本研究旨在探討不同生成式人工智慧（GAI）引導模式融入社會性科學議題（SSI）教學後，對國小五年級學生非形式推理、GAI論證及反思表現之影響，並分析驗證信念與各項學習表現之關係。本研究採準實驗之不等組前後測設計，以某一國民小學五年級兩班共60名學生為研究對象，分別接受Belief mode與Design mode引導。研究資料包括驗證信念量表、非形式推理前後測、GAI論證對話及反思紀錄，並以成對樣本t檢定、共變數分析、多元回歸分析及變異數分析進行處理。主要研究結果如下： 一、兩種GAI引導模式皆能促進部分非形式推理表現，其中，接受Design mode引導之組別進步指標較多。而控制前測差異後，第一循環中Design mode組別的論點 總數較高；第二循環中，其反方論點數與推理面向總數亦較高。 二、控制前測表現、GAI引導模式及其他驗證信念向度後，三項驗證信念均未能顯著預測兩個循環之非形式推理後測表現，GAI引導模式亦未顯著調節兩者關係；相 較之下，學生原有推理表現，尤其在第二循環中，較能預測後續表現。 三、不同驗證信念傾向學生在多數GAI論證表現上未呈現顯著差異。Belief mode組別於第一循環的證據支持比例及第二循環的反方回應比例較高；第一循環另出現GAI 引導模式與驗證信念在觀點整合比例上的交互作用。 四、反思表現方面，第一循環中Design mode組別在立場與原有想法回顧表現上較 佳；第二循環中，多重來源傾向學生的AI影響覺察高於個人來源傾向學生，但納入GAI引導模式後未達顯著。 由本研究產出的結果可知，不同GAI引導模式的影響具有選擇性與情境性，較易促進特定推理、論證或反思行為，而未對所有指標產生全面且一致的效果。而驗證信 念可作為了解學生知識判斷傾向的參考，但尚不足以單獨預測其學習表現。;This study examined the effects of two generative artificial intelligence (GAI) guidance modes on fifth-grade students′ informal reasoning， argumentation， and reflection in socioscientific issues (SSI) instruction. A quasi-experimental nonequivalent pretest–posttest design was adopted with 60 students from two intact classes. The groups received either Belief mode or Design mode guidance. Data included a justification beliefs questionnaire， informal reasoning pre- and posttests， GAI argumentation dialogues， and reflection records. Results showed that both modes improved some aspects of informal reasoning， while the Design mode group improved across more indicators. After controlling for pretest differences， the Design mode group produced more arguments in the first cycle and more counterarguments and reasoning dimensions in the second cycle. After controlling for corresponding pretest performance， GAI guidance mode， and the other justification belief dimensions， none of the three justification belief dimensions significantly predicted informal reasoning posttest performance in either cycle. Moreover， no significant moderating effects of GAI guidance mode on the relationship between justification beliefs and informal reasoning performance were found. Most argumentation indicators did not differ by justification belief orientation. The Belief mode group showed higher evidence-support and counterargument-response proportions， whereas the Design mode group performed better in reviewing initial positions. A significant interaction between guidance mode and justification beliefs was found for viewpoint integration in the first cycle. According to the results， the effects of GAI guidance were selective and context dependent. Justification beliefs helped explain students′ knowledge-evaluation tendencies but did not independently predict learning performance.
