近年來，生成式人工智慧廣泛應用於教育場域，如何辨識AI生成文本已成為自然語言處理與AI安全領域的核心議題。然而，現有偵測方法對非母語（L2）寫作者存在系統性偏誤：L2 文本在詞彙多樣性、句法複雜度與流暢度上常與語言模型生成文本分佈相近，易被誤判為AI生成，影響學術評量公平性；此問題在偵測資源仍以英文為主的繁體中文情境下尤為關鍵。本研究旨在降低繁體中文情境下對 L2 寫作者的誤報風險。 本研究提出 GUARD（Generated-text Understanding And Robust Detection）框架，整合 RoBERTa 統計型編碼器與 TAIDE 邏輯型大型語言模型，並透過不確定性感知動態閘控進行晚期融合，以兼顧微觀文體特徵與宏觀語意判讀。資料上，本研究蒐集繁體中文人類語料（含 L2 子群）與多模型 AI 生成文本，並建構字符、句法與語義三層級對抗擾動樣本。 領域內（In-domain）評估中，GUARD 於 L2 子集（n＝258）未產生假陽性；惟此應視為「零觀察誤報」而非真實誤報率為零，其 Wilson 95% 上界約 1.46%，僅為特定繁體中文 L2 語料下的初步公平性證據。動態閘控於衝突樣本正確率達 99.2%，消融亦顯示不確定性校準有助降低過度自信誤判。然而，OOD 測試顯示系統在固定閾值下過於保守，召回率降至 0.1530。雖然模型仍具排序能力（ROC-AUC = 0.9449，AP = 0.9391），但召回率恢復需依賴目標域閾值重校準；因此，本研究將 OOD 結果定位為跨域部署與校準限制，而非直接可部署之效能宣稱。本研究另建構結合直接偏好最佳化（DPO）與檢索增強生成（RAG）之擬人化閉環（Humanizer），於 RAG_DPO 設定下規避成功率達 72.33%，但伴隨可測品質降幅。本研究為繁體中文 AI 文本偵測提供兼具公平性意識與對抗性評估之實證參考。;The proliferation of generative AI has transformed text production in education， challenging academic integrity. Existing AI-generated text (AIGT) detectors often exhibit systematic bias against non-native (L2) writers， whose lower lexical diversity， simpler syntax， and distributional overlap with machine-generated text leave them vulnerable to high false positive rates. This thesis proposes GUARD (Generated-text Understanding And Robust Detection)， a dual-track detection framework specifically designed for Traditional Chinese contexts. GUARD integrates a statistical expert path based on a fine-tuned RoBERTa encoder optimized with Supervised Contrastive Learning (SupCon)， and a logical expert path leveraging TAIDE (a Traditional Chinese Large Language Model) as an LLM-as-a-Judge. The two paths operate in parallel and are merged through an Uncertainty-Aware Dynamic Gating mechanism that performs uncertainty quantification via binary entropy and learns adaptive fusion weights through logistic regression. On in-domain evaluations， GUARD outperforms the baseline AIGCDetector-ZH-v2 across all metrics (Accuracy 0.9932， F1-macro 0.9931)， reducing the False Positive Rate (FPR) from 0.2318 to 0.0040. Notably， GUARD yields zero observed false positives on the L2 subgroup (n = 258; Wilson 95% upper bound ≈ 1.46%) — preliminary fairness evidence under the evaluated Traditional Chinese L2 corpus rather than a true zero false-positive rate. Dynamic gating resolves expert conflicts with 99.2% accuracy on borderline samples. However， in out-of-domain (OOD) settings based on Taiwan Legislative Yuan transcripts， the system becomes highly conservative under the fixed operating threshold， with recall dropping to 0.1530. A subsequent threshold-sensitivity analysis shows that the model still retains ranking ability under OOD conditions (ROC-AUC = 0.9449， AP = 0.9391)， but recall recovery requires target-domain threshold recalibration. Therefore， the OOD results are interpreted as evidence of cross-domain calibration limitations rather than as a directly deployable cross-domain operating point. An adversarial closed-loop system (Humanizer) combining Direct Preference Optimization (DPO) and Retrieval-Augmented Generation (RAG) reaches a 72.33% attack success rate under RAG_DPO， though at measurable quality cost. These findings inform the design of fairness-aware AIGT detection for Traditional Chinese educational contexts.
