Pinning down an octopus: towards an operational definition of AI systems in the EU AI Act
本文分析欧盟AI法案对AI系统的定义过于概念化，不足以有效监管，并提出一个聚焦于决策模型、数据和界面设计三要素的可操作框架，为未来修订提供参考。
This article examines the recently adopted EU AI Act (2024), which lays the groundwork for AI regulation in the EU. We argue that the Act's problem definition for AI systems is mainly conceptual, establishing jurisdictional authority; however, this definition alone is insufficient for effective AI regulation in practice. To contribute to discussions on future amendments to this regulatory framework, we propose an operational framework for the regulation of AI. We argue that defining AI systems should focus on their immutable, rather than mutable components. Our framework decomposes the development and use of AI products into three main components to enable adequate problem structuring: (1) decision models, (2) data, and (3) interface design. Within these components, we identify nine specific issues that warrant focused regulatory attention if we are to uphold the fundamental principles and values derived from the broader EU institutional framework, the values set out in the EU AI Act itself, and the traditional rationales – both social and economic – for regulatory intervention. Comparing the EU AI Act with our proposed framework, our analysis reveals and discusses a number of areas where the current EU framework for the regulation of AI could be amended.