本研究旨在探討GPT-4解題錯誤分析學習活動是否能夠提升五年級學生解決真實數學問題的能力，並增強他們的數學信心。真實數學問題將數學與生活情境相連，賦予數學學習更深的意義。然而，學生常常難以理解真實數學問題的複雜性以及其冗長的文字敘述，導致數學信心不足。本研究採用準實驗設計，對象為台灣北部某小學的59位五年級學生，實驗組參與了GPT-4解題錯誤分析學習活動，對照組則接受傳統教學。兩組均進行了同儕解釋活動，但實驗組聚焦於討論GPT-4解題錯誤，而對照組則由高成就學生分享解題方案。實驗組進行了登台分享活動，而對照組在小組內分享後由教師講解答案。兩組均使用相同的真實數學問題教材。研究透過解決真實數學問題的測驗與數學信心量表進行量化評估，並輔以半結構式訪談收集質性數據。結果顯示，實驗組在解決真實數學問題上有顯著提升，無論是組內前測與後測比較，還是與對照組後測比較均顯著優於對照組。此外，實驗組中的低成就學生在解決真實數學問題上的進步顯著高於對照組。實驗組中無論是高成就或低成就學生，其數學信心均顯著高於對照組。此研究確認了GPT-4在數學教育中的有效性，為教育工作者和研究人員提供了新的教學策略與研究方向。;Authentic mathematical problems connect mathematics to real-world scenarios, making math learning more meaningful. However, students often find it challenging to comprehend the complexity and extensive textual descriptions of authentic mathematical problems, resulting in a lack of mathematical confidence. This study aims to investigate whether error analysis learning activity of GPT-4 solutions can improve the skill of fifth-grade students to solve authentic mathematical problems and foster their mathematical confidence. A quasi-experimental design was employed, involving 59 fifth-grade students from a primary school in northern Taiwan. Both the experimental group, which participated in GPT-4-based error analysis learning activities, and the control group, which received traditional instruction, engaged in peer explanation activities. However, the experimental group focused on discussing GPT-4’s errors, while the control group involved high-achieving students sharing their solutions. Additionally, the experimental group conducted staging activities, whereas the control group transitioned directly to teacher-led explanations after group sharing. Both groups also utilized the same authentic mathematical problem materials. Quantitative assessments were conducted through tests on solving authentic mathematical problems and a mathematical confidence scale, complemented by qualitative data collected via semi-structured interviews. The results revealed that the experimental group showed significant improvement in solving authentic mathematical problems, both in pre- and post-test comparisons within the group and in post-test comparisons between groups. Furthermore, the low-achieving students in the experimental group showed a significant improvement in solving authentic mathematical problems compared to the control group. Additionally, the mathematical confidence of both high- and low-achieving students in the experimental group was significantly higher than that of the control group. This study confirms the effectiveness of GPT-4 in mathematics education, offering new teaching strategies and research directions for educators and researchers.
