本研究旨在探討不同角色互動模式之 AI 聊天機器人對使用者情緒的影響，包括（1）由 Chatbot 協助使用者與由使用者協助 Chatbot 的角色差異，是否會影響使用者的焦慮程度、正向情感與負向情感；（2）在 Chatbot 協助使用者的情境中，朋友式與積極式互動風格是否會產生不同的情緒調節效果；（3）在使用者協助 Chatbot 的情境中，模擬對話情境與文章呈現方式是否會影響使用者的情緒反應與投入程度；（4）評論與分數及只有分數兩種回饋方式，是否會對使用者的焦慮程度、正向情感與負向情感產生不同影響；以及（5）運用腦波技術，探討使用者在實驗前、Chatbot 互動中及實驗後的腦波變化，以補足傳統問卷測量在即時情緒與認知反應上的侷限。 本研究採用實驗法，受試者於實驗過程中配戴腦波儀，依序完成前測問卷、AI Chatbot 互動任務及後測問卷。正式實驗設計包含三組比較，分別為1（由 Chatbot 協助使用者）×2互動風格（積極式 vs. 朋友式）、1（使用者協助 Chatbot）×2呈現方式（模擬對話情境 vs. 文章呈現），以及1（使用者協助 Chatbot）×2回饋方式（評論與分數 vs. 只有分數），共形成六種實驗情境。受試者分配至其中一組進行互動，並透過正負向情感量表（PANAS）、廣泛性焦慮量表（GAD-7）及人機互動舒適度量表衡量其主觀情緒反應。腦波資料則依實驗流程區分為實驗前（S1）、互動中（S2）及實驗後（S3）三個階段。本研究總計取得146份有效樣本，資料分析分為問卷分析、腦波資料分析，以及問卷與腦波結果的綜合比較，並運用敘述性統計、多變量變異數分析、配對樣本 t 檢定、組間變異數分析及重複量數變異數分析進行檢驗。 本研究結果顯示，（1）不同實驗組別在整體結果變項組合上具有顯著差異，表示不同角色定位、互動風格、呈現方式與回饋機制確實可能造成不同的心理與互動反應；（2）朋友組在焦慮程度、正向情感與負向情感方面皆呈現顯著改善，為六組中情緒改善效果最完整的互動模式；（3）積極組的焦慮程度與負向情感顯著降低，正向情感則呈現提升趨勢，顯示積極式回應較能協助使用者降低負向情緒；（4）情境組與評論及分數組的正向情感顯著提升，表示具體互動情境及帶有語意內容的回饋，有助於提高使用者的參與感與正向感受；（5）文章組與只有分數組在焦慮程度、正向情感及負向情感方面皆未達顯著差異，顯示單向閱讀或僅提供量化分數，較難形成穩定的情緒調節效果；（6）腦波結果顯示，部分實驗組在 High Beta、Low Gamma、Mid Gamma、Low Alpha 及 Low Beta 等頻段出現顯著變化，且部分高頻腦波呈現互動中下降、互動後回升的趨勢，可能反映受試者在不同互動階段中的注意力、資訊處理與認知整合活動有所改變。 本研究之研究貢獻如下：（1）將情緒支持型 Chatbot 的研究由傳統的「AI 協助使用者」延伸至「使用者協助 AI」的角色反轉情境，補充過去研究對人機角色關係探討之不足；（2）結果顯示朋友式互動較能同時降低焦慮與負向情感並提升正向情感，說明社會臨場感、陪伴感與關係性回應在情緒支持型 Chatbot 中的重要性；（3）發現使用者協助 Chatbot 並不必然產生情緒改善，其效果仍取決於互動情境是否具體、使用者是否具有參與感，以及回饋是否具有意義；（4）透過評論與分數組及只有分數組的比較，指出回饋效果不只取決於是否提供評價，更與回饋是否具有語意說明及情感回應有關；（5）結合主觀問卷與客觀腦波資料，從情緒感受與生理反應兩個層面驗證不同 Chatbot 互動模式的效果，並顯示腦波變化未必直接等同於情緒改善，而可能同時反映注意力與認知處理歷程；（6）本研究結果可作為未來情感智能型 Chatbot 進行角色設定、語氣設計、情境呈現及回饋機制規劃之實證參考。;This study investigates how AI chatbots employing different role-based interaction modes influence users’ emotions. Specifically， it examines: (1) whether assigning the chatbot or the user as the helper affects users’ anxiety， positive affect， and negative affect; (2) whether friend-style and encouraging-style interactions produce different emotion-regulation effects when the chatbot assists the user; (3) whether scenario-based dialogue and article-based presentation influence users’ emotional responses and engagement differently when users assist the chatbot; (4) whether two feedback formats—comments with scores and scores only—have different effects on users’ anxiety， positive affect， and negative affect; and (5) how users’ electroencephalographic activity changes before， during， and after interacting with the chatbot. EEG measurements were incorporated to address the limitations of conventional questionnaires in capturing immediate emotional and cognitive responses. An experimental method was adopted. Participants wore an EEG device while completing a pretest questionnaire， an AI chatbot interaction task， and a posttest questionnaire. The experimental design consisted of three comparisons: one chatbot-assists-user condition × two interaction styles (encouraging vs. friend-style)， one user-assists-chatbot condition × two presentation formats (scenario-based dialogue vs. article presentation)， and one user-assists-chatbot condition × two feedback formats (comments with scores vs. scores only). These comparisons resulted in six experimental conditions， with each participant assigned to one condition. Subjective emotional responses were measured using the Positive and Negative Affect Schedule (PANAS)， the Generalized Anxiety Disorder Scale (GAD-7)， and a human–computer interaction comfort scale. EEG data were divided into three stages: before the experiment (S1)， during the interaction (S2)， and after the experiment (S3). A total of 146 valid cases were obtained. The analysis comprised questionnaire analysis， EEG analysis， and an integrated comparison of the questionnaire and EEG findings. Descriptive statistics， multivariate analysis of variance， paired-samples t-tests， between-group analysis of variance， and repeated-measures analysis of variance were employed. The results indicate that: (1) the experimental groups differed significantly in the overall combination of outcome variables， suggesting that role assignment， interaction style， presentation format， and feedback mechanism may generate different psychological and interaction responses; (2) the friend-style group showed significant improvements in anxiety， positive affect， and negative affect， representing the most comprehensive emotional improvement among the six conditions; (3) the encouraging-style group exhibited significant reductions in anxiety and negative affect， while positive affect showed an upward trend， indicating that encouraging responses were particularly effective in alleviating negative emotions; (4) positive affect increased significantly in both the scenario-based group and the comments-with-scores group， suggesting that concrete interaction scenarios and semantically meaningful feedback can enhance engagement and positive feelings; (5) no significant changes in anxiety， positive affect， or negative affect were observed in the article-based or scores-only groups， indicating that one-way reading or quantitative scores alone may be insufficient to produce stable emotion-regulation effects; and (6) significant changes were observed in High Beta， Low Gamma， Mid Gamma， Low Alpha， and Low Beta activity in some experimental groups. Certain high-frequency EEG bands decreased during the interaction and subsequently rebounded after the interaction， possibly reflecting changes in attention， information processing， and cognitive integration across the different interaction stages. This study makes the following contributions: (1) it extends research on emotionally supportive chatbots from the conventional chatbot-assists-user model to a role-reversal model in which users assist the AI， thereby addressing the limited attention previously given to role relationships in human–AI interaction; (2) it demonstrates that friend-style interaction is more effective in simultaneously reducing anxiety and negative affect while increasing positive affect， highlighting the importance of social presence， companionship， and relational responses in emotionally supportive chatbots; (3) it shows that assisting a chatbot does not necessarily improve users’ emotions， as its effectiveness depends on whether the interaction scenario is concrete， whether users experience a sense of engagement， and whether the feedback is meaningful; (4) by comparing comments with scores against scores only， it shows that the effectiveness of feedback depends not merely on the presence of an evaluation but also on whether the feedback provides semantic explanations and emotional responses; (5) by integrating subjective questionnaires with objective EEG data， the study evaluates different chatbot interaction modes from both emotional and physiological perspectives and demonstrates that EEG changes do not necessarily represent emotional improvement directly， but may also reflect attentional and cognitive processing; and (6) the findings provide empirical guidance for the design of roles， conversational styles， scenario presentations， and feedback mechanisms in future emotionally intelligent chatbots.
