2023).

Despite recent advancements, two key gaps remain. Methodologically, there is a scarcity of quasi-experimental studies that isolate the causal impact of heutagogical structures on specific meta-cognitive dimensions such as RoC (Gillaspy and Vasilica, 2021). A gap in theory is that current models describe how AI influences learning but do not clarify the mediating role of heutagogical agency in directing cognitive engagement and preventing default surface approaches.

Figure 1 presents the theoretical model of the HDL framework, illustrating how heutagogical design influences deep learning outcomes indirectly through double-loop reflection and metacognitive regulation. The present study addresses this gap by testing a model in which self-determined scaffolding operates as the primary regulator of students’ interaction with AI, thereby linking instructional design, meta-cognitive regulation, and deep learning outcomes.

A quasi-experimental, non-equivalent groups, pre-test or post-test control group design was employed. The study spanned an eight-week duration during a compulsory third-year module on “Applied Data Science and Ethical Practices” at a major technological university in China. This duration was selected for two reasons. First, it aligns with a standard mid-semester instructional cycle, allowing sufficient time to implement the intervention without disrupting the regular curriculum; Second, prior research indicates that a minimum exposure period of 4–8 weeks is required to observe measurable changes in self-regulated learning behaviors and meta-cognitive strategies without fostering short-term novelty effects or tool dependency (Hemmler and Ifenthaler, 2024; Johnson et al., 2021; Gillaspy and Vasilica, 2021). The mixed-methods approach integrated validated quantitative surveys (R-SPQ-2F and MAI) with qualitative Framework Analysis of student reflective journals and interviews, ensuring triangulation of findings (Johnson et al., 2021; Gillaspy and Vasilica, 2021).

The study population consisted of 300 undergraduate students (n = 300) enrolled across three sections of a mandatory module. Individual-level randomization was not feasible due to institutional and administrative constraints. Students are enrolled in “intact class” cohorts based on their major, and altering these groups would have disrupted established scheduling and enrollment policies. Additionally, randomizing within the same sections could have caused instructional contamination, as students frequently share resources and collaborate outside class. Consequently, existing sections were used as non-randomly assigned comparison groups: two sections formed the HDL Intervention Group (n = 150), which received the HDL curriculum, and one section formed the Traditional Control Group (n = 150), which received the standard curriculum.

To ensure baseline equivalence and mitigate selection bias, independent samples t-tests were conducted for continuous variables, and chi-square tests were performed for categorical variables. As shown in Table 3, no statistically significant differences were observed between the groups in demographic characteristics, prior academic performance, or pre-test scores on key dependent measures (p > 0.05). These results confirm the initial equivalence of the non-randomized groups, providing a robust foundation for attributing post-intervention outcomes to the HDL framework.

Instructors in the HDL group shifted from being primary knowledge sources to “orchestrators” of learning processes (Blaschke, 2012). This role was operationally defined by the active relinquishing of curricular ownership and the facilitation of a shift in learning responsibility. Concrete pedagogical actions included: (a) facilitating “Goal-Negotiation” sessions where students co-defined project criteria; (b) providing adaptive scaffolding only when learners encountered “zones of uncertainty” exceeding their current capability; and (c) curating a suite of digital resources (e.g., GitHub, specialized LLM agents) to support learner-generated content rather than providing textbook solutions (Saleem et al., 2024).

The intervention (Weeks 2–8) consisted of three interdependent components designed to operationalize heutagogical principles by promoting autonomy, constructive AI use, and reflective practice.

Figure 2 illustrates the cyclical implementation flow of the HDL framework, highlighting how goal negotiation, non-linear exploration, constructive GenAI interaction, and double-loop reflection iteratively contribute to learners’ capability development.

Students in the HDL group were required to define their capstone project’s specific research question, methodology, and the criteria for assessing its real-world applicability, a core principle of heutagogy. Educators shifted roles to become “orchestrators” (Blaschke, 2012), providing guidance and resources for non-linear learning paths, but explicitly avoiding prescription of knowledge or methods (Gillaspy and Vasilica, 2021).

All students were granted access to GenAI coding and analysis tools. However, HDL students were mandated to follow a structured protocol enforcing Constructive Use (Lee et al., 2024). For instance, in data analysis tasks, students were explicitly instructed to use GenAI to: (a) generate comparative data models, (b) draft initial synthesis summaries, and (c) identify potential ethical constraints. Critically, direct GenAI output was not accepted as a final deliverable. This protocol guided students to move beyond simple task completion and engage in critical thinking, synthesis, and ethical evaluation. Example prompts are shown in Table 4.

The core of the intervention was the mandatory submission of a Weekly Reflective Portfolio. The assessment criteria are shown in Table 5. This portfolio required a two-part reflection (Wan et al., 2025).

A2 (Group: HDL & Control) × 2 (Time: Pre-test & Post-test) Mixed-Design Analysis of Variance (ANOVA) was conducted for each of the six dependent variables (DM, DS, SM, SS, KoC, RoC). The level of significance was set conservatively at p < 0.01 to minimize Type I error given the non-randomized design. The level of significance was set conservatively at p < 0.01 rather than the conventional p < 0.05 in order to control for inflated Type I error arising from multiple comparisons across several dependent variables within the mixed-design ANOVA framework. Given the quasi-experimental, non-randomized design and the simultaneous testing of six outcome variables, a stricter significance threshold was adopted as a pragmatic correction to enhance statistical rigor and reduce the likelihood of false-positive findings. Partial eta squared (ηp²) was used to report effect sizes.

The Mixed-Design ANOVA revealed highly significant Group x Time interaction effects across all four R-SPQ-2F subscales, providing strong support for Hypothesis H1 (see in Table 6).

The HDL Group showed a significant increase in both Deep Motive and Deep Strategy (DM: +0.76; DS: +0.77), whereas the Control Group’s increases were minimal (DM: +0.13; DS: +0.13). The large, positive interaction effect for the HDL group suggests that the intervention was highly effective in fostering both the intention and the actual practice of deep learning (Biggs et al., 2001).

Conversely, the HDL Group showed a substantial and significant reduction in both Surface Motive and Surface Strategy (SM: -0.85; SS: −0.90), indicating a fundamental shift away from learning approaches centered on rote memorization and procedural completion. The Control Group’s scores remained relatively stable, confirming that the simple presence of GenAI without structural intervention does not significantly alter the established surface learning habits (Habib et al., 2025).

The analysis of MAI scores provided the most compelling evidence regarding the mechanism of action, confirming Hypothesis H2 (see in Table 7).

A highly significant Group x Time interaction effect was found for RoC (p < 0.001), with a strong effect size (ηp² = 0.097). The HDL Group’s increase in RoC score (+0.77) was substantial and significantly higher than the Control Group’s (+0.08). This finding confirms that the mandatory process documentation and critical justification protocol successfully enhanced the students’ ability to monitor, evaluate, and adapt their learning strategies, the core of self-regulation (Schraw and Dennison, 1994; Wan et al., 2025).

No significant interaction effect was observed for KoC (p = 0.323). This suggests the intervention did not merely increase awareness of strategies (KoC), but rather improved the active application and use of these strategies (RoC), validating the efficacy of the process-focused, scaffolding approach (Wan et al., 2025).

The qualitative analysis of the reflective portfolios and interviews provided the necessary contextual depth to interpret the quantitative shifts, specifically confirming the mandatory adoption of double-loop learning and constructive AI use (Gillaspy and Vasilica, 2021).

A Pearson correlation analysis was conducted to examine the relationship between the depth of double-loop reflection and post-intervention RoC scores. A strong, significant positive correlation was identified (r = 0.76, p < 0.001), confirming that the reflective process is the primary mechanism driving meta-cognitive growth. This validated H3 (see in Table 8).