The rapid integration of Conversational Generative Artificial Intelligence (GenAI) into higher education has sparked intense pedagogical debates regarding its impact on student cognition. This study aims to deconstruct the micro-level interactive processes through which university students engage with large language models (LLMs), mapping the systemic socio-technical mechanisms that shape, extend, or restrict the development of independent critical thinking skills.
Adopting an exploratory qualitative research paradigm, this study used a combination of purposive and maximum variation sampling strategies to recruit n = 55 undergraduate and postgraduate students across diverse disciplinary tracks (STEM, medicine, humanities and social sciences) at a public university. Data collection was executed through semi-structured depth interviews incorporating stimulated recall methodologies and live interaction histories. The qualitative texts were analyzed using Braun and Clarke’s inductive thematic analysis methodology guided by Grounded Theory coding protocols via NVivo, augmented by cross-dimensional matrix queries across student subgroups.
The analysis generated an integrated socio-technical framework comprised of four core themes. The Paradox of Cognitive Offloading vs Extension, where automated text consolidation yields significant efficiency dividends but risks fostering cognitive inertia and domain-specific skill atrophy, particularly in technical code and math execution. The Awakening of Algorithmic Discernment through Epistemic Vigilance, revealing that model imperfections, flawed logic and fabricated bibliographic citations serve as unexpected cognitive catalysts that disrupt passive trust and force active multisource validation loops. The Regulatory Role of Individual Prompt Literacy, distinguishing low-literacy single-turn commands from advanced interactive conditioning workflows that deconstruct complex problems; and the Scaffold Effect of Process-Oriented Pedagogical Interventions, demonstrating that instructor-imposed structural constraints create “desirable difficulties” that effectively preserve human cognitive agency.
The research relies on qualitative self-reported reflections during depth interviews and represents a cross-sectional snapshot of rapid technological evolution. Future research should integrate objective backend telemetry data with multi-semester longitudinal tracking.
Grounded directly in empirical student interaction data, the paper provides an actionable blueprint for higher education governance, advocating for a transition from unenforceable prohibition policies toward structured “Critical Integration.” Key recommendations include rewriting curriculum assignments to evaluate process-oriented modification logs rather than final written products, embedding formal prompt literacy into core curricula and training students to deploy multisource triangulation protocols against digital hallucinations.
Moving beyond descriptive behavioral intention surveys and binary debates labeling GenAI as either a destructive shortcut or a perfect accelerator, this study opens the cognitive “black box” of active human–AI interaction. By analyzing data from a large qualitative sample (n = 55), this inquiry extends classical Cognitive Offloading Theory into high-order semantic and logical outsourcing, providing an empirical framework for process-focused pedagogy in smart learning environments.
