論人工智慧之透明度監理法制—以銀行授信之應用為中心

Abstract

人工智慧技術之發展改變了人們的生活，亦為各行各業帶來前所未有之機遇與挑戰，而銀行授信作業作為最早導入人工智慧的領域之一，藉此迎來了服務效率與準確度之全面提升，惟同時亦須面臨因演算法不透明所衍生之黑盒子（Black Box）風險。透明度監理之目的即在於使利害關係人，包含監理機關與終端消費者等，能夠了解人工智慧運作之邏輯，以及人工智慧所做成的決策將帶來什麼樣的影響，藉此達到控制人工智慧風險以及保護消費者之效果。然而，在法制仍存在許多模糊與缺漏之狀態下，應如何調整以引導金融機構回應不同利害關係人對於透明度之需求與保障，即為本文主要之研究問題。 本文從專家透明度與個體透明度之雙重切入角度出發，並透過對我國現行法制之觀察得知，我國法制雖已初具透明度監理之雛形，惟仍存在許多可強化之處，例如揭露內容要求模糊、自我評估機制易生利益衝突、缺乏對消費者之保障機制等，使其難以滿足消費者對人工智慧透明度之要求。經由對美國與歐盟法制之探討，尤其著重於其監理思維與規範模式之差異，以及所對應之具體手段，同時彙整相關學術理論與實務作法，本文認為應分別從專家透明度與個體透明度角度重新思考我國法制，就前者，應加強法定資訊揭露義務，引進獨立第三人審查機制，並考量是否需要在整體規範架構層面上加強監理機關之權限；就後者，則應在現有法制框架上重構事前資訊揭露之內容，並建立事後保障機制，以保障消費者異議與救濟之權利。

The development of artificial intelligence (hereinafter “AI”) technologies has significantly transformed modern life and brought unprecedented opportunities and challenges across various sectors. In particular, credit granting operations in the banking industry—one of the earliest adopters of AI—have benefited from enhanced efficiency and accuracy. However, this transformation also introduces the inherent risk of algorithmic opacity, commonly referred to as the “black box” problem. The objective of transparency regulation is to enable stakeholders, including regulatory authorities and end users, to understand the logic underlying the operations of AI and the potential impact of AI-driven decisions, thereby facilitating effective risk control and ensuring consumer protection. In light of persistent ambiguities and gaps in the existing legal framework, this thesis seeks to explore how regulatory systems may be adjusted to ensure that financial institutions respond appropriately to the diverse transparency needs and interests of different stakeholders. This thesis adopts a dual-perspective approach by analyzing both expert transparency and individual transparency. Through a detailed examination of the current legal framework in Taiwan, the thesis finds that, although the regulatory structure for AI transparency has begun to take shape, substantial deficiencies remain. These include vague disclosure requirements, potential conflicts of interest in self-assessment mechanisms, and a lack of protective measures for consumers, all of which hinder the fulfillment of transparency expectations. By conducting a comparative analysis of the regulatory approaches in the United States and the European Union, especially highlighting their divergent supervisory philosophies and regulatory techniques, and synthesizing relevant academic theories and practical measures, the study proposes a framework for rethinking Taiwan’s legal system in light of the two transparency dimensions. With respect to expert transparency, the thesis recommends enhancing statutory disclosure obligations, introducing independent third-party audit mechanisms, and considering the expansion of regulatory authority at a structural level. For individual transparency, the thesis suggests reconstructing ex ante disclosure requirements and establishing ex post redress mechanisms, in order to safeguard consumers’ rights to contest and understand AI-based credit decisions.