Learning-to-learn in the age of generative AI: A scoping review and conceptual framework

Synthesis: This paper presents a scoping review of learning-to-learn (L2L) definitions within pedagogical and psychological literature, identifying 21 relevant publications via PRISMA-ScR. It proposes a novel three-layered framework organized by conceptual broadness: Dimensions (cognitive and metacognitive skills), Processes (self-AI Regulation in Education), and Tools (retrieval practice). The framework maps L2L components to GenAI application use cases in higher education, positioning L2L as key to reducing GenAI overreliance and fostering learner agency.

Key Findings

• A scoping review of learning-to-learn (L2L): Identifies 21 relevant publications via PRISMA-ScR, mapping how L2L is defined across pedagogical and psychological literature.
• A three-layered framework: Organized by conceptual broadness — Dimensions (cognitive and metacognitive skills), Processes (self-regulation), and Tools (retrieval practice).
• Mapping L2L to GenAI use cases: The framework maps L2L components to GenAI application use cases in higher education, positioning L2L as key to reducing GenAI overreliance and fostering learner agency — a contribution to AI-mediated learning and AI literacy.

What this means for practice

• Instructors. Embed GenAI support in evidence-informed instructional routines that target learning processes rather than deliver answers — for example, prompting self-explanations before supplying hints.
• Instructors. Distinguish the three layers when designing an intervention: work on goals and beliefs at the Dimensions layer, strategy selection and monitoring at the Processes layer, and concrete learning strategies at the Tools layer.
• Instructors. Foster critical awareness of GenAI use with confidence-estimation visualizations, ethical reflection, and prompts that surface the risk of over-reliance.
• Instructional designers. Design GenAI systems to assist rather than replace self-regulation — Socratic or decision-tree dialogue structures and Open Learner Models — preserving learner agency and independence.
• Researchers. Address the social dimension explicitly: current AI systems cannot replicate the interpersonal dynamics of human interaction, so investigate how GenAI complements rather than substitutes for human collaboration.

Limitations

• Only 21 studies met the PRISMA-ScR criteria from 787 records (searched May 16, 2025), and roughly 43% (n = 9) were empirical; their sample sizes ranged from 67 to 13,500.
• English-language only, and the Scopus and Web of Science searches deliberately excluded computer science and engineering subject areas, a decision the authors concede may have omitted interdisciplinary work.
• Screening, extraction, and coding were handled by the first and second authors, with the first author reading each paper multiple times; no multiple independent encoders were involved.
• L2L remains difficult to measure reliably as a complex, multidimensional construct, and the review captures literature only up to its search date in a fast-moving field.

Citation

Schorr, I., Bardach, L., Bühler, B., & Kasneci, E. (2026). Learning-to-learn in the age of generative AI: A scoping review and conceptual framework. Computers and Education: Artificial Intelligence.