Students' engagement with generative AI in academic learning: A self-determination theory and epistemic network analysis study

Synthesis: Isaeva et al. (2026) examine undergraduate students' engagement with generative AI (GenAI) in academic learning at an English-medium university, using self-determination theory (SDT) as the interpretive framework and epistemic network analysis (ENA) to model the structural relationships among themes. Analysis of 23 semi-structured interviews revealed that students frequently described GenAI as supporting efficiency and conceptual understanding, yet their accounts exposed persistent tensions concerning Creativity, trust, and academic integrity. The ENA results showed these concerns were systematically interconnected — discussions of learning support consistently co-occurred with verification practices, reflecting a "trust-but-verify" repertoire through which students calibrated their reliance on AI while maintaining epistemic control.

Key Findings

• Students' engagement with GenAI is best understood as a motivated, value-oriented learning practice shaped by the negotiation of the three SDT needs: autonomy, competence, and relatedness — not simply a technology-acceptance decision.
• Students valued GenAI for efficiency and conceptual support (19/23 described time-saving; 15/23 conceptual understanding; 12/23 skill development), but 8 participants worried excessive reliance would diminish creativity or critical thinking: "When we ask it to write for us, it's not our creativity."
• The ENA revealed a central cluster interconnecting conceptual support, skill development, and prompting/personalization with cross-checking with sources — a "trust-but-verify" repertoire where students incorporated AI outputs into Active Learning with verification, rather than fully delegating cognition.
• A second key relationship co-occurred creativity-vs-dependency with verification practices, reflecting autonomy negotiation: students attempted tasks independently before consulting AI or verified outputs against external sources.
• 16/23 raised academic-integrity concerns and distinguished legitimate assistance from cheating; 12/23 noted the lack of explicit institutional AI policies ("University doesn't have a clear and unified policy yet"); 19/23 positioned AI as an assisting tool rather than a substitute for human instruction.
• Beyond instrumental concerns, students raised privacy (7), algorithmic bias (3), and environmental Sustainability (6, e.g., water/energy use of large AI systems) — indicating broader value- and norm-related considerations.

Study Design & Method

This qualitative case study was conducted at an English-medium university, examining 23 undergraduate students via semi-structured interviews. The authors used self-determination theory (SDT) as the primary interpretive framework, treating technology-acceptance perceptions (usefulness, ease of use) as descriptive cues rather than explanatory constructs. Data were analyzed using reflexive thematic analysis, complemented by epistemic network analysis (ENA) — an educational data-analytics method that models the structural co-occurrence of themes in discourse. 1,015 paragraph-level stanzas from the 23 interviews were coded (with automated coding validated against a manually coded subset); 23.0% contained two or more co-occurring codes. A permutation test (N=500) confirmed the co-occurrence structure significantly deviated from random expectations (p = .002). ENA edges denote the weighted strength of co-occurrence and are interpreted as structural relationships in reasoning, not causal effects.

What this means for practice

• Learners. Make verification part of the routine rather than an occasional check: students whose accounts paired GenAI use with cross-checking against sources were the ones who described keeping conceptual work and judgment in their own hands.
• Learners. Attempt a task independently before consulting the AI when the point is learning rather than speed — the "trust-but-verify" repertoire and the creativity-versus-dependency worries (eight participants) both turned on negotiating autonomy.
• Instructors. Design assessment so the reasoning process is visible, because the central ENA cluster links conceptual support and skill development to cross-checking rather than to delegation.
• Instructors. Put a unified AI policy in writing: twelve participants named the absence of clear institutional guidance as a live problem, and students were left to distinguish legitimate assistance from cheating on their own.

Limitations

The study is a qualitative case study in a single English-medium university setting (with evidence from Azerbaijan, a post-Soviet context where institutional AI policies are underdeveloped), bounding generalizability. The sample is 23 students, and the ENA is exploratory — co-occurrence is modeled as structural relationship, not causation. The authors rely on students' self-reports of their AI practices rather than direct observation. The study was conducted before and during ongoing institutional policy development, so students' accounts may reflect a transitional context.

Citation

Isaeva, R., Caner, H. N., Caner, M., Giray, L., & Karadag, E. (2026). Students' engagement with generative AI in academic learning: A self-determination theory and epistemic network analysis study. Computers and Education: Artificial Intelligence.