JMIR Publications

ABSTRACT

Background:

Early clinical exposure provides medical students with experiences that require reflection and interpretation. Generative artificial intelligence (GenAI) may support these processes; however, its effects on the underlying learning process remain unclear.

Objective:

This study examined how a NotebookLM-supported, source-grounded GenAI approach shaped students’ learning processes after early clinical exposure, with a focus on students’ perceptions, information-handling practices, and final presentation quality.

Methods:

We conducted a convergent mixed-methods study of a third-year early clinical exposure course at a Japanese medical school. Of 117 enrolled students, 98 completed matched precourse and postcourse survey responses. After peer reflection, students used NotebookLM-supported GenAI to organize and integrate their clinical experiences with related information. We analyzed Likert-scale responses and free-text comments using quantitative analysis and qualitative content analysis, respectively. Three faculty members rated slide decks from the 2023 cohort before GenAI integration (n=15) and the 2025 cohort after GenAI integration (n=14) using a 7-item instrument. Findings were integrated using a joint display.

Results:

Most students (85/98, 86.7%) had never used NotebookLM. Confidence in organizing a topic into slides increased from a mean of 3.3 (SD 1.0) to 3.7 (SD 0.9), and self-rated ability to plan a slide structure increased from 3.0 (SD 1.1) to 3.5 (SD 1.0; both P<.01). The other five paired items showed no significant changes. After the program, 82.7% of students reported that NotebookLM helped them understand and identify key points, 80.6% reported reduced workload, 78.6% reported improved slide-preparation efficiency, and 91.8% agreed that AI outputs required critical evaluations. Qualitative findings indicated that students used GenAI-supported initial structures as starting points for review and revision and that perceived efficiency gains allowed more time for reflection and content refinement. Joint display integration showed convergence between improved self-evaluations and reports of AI-supported organization and revision and between perceived efficiency gains and redistribution of time toward deeper learning. However, these perceived changes were only partly reflected in the final products: only the use of visual materials was rated higher in 2025 than in 2023 (mean 4.1, SD 0.4 vs mean 3.8, SD 0.3; P=.04), whereas evidence-support ratings were unchanged.

Conclusions:

A NotebookLM-supported, source-grounded GenAI approach may provide temporary scaffolding for organizing and integrating experiences and information after early clinical exposure. Students perceived gains in efficiency and changes in how they reviewed and reconstructed information; however, these changes did not correspond to broad improvements in presentation quality. Therefore, NotebookLM-supported learning should retain peer dialogue, critical appraisal, source verification, and reconstruction as explicit learning activities.