The Impact of Artificial Intelligence-Powered Chatbots on Self-Directed Learning in Health Sciences Students: A Systematic Review and Meta-Analysis

Abstract

To evaluate the effectiveness of AI-powered chatbots on self-directed learning (SDL; primary outcome), satisfaction, confidence, and critical thinking in health professions students. We systematically searched five databases Library up to July 2026. Two reviewers independently performed study selection, data extraction, and risk-of-bias assessment (RoB 2 for RCTs; ROBINS-I for quasi-experimental studies). Random-effects meta-analyses using REML estimation were performed, with Hartung-Knapp-Sidik-Jonkman confidence intervals and prediction intervals where feasible. Certainty of evidence was assessed using GRADE. Fourteen studies (11 RCTs, 3 quasi-experimental; N=1,182) met inclusion criteria. Of these, 5 studies (365 participants) contributed to the primary meta-analysis of self-directed learning, 5 studies (320 participants) to satisfaction, and 3 studies (227 participants) to confidence. AI chatbots may improve self-directed learning (Hedges’s g=0.85; 95% HKSJ CI: 0.50-1.21; p=0.003; 5 studies; I 2 =29.1%; prediction interval: 0.21-1.50; low certainty). Leave-one-out sensitivity analyses confirmed robustness for self-directed learning (SMD range: 0.75–0.93; all p < 0.05). Chatbots may improve satisfaction (g=1.05; 95% CI: 0.36-1.75; p=0.014; 5 studies; I 2 =75.9%; prediction interval: -0.67 to 2.77; low certainty). The confidence finding was methodologically fragile: the Hartung-Knapp-Sidik-Jonkman (HKSJ) 95% CI crossed the null (-0.37 to 1.98; p=0.098), whereas the conventional Wald-type CI suggested significance (0.27 to 1.35; p=0.003), underscoring the importance of using HKSJ correction with sparse data. Meta-analysis was not performed for critical thinking and self-efficacy due to extreme heterogeneity. AI chatbots may enhance self-directed learning and may improve satisfaction in health professions education. However, the certainty of this evidence is low due to indirectness (variability in SDL measurement tools) and imprecision (sample size below optimal information size). These findings support chatbot integration into pedagogically sound designs, though future research requires longer-term evaluations and standardized measures. In particular, future studies should adopt validated, multi-item SDL scales (e.g., Self-Directed Learning Scale, SRSSDL) to enhance comparability across studies.