在生成式人工智慧快速發展的背景下，AI 已逐漸改變傳統寫作模式，使寫作方式由「從空白開始撰寫」轉向「以 AI 生成內容為基礎，進行持續修訂與優化」。在此脈絡下，如何運用 AI 支持學習者進行反思、修正與強化思考，而非僅提升寫作效率，已成為高等教育的重要議題。反思雖具有促進學習理解與自我反思之價值，但受限於教師回饋時間與資源，持續提供具深度與個別化之支持仍具挑戰。 本研究開發一套以大型語言模型為核心之「反思回饋系統」。學生於系統外完成反思作業初稿後，將文本上傳至系統，由系統提供 AI 評分、分項回饋與多輪挑戰式對話功能，協助學習者在草稿提交後的修訂歷程中進行檢視、追問與深化反思，並保留教師監督介面以支持教學者掌握學生互動歷程。 本研究採用「個案研究法」，以兩門研究所課程為個案，共 33 名研究生為對象。並採「問卷調查法」評估學生對系統之科技接受度與各功能模組之知覺有用性；並針對學生於系統中累積之 AI 評分結果與互動歷程資料，檢視系統使用對學生批判反思表現之影響。 研究結果顯示，學生對本系統之整體接受度與各功能評價均達高度正向；於 28位實際於系統中完成反思作業之學生中，其批判反思表現隨作業提交次數之增加而呈現上升趨勢；進一步比較學生於課程初期與課程後期之提交所撰寫之反思作業，後期作業於三向度與總分上呈現高於課程初期之趨勢，初步反映學生於持續使用系統之歷程中批判反思表達之變化，顯示學生之批判反思表達於持續使用系統之歷程有所提升。;Generative AI has shifted writing practices from drafting on a blank page to iteratively revising AI-generated content. In this context， leveraging AI to support deep reflection， rather than merely boosting writing efficiency， is a crucial higher education issue. While reflection promotes learning and self-understanding， providing sustained， individualized feedback remains challenging due to limited instructor time and resources. This study developed a Large Language Model (LLM)-based reflective feedback system. Students upload their initial drafts to receive AI scoring， dimension-specific feedback， and engage in multi-turn dialogic challenges. These features help learners review， question， and deepen their reflections during the revision process， while a monitoring interface allows instructors to track student interactions. Adopting a case study approach， this research involved 33 students from two graduate-level courses. Surveys assessed students′ technology acceptance and the perceived usefulness of system modules. Additionally， AI scoring results and interaction logs were analyzed to evaluate the system′s impact on critical reflection performance. Results indicated high overall system acceptance and positive evaluations of its features. Among the 28 students who completed the assignments， critical reflection performance trended upward with increased submissions. Furthermore， comparing early and late-course assignments revealed that later submissions scored consistently higher across all three dimensions and total scores. This suggests that sustained engagement with the system is associated with improvements in students′ critical reflection expression， though causal claims require further investigation with controlled designs.
