Analysis of Generative AI's Diagnostic and Feedback Capabilities Regarding Fraction Multiplication Errors: Focusing on ChatGPT, Gemini, and DeepSeek

This study aims to explore the potential of generative AI in mathematics education by analyzing its ability to diagnose student errors in solving fraction multiplication problems and provide feedback. Drawing on prior research, this study first categorized errors in fraction multiplication into computational errors and word problem errors. Prompts reflecting these error types were input into three generative AI models—ChatGPT-4, Gemini-2.0 Flash, DeepSeek-R1—and the responses from each model were collected for analysis. The results showed that the three generative AI models exhibited complex and inconsistent patterns in error diagnosis overall. In the case of computational errors, all three models demonstrated low diagnostic accuracy, with minimal differences in performance between models. In contrast, for word problem errors, the models showed relatively high diagnostic accuracy, and clear differences in capabilities were observed among the models. In terms of the feedback, the feedback for word problem errors was found to be more appropriate and specific than that for computational errors. This suggests that the strengths and weaknesses of AI models vary depending on the instructional context, and that the educational success of AI depends not on the adoption of the technology itself, but on the professional expertise of teachers who select and guide its use.