人工智慧應用於保險詐欺偵測之研究--以模型偏誤與演算法公平性為中心

Abstract

人工智慧技術近年於保險業之導入已自輔助行政事務逐步擴及理賠決策核心，其中保險詐欺偵測尤為各國保險業者著力應用之領域。我國保險業者於壽險與產險業務中運用人工智慧之比率已過半，相關技術導入確有提升偵測效率與降低理賠成本之實益。然而，演算法所驅動之詐欺風險評分一旦出現誤判，其所衍生之保戶權益侵害並非傳統理賠審查爭議所能完整涵攝，現行法制對此一新興風險之回應能力，遂成為值得深入檢討之課題。 本文先就人工智慧詐欺偵測之技術運作加以梳理，繼而剖析其所衍生之三項根本性課題。模型偏誤源於訓練資料之歷史偏差、模型設計目標之選擇及部署後之回饋循環，黑箱問題使保戶與監理機關均難以審視決策依據，演算法公平性不可能定理則揭示不同公平性定義之間無法同時被滿足之根本限制。三者交互作用，使誤判風險難以單純仰賴技術進步而獲解決。就現行法制而言，保險法欠缺針對人工智慧介入理賠決策之特別規定，個人資料保護法之資料主體權利於人工智慧情境中適用困難，自律規範則規範密度與強制力俱有不足。 基於上述分析，本文以2026年1月14日施行之人工智慧基本法為規範依據，就高風險應用之風險定性、公平性測試、透明度義務與保戶說明請求權、申訴複核機制及人機協作監督等五個面向，提出具體治理建議，並主張我國應由現行以同業公會自律規範為主體之軟性治理，逐步轉向「主管機關規範主導、自律規範輔助補充」之雙軌格局，期能兼顧技術創新與保戶權益之保障。

Artificial intelligence has moved from peripheral administrative tasks into the core of insurance claims decision-making, with fraud detection now among its most actively pursued applications worldwide. In Taiwan, more than half of all life and non-life insurers have adopted AI systems, and the gains in detection efficiency and cost reduction are not in dispute. What remains unresolved is the legal response to algorithmic misclassification: when a fraud risk score wrongly flags a legitimate claim, the resulting harm to the policyholder falls outside the contours of conventional claims dispute resolution, and it is far from clear that the existing regulatory framework is equipped to address it. This thesis approaches the problem in three steps. After setting out how AI fraud detection systems are built and deployed, it identifies three fundamental issues that recur across them. Model bias stems from historical distortions in training data, from the choices embedded in model design, and from feedback loops that reinforce themselves once the system is in operation. Opacity prevents both policyholders and regulators from meaningfully scrutinizing the basis of any given decision. The impossibility theorem of algorithmic fairness shows, further, that competing definitions of fairness cannot all be satisfied at once. Taken together, these problems are not artifacts of immature technology that further engineering will dissolve. Yet Taiwan's current legal framework offers little traction against them. The Insurance Act contains no provisions tailored to AI-mediated claims decisions; data subject rights under the Personal Data Protection Act translate poorly into the AI context; and the prevailing self-regulatory instruments are deficient in both density and binding force. Drawing on the AI Basic Act, which took effect on January 14, 2026, the thesis proposes a governance framework organized around five elements: classifying fraud detection systems as high-risk applications, mandating fairness testing and the management of sensitive attributes, imposing transparency duties together with a policyholder's right to explanation, strengthening complaint and review mechanisms, and securing the substantive independence of human oversight alongside continuous post-deployment monitoring. The broader argument is that Taiwan should move beyond its current reliance on industry self-regulation toward a dual-track model, in which binding rules issued by the competent authority establish the regulatory floor and self-regulatory norms supply operational detail above it, so as to reconcile technological innovation with the protection of policyholders' interests.