Assessing Generative AI Adoption, Tool Preferences, and Cognitive Reliance Among Medical Students: A Cross-Sectional Study
TL;DR
This is among the first studies to jointly profile students’ tool-specific reliability perceptions and to identify verification behavior as a moderator that buffers familiarity-driven overdependence.
Abstract
Background and Objectives: Generative artificial intelligence (AI) chatbots have entered medical education faster than guidance for their responsible use. Although a rapidly expanding 2024–2026 literature has examined generative AI adoption, attitudes, and AI literacy among healthcare students, comparatively little is known about which specific tools medical students prefer or whether reliance on them carries measurable cognitive risks. We characterized adoption patterns, tool preferences, perceived benefits, and determinants of cognitive overdependence among medical students. Methods: A single-center cross-sectional survey was administered to 141 medical students across preclinical and clinical years at a single institution. A 28-item instrument captured usage patterns, perceived learning benefit, output trust, verification behavior, and AI overdependence risk. Analyses included t-tests, ANOVA, chi-square tests, Pearson correlations, and hierarchical regression. Results: Unless otherwise specified, values are reported as group mean scores on 1–5 Likert agreement scales or as percentages of respondents. ChatGPT was the primary tool for 69.5% of respondents, followed by Claude (12.1%). Daily users reported greater perceived learning benefit than infrequent users (4.14 vs. 3.36; p < 0.001). Clinical students verified AI outputs more often than preclinical students (3.69 vs. 3.21; p < 0.001), while preclinical students showed higher reliance (p = 0.002); verification moderated overdependence risk across academic years (interaction p = 0.041). AI familiarity (β = 0.31) and verification habit (β = −0.22) were the strongest predictors of integration acceptance (R2 = 0.34). Conclusions: Reliance and verification habits diverge by training stage; curricula should pair AI literacy with explicit verification training to mitigate overdependence. As a single-center, self-report study, these findings require multi-center confirmation; nonetheless, to our knowledge, this is among the first studies to jointly profile students’ tool-specific reliability perceptions and to identify verification behavior as a moderator that buffers familiarity-driven overdependence.