Dechao Lu , Jingbin Xue , Yuyang Qi , Wangyi Xu *
Hangzhou Normal University, Hangzhou, Zhejiang , China
*Authors to whom correspondence should be addressed.
Received: 2026-1-12 / Accepted: 2026-4-6 / Published: 2026-4-20
DOI: https://doi.org/10.37906/real.2026.2
Abstract This study investigates the performance of ChatGPT-5 in solving physics conceptual understanding and reasoning problems, with a focus on its implications for education. The research employed two widely used assessment tools, the Force Concept Inventory (FCI) and the Conceptual Survey of Electricity and Magnetism (CSEM), to evaluate ChatGPT-5's conceptual understanding, including both text-based and image-based questions. Additionally, eight primitive physics problems were used to assess its reasoning ability based on self-organizing representation theory. Results show that ChatGPT-5 achieved nearly perfect accuracy (100%) on text-based problems and substantial improvement (71.7%) on image-based items compared to earlier versions. In reasoning tasks, ChatGPT-5 reached an overall accuracy rate of 87.5%, outperforming both students and ChatGPT-4, particularly in abstraction, methodology, and physical representation dimensions. However, persistent weaknesses were identified in mathematical representation and in handling complex visual reasoning. These findings highlight ChatGPT-5's potential as a powerful educational tool for conceptual learning support and misconception clarification, while also revealing its current limitations in multimodal reasoning. The study underscores the necessity of integrating AI tools with teacher guidance to enhance students' comprehensive physics competencies.
Research Areas: Scientific reasoning & problem solving, Technology
Keywords: ChatGPT-5, problem solving, conceptual understanding, scientific reasoning
