Experiential Versus Instructional Approaches for Eliciting Metacognitive Awareness in AI-Assisted Learning

Synthesis: A quasi-experimental, short-term longitudinal study with 126 first-year engineering students comparing two ways of teaching students how to learn with generative AI: an experiential, hands-on session versus a classical instructional lecture. Metacognitive awareness — both knowledge of cognition (understanding effective AI-use strategies) and AI Regulation in Education of cognition (applying that knowledge in practice) — was measured before and after a two-hour session and tracked longitudinally. The design directly addresses a gap flagged across the knowledge base's Metacognition thread: interventions that build awareness of one's own AI-assisted thinking, rather than just policing AI use.

The study speaks to the Cognitive Offloading and Over-Reliance literature by testing whether metacognitive skills for AI-assisted learning are better acquired by doing (experiencing GenAI's strengths and failure modes firsthand) or by being told. It extends Mapping the Scaffolding of Metacognition and Learning by AI Tools in STEM Classrooms: A Bibliometric-Systematic Review with primary quasi-experimental evidence in an engineering-course context and connects to Self-Regulated Learning as a proximal training target for durable, transferable AI-use strategies.

What this means for practice

• Learners. Work through productive and unproductive GenAI use cases hands-on rather than hearing them described: the experiential condition had students attempt two productive and two unproductive collaboration scenarios in roughly 5-10 minute cycles followed by structured reflection.
• Learners. Do not treat one session as sufficient: both conditions reached comparable levels of self-reported AI engagement by the five-week follow-up (M = 3.93), so schedule repeated cycles of independent GenAI use with reflection across a term.
• Learners. Attend to AI Regulation in Education of cognition separately from knowledge of cognition, as the study measured them: knowing which AI-use strategies work and actually applying them during a task were distinct components of metacognitive awareness.
• Learners. Judge progress against what you can do unaided rather than against how the session felt, because the intervention asked participants to reconstruct their pre-task understanding after the activity rather than measuring it at baseline.

Limitations

• Metacognition was assessed only through repeated self-report administrations of the same questionnaire (MAI-AI), which the authors note carries social-desirability bias, varying accuracy of self-assessment, possible familiarity effects from repeated exposure, and a still-pending validation.
• The final analyzed sample of N = 126 (61 in the instructional group, 65 in the experiential group) fell short of the 128 participants required by the a priori power analysis for a medium effect, so the non-significant five-week difference may reflect limited power rather than a true absence of difference.
• All participants were first-year engineering students in one course at Universitat Pompeu Fabra, Barcelona, and no demographic data were collected, so the authors warn that generalization to other disciplines, educational levels and cultural contexts is unestablished.
• The design is quasi-experimental, with conditions assigned by class schedule rather than individually randomized, and students could use GenAI as they personally saw fit between the initial session and the five-week measurement, leaving that independent use uncontrolled.

Citation

Benazet i Montobbio, P., Rotter, J., & Hernández-Leo, D. (2026). Experiential Versus Instructional Approaches for Eliciting Metacognitive Awareness in AI-Assisted Learning. arXiv preprint (cs.CY/cs.HC).