Integrating generative AI into cognitive apprenticeship–based problem-based learning in biochemistry: a quasi-experimental study

Abstract Background Problem-Based Learning (PBL) is extensively employed in medical education to promote advanced cognitive skills. At the foundational level of medical education, novice learners frequently encounter difficulties in forming cohesive knowledge structures due to an excessive cognitive load. Generative artificial intelligence (GenAI) offers new opportunities for learning support, yet its pedagogical role and integration mechanisms in medical education remain insufficiently explored. Methods This study developed an AI-supported problem-based learning model grounded in the cognitive apprenticeship framework (AI-CAS-PBL). Using the “Lipid Metabolism” unit in a biochemistry course, a quasi-experimental design was used to compare AI-CAS-PBL with conventional PBL and lecture-based learning (LBL). The same instructor taught all three groups. Learning objectives, core content, scheduled contact time, case materials, and research outcome assessment procedures were standardized across groups, whereas instructional approaches, learning support, and out-of-class activities differed.The AI learning agent was developed on the Coze platform and powered by the Doubao 1.5 Pro 32 K large language model. Learning outcomes included lipid metabolism knowledge and case-based clinical reasoning. Technology acceptance, perceived cognitive workload measured using the NASA Task Load Index (NASA-TLX), self-efficacy, and cognitive apprenticeship experience were also assessed. Results Students in the AI-CAS-PBL group achieved significantly higher post-test lipid metabolism knowledge and clinical application analysis scores than both the conventional PBL and LBL groups (all P < 0.05). Students reported high perceived usefulness and perceived ease of use of the AI learning agent. Compared with the conventional PBL group, students in the AI-CAS-PBL group reported lower perceived cognitive workload and higher self-efficacy. Conclusion The AI-CAS-PBL group demonstrated higher post-intervention lipid metabolism knowledge and case-based clinical reasoning scores than the conventional PBL and LBL groups. Compared with conventional PBL, students in the AI-CAS-PBL group reported lower perceived cognitive workload and higher self-efficacy. These findings are consistent with the possibility that AI-supported questioning and instructor modeling may provide complementary forms of instructional support. However, because the interaction between AI and instructor scaffolding was not directly examined, the proposed dual-scaffolding framework should be regarded as a theoretically informed interpretation rather than a confirmed causal pathway. The model represents a promising approach to human–AI collaborative teaching that warrants further evaluation in larger, multicenter, and multi-class studies.