Enhancing Clinical Reasoning Development: An AI-Assisted Learning Framework for Unstructured Problem-Solving in Medical Education

Abstract Background Clinical reasoning for unstructured problems—those lacking clearly defined initial conditions, standard protocols, or singular correct solutions—remains a persistent challenge in medical education. While artificial intelligence (AI) tools have shown promise in structured knowledge tasks, their integration into frameworks designed for unstructured clinical reasoning remains underexplored. Methods A systems thinking assessment framework was adapted for medical education and validated through expert review and inter-rater reliability testing (ICC = 0.801, 95% CI: 0.725–0.866). Fifty-five medical students were randomly assigned to three groups: an AI-assisted learning group (AIL, n = 18) using ChatGPT 4.0 as a guided reasoning aid, a conventional learning group (CL, n = 18) using textbooks and non-AI online resources, and a manual group (MG, n = 19) without any supplementary tools. All participants responded to the same unstructured clinical case comprising seven tasks (Q1–Q7), each scored on a 0–3 scale by five blinded raters. As an exploratory benchmark, ChatGPT 4.0 responses to the identical case were independently scored using the same framework to characterize the AI model's performance profile and limitations. Results The AIL group achieved the highest mean total score (12.11 ± 1.78), significantly outperforming the CL group (8.94 ± 1.08, p < 0.0001) and the MG (9.97 ± 2.37, p < 0.01). Significant advantages were observed in Q2 (information needs), Q5 (scenario development), Q6 (challenges), and Q7 (unintended consequences). Subgroup analysis revealed no significant interaction between learning tools and baseline academic performance (F = 0.213, p = 0.930, η² = 0.018), suggesting consistent benefits across GPA levels. The ChatGPT 4.0 benchmark produced uniformly identical scores (total score 16.5 across all trials), revealing characteristic limitations in response variability and template dependency. Conclusions An AI-assisted learning framework grounded in a validated systems thinking assessment tool can enhance medical students' clinical reasoning for unstructured problems, with effects that appear independent of baseline academic performance. The ChatGPT benchmark highlights both the capabilities and the rigidities of current AI models, underscoring the need for hybrid human–AI collaboration rather than AI substitution in clinical reasoning education. Trial registration: Chinese Clinical Trial Registry ChiCTR2600133444, 2026-09-28, Retrospectively registered