人工智慧透明性義務之法制研究-以電商平台詐欺防制為中心

Abstract

我國電子商務活動於2020年代持續擴張，隨之衍生之詐欺犯罪已成為影響數位經濟秩序之問題。主流電商平台已自人工審查為主之防詐模式轉向以人工智慧（AI）系統為主之自動化治理；惟AI系統之黑箱特性使既有法制於透明性義務之具體化上面臨結構性挑戰。我國雖於 2024 年 7 月公布施行《詐欺犯罪危害防制條例》、於2026年1月14日公布施行《人工智慧基本法》，惟其透明性條款多止於原則宣示，子法訂定工作正進入關鍵 24 個月檢討期，立法建議與實踐途徑之時序急迫性具實務意義。 本論文以三組研究問題為核心：其一，於電商平台應用AI防詐時，《詐欺犯罪危害防制條例》第36條第1項所稱「善良管理人之注意義務」與「合理措施」為何；其二，於電商平台應用AI防詐時，其透明度義務之內涵具體內容為何；其三，針對我國電商平台應用AI防詐應負之透明度義務，相關監理框架應如何設計。 本論文考量技術可行性與規範明確性等面向後，參考國際組織、比較法以及相關技術標準後，提出六項具體立法建議：建議一，人工智慧基本法第 4 條第 5 款之透明度指引，建議由數位發展部依同法第 16 條第 1 項後段以「協助」定位草擬之，並賦予AI評測中心明確定位與角色；建議二，該指引應針對透明度之程度、時點及對象等三面向之具體要件有明文化規定；建議三，制定平台AI透明度專法，設「一般解釋權」與「反對自動化決策權」專章（個人資料保護法修法得為評估之替代途徑），配套規範資料保護影響評估（Data Protection Impact Assessment，DPIA）等具體要件，並參考歐盟Regulation（EU）2019/1150及人工智慧法第86條涵蓋對商業使用者（含3P企業賣家）之透明度規範，填補企業賣家救濟缺口；建議四，制定人工智慧基本法第 17 條高風險 AI 責任歸屬之子法，並由各目的事業主管機關依原有授權訂立各業別之特定AI管理規範；建議五，由公平交易委員會依公平交易法第25條之授權，就數位平台AI不透明決策之適用發布解釋令，擴張處理原則第7點第（四）款「不當利用相對市場優勢地位」至數位平台情境；建議六，建立跨國合規相容性政策，透過AI安全機構（AI Safety Institutes，AISI）國際合作、ISO/IEC 42001認證互認、經濟合作暨發展組織（OECD）與七大工業國組織（G7）政策對話等面向進行。 前揭六項建議將建立於四層立法架構：其一，由人工智慧基本法設定原則性框架與授權主管機關；其二，制定平台AI透明度專法建構對業者之解釋權基礎；其三，由數位發展部制定AI透明度指引作為對業者之非拘束性技術規範參考；其四，各目的事業主管機關依原有授權訂立各業別之特定 AI 管理規範。希冀於人工智慧基本法第18條規定之二年法規檢討期限屆滿前，為我國電商平台AI防詐治理法制建構提供方向性指引建議。

E-commerce activities in Taiwan have continued to expand throughout the 2020s, and the resulting fraud has become a structural challenge to the order of the digital economy. Mainstream e-commerce platforms have shifted from a fraud-prevention model centered on manual review to automated governance driven primarily by artificial intelligence (AI) systems; however, the black-box nature of such systems poses structural challenges to the concretization of transparency obligations under the existing legal framework. Although Taiwan promulgated the Fraud Crime Hazard Prevention Act in July 2024 and the Artificial Intelligence Basic Act on 14 January 2026, its transparency provisions remain at the level of principled declarations. As secondary legislation enters a critical 24-month review period, the timing urgency of legislative recommendations and implementation pathways carries substantive practical significance. This thesis is organized around three sets of research questions: first, how the duty of care of a “good-faith administrator” and the notion of “reasonable measures” under Article 36, paragraph 1 of the Fraud Crime Hazard Prevention Act should be understood in AI-driven fraud prevention contexts; second, how the transparency obligations should be understood in the said context; and third, how to establish the regulatory framework for the transparency obligations of Taiwan’s e-commerce platforms in the said context. After considering aspects of technological feasibility and regulatory certainty and taking reference from international organizations, comparative laws, and technical standards, this thesis proposes six specific legislative recommendations: Recommendation 1-the Ministry of Digital Affairs may issue the transparency guideline under Article 4, Paragraph 5 of the Artificial Intelligence Basic Act per Article 16, Paragraph 1 as a non-binding, yet "supportive" instrument covering technical specifications, together with a clearly defined role for the AI Evaluation Center; Recommendation 2-the said guideline shall specify the specific requirements of the three dimensions of the transparency obligations, degree, timing, and subject matter; Recommendation 3-enacting a special Platform AI Transparency Act with dedicated chapters on the general right to explanation and the right to object to automated decision-making (an amendment to the Personal Data Protection Act as an alternative pathway), supplemented by Data Protection Impact Assessment (DPIA) requirements, and extending transparency obligations to business users (including 3P commercial sellers) by reference to Regulation (EU) 2019/1150 and Article 86 of the EU Artificial Intelligence Act, thereby filling the gap in remedies for commercial sellers; Recommendation 4-relevant competent authorities shall enact the regulation on high-risk AI liability attribution under Article 17 of the Artificial Intelligence Basic Act, with sector-specific AI administration rules; Recommendation 5-the Fair Trade Commission shall issue an interpretive order on the application of Article 25 of the Fair Trade Act to opaque AI decision-making on digital platforms, which extends Item (4) of Point 7 of its Enforcement Guidelines (“improper exploitation of relative market advantage”) to digital-platform scenarios; Recommendation 6-implementing cross-border regulatory harmonization policies through three parallel tracks: international cooperation among AI Safety Institutes (AISI), mutual recognition of ISO/IEC 42001 certifications, and policy dialogues with the OECD and G7. The six recommendations rest upon a "four-layer legislative architecture": (i) the Artificial Intelligence Basic Act establishes the framework and designates competent authorities; (ii) a special Platform AI Transparency Act establishes the right to explanation against operators; (iii) the AI Transparency Guidelines serve as non-binding technical reference standards for operators; and (iv) competent authorities issue sector-specific AI administration rules under their existing statutory mandates. It is anticipated that, prior to the expiration of the two-year regulatory review period under Article 18 of the Artificial Intelligence Basic Act, this thesis provides directional recommendations for the construction of AI-driven fraud prevention governance for Taiwan’s e-commerce platforms.