The Chinese Regulation on Generative Artificial Intelligence: A Comparative Study of the Administrative Measures for the Management of GAI Services and Their Alignment with the EU AI Act

当今，诸如人工智能（AI）、生成式人工智能（GAI）及其产出的概念正变得越来越普及。尽管很难为这些概念建立全面且标准化的定义，但以下定义可能较为合适：AI 是一种高度可适应的机器系统，具有不同程度的自主性，能够通过处理接收到的输入来生成输出；而 GAI 作为 AI 的一个子集，则使用深度学习来生成文本、图像和音乐。作为这一领域的两个重要国际角色，中国和欧盟迅速意识到，尽管信息量丰富，但基础概念的清晰度依然不足。因此，为了在发展与安全之间寻求平衡，这两个司法辖区率先制定了相关监管框架，不仅澄清了复杂概念，还提供了解决潜在违规行为和保护公民的机制。 本论文结构分为三个主题章节，旨在澄清与 AI（及 GAI）相关的各种概念，解决不确定性问题，并分析目前已实施的两个监管框架——《欧盟人工智能法案》和《生成式人工智能服务管理办法》的法律和比较方面。本研究揭示了欧盟和中国在该领域的一些有趣的相似点与差异点，并有别于将两者仅仅视为对立世界的一般看法。 第一章从 AI 和 GAI 的概念框架入手，概述其可能的应用以及该领域漫长的构想和发展过程中的关键历史里程碑。随后，对欧盟在这些主题上的方法进行了详细分析，追踪了起草《欧盟人工智能法案》的关键步骤。从风险导向的框架到监管沙盒的引入，该法案的主要特点展示了其如何代表了欧洲乃至全球技术未来的一次重要飞跃。 第二章是本论文的核心，完全聚焦于中国的视角。在全面概述中华人民共和国人工智能发展的历史后，讨论了中国的“垂直-迭代”治理模式，并通过对主要人工智能相关法规的简要分析加以说明。接下来，重点介绍了《生成式人工智能服务管理办法》。通过对草案版本和最终版本的对比，突出了关键的相似点和变化，例如通过更明确的定义框架实现了更精炼的适用范围、某些要求的删除，以及国际合作作为未来发展核心支柱的出现。总体来看，该监管文本在发展与监管之间找到了理想的平衡，充分契合了以垂直治理为主的模式。 第三章建立了对这两个监管文本的比较分析。首先，重点放在与每个司法辖区主要治理模式直接相关的差异上。随后探讨了具有更广泛意义的差异——这些元素独立于不同治理模式而存在，包括截然不同的惩罚框架、不同的监督机制以及对特定人工智能系统的禁止标准的差异。最后，该章节通过对两份文本相似点的研究收尾，强调了双方在创新和国际合作开放性等方面的共同点，为未来的合作和共同发展提供了潜力。

Nowadays, concepts such as Artificial Intelligence (AI), Generative Artificial Intelligence (GAI), and their outputs have become increasingly common. Although it is challenging to establish comprehensive and standardized definitions, the following synthesis may be appropriate: AI refers to a highly adaptable machine system with varying degrees of autonomy that can produce outputs through processing the received inputs, while GAI, a subset of AI, uses deep learning to generate texts, images, and music. The People’s Republic of China and the European Union, two world’s key players in this domain, quickly realized that despite an abundance of information, there is still a lack of foundational clarity. Consequently, in seeking a balance between development and safety, these two jurisdictions became the first to establish regulatory frameworks that clarify complex concepts and provide mechanisms to address potential violations and safeguard citizens. This thesis, structured in three thematic chapters, aims to clarify various concepts related to AI (and GAI), address uncertainties, and analyze the legal and comparative aspects of two regulatory frameworks currently in force, namely the EU AI Act and the Administrative Measures for the Management of Generative Artificial Intelligence Services. This analysis reveals both interesting similarities and discrepancies between the EU and China, distancing itself from the general tendency to consider them solely as opposing worlds. The first chapter begins with a conceptual framework for understanding AI and GAI, outlining possible applications and the key historical milestones in the field’s lengthy process of ideation and development. Subsequently, a detailed analysis is presented on the EU’s approach to these topics, tracing the essential steps leading up to the drafting of the EU AI Act. Key features of this legislative text, which range from a risk-based approach to the inclusion of regulatory sandboxes, illustrate how this Act represents a significant leap forward for Europe’s - and the world’s - technological future. The second chapter, the core of this thesis, is entirely devoted to the Chinese perspective. After a comprehensive overview of the historical development of AI in the People’s Republic of China, the focus shifts to the vertical-iterative governance model, illustrated through a brief examination of major Chinese regulations related to AI. This section is followed by an introduction to the Administrative Measures for the Management of Generative Artificial Intelligence Services. A comparison between the draft and final versions of this text highlights key similarities and changes, such as a more refined scope of action enabled by a clearer definitional framework, the removal of certain requirements, and the emergence of international cooperation as a central pillar for the future. Overall, this regulatory text strikes an ideal balance between development and oversight, fitting well within a predominantly vertical governance model. The third chapter establishes a comparative analysis of the two regulatory texts. Initially, the emphasis is on the difference directly tied to the prevailing governance models in each jurisdiction. This is followed by an exploration of divergences with broader implications - those elements that do not align independently of the differing governance models. These include distinct penalty frameworks, different oversight mechanisms, and varying criteria for prohibiting specific AI systems. The chapter concludes with an examination of similarities between the two texts, highlighting common ground that offers potential for future collaboration and mutual development, such as the emphasis on innovation and openness to international cooperation.