本研究旨在開發一套結合生成式人工智慧之迷思概念診斷平台（SciLens），並對其進行初步評估。系統採兩階段制診斷，第一階段以二階層（two-tier）診斷題型，由學生先選答案、再選理由，作答即自動推定其可能持有之迷思；第二階段以生成式人工智慧進行多輪對話追問，以蘇格拉底式提問為基礎，引導學生說出並檢視自己的想法，並追溯其迷思之成因；最後以學生端診斷報告與教師端學生診斷分析儀表板提供診斷回饋。考量人工智慧之判讀並非絕對正確，系統將對話紀錄完整留存供教師複核，並以教師之專業判斷為最終把關，以降低誤判之影響。本研究以國小六年級三個班級共 76 名學生為受測對象，學生以自然科學「水溶液」單元之診斷題組實際使用系統後，以科技接受度模型（TAM）問卷蒐集其知覺有用性、知覺易用性與使用意願，並蒐集其系統建議與回饋。結果顯示，學生對系統之整體知覺有用性（M=4.51）、知覺易用性（M=4.38）與使用意願（M=4.39），以及三項核心功能之有用性與易用性（M 介於 4.51 至 4.84）均高於六點量表之中點；各構面信度良好（Cronbach′s α 介於 .86 至 .96）。開放式回饋顯示系統整體獲得正向評價，並指出診斷報告清晰度、互動趣味性與人工智慧判讀正確性等仍有改善空間，期能作為本系統後續改善與未來研究之參考。;This study developed and conducted an initial evaluation of SciLens， a misconception-diagnosis platform built with generative AI. The system applies a two-stage diagnosis: in the first stage， a two-tier diagnostic item design， in which students first select an answer and then a reason， lets each student’s response automatically infer a likely misconception; in the second stage， a large language model conducts multi-turn follow-up dialogue， based on Socratic questioning， to guide students to articulate and examine their own thinking and to trace the causes of their misconceptions; the results are then provided as diagnostic feedback through a student diagnostic report and a teacher dashboard for analyzing students’ diagnoses. Because AI judgements are not infallible， the system retains complete dialogue logs for teacher review， with the teacher’s professional judgement as the final safeguard against misjudgement. Seventy-six sixth-grade students in three classes used the system with diagnostic items on the natural-science unit of “aqueous solutions”; afterwards， they completed a Technology Acceptance Model (TAM) questionnaire (adapted from Yang (2024)， itself a modification of Yuen & Ma’s (2008) TAM， with items adjusted to this system’s functions) measuring their perceived usefulness， perceived ease of use， and intention to use， together with their suggestions and feedback. Students rated the system’s overall perceived usefulness (M=4.51)， perceived ease of use (M=4.38)， and intention to use (M=4.39)， as well as the usefulness and ease of use of its three core functions (M=4.51 to 4.84)， all above the midpoint of the six-point scale; reliability was good (Cronbach’s alpha between .86 and .96). Students’ open-ended feedback was generally positive and pointed to areas for improvement such as report clarity， interactive engagement， and AI diagnostic accuracy， which can inform future refinement of the system and further research.
