从算法到模型，人工智能正跨越一个分水岭——从依靠既定规则的智能，走向自我学习进化的智能。同 步，一个全新的治理命题浮现：如何在推动创新的同时，确保模型的安全与可控？近期，围绕模型治理 ——特别是通用、前沿大模型的治理，中美欧交出不同答卷。2025年7月，欧盟发布自愿性指导性文件 《通用人工智能实践准则》（以 下简称《 实践准则》），旨在帮助模型提供者遵守 《人工智能法案》 中有关通用人工智能模型的义务规范。 [1] 9月，美国加州州长签署《前沿人工智能透明法案》 （SB 53） ，聚焦于"前沿模型" （Frontier Models） 的透明度义务。 [2] 同月，中国发布《人工智能安全治理 框架2.0》，尽管并非针对模型，但作为指导性政策文件，其明确了模型层面的风险，并提出了相应的 风险应对指引。 [3] 如何治理模型：欧盟、美国加州与中国的探索 中美欧在模型治理上形成了三种差异化路径：欧盟构建层层叠叠的风险类别，配套高密度的义务；美国 以加州为代表，选择小范围、轻监管模式，强调企业自律；中国则从应用场景出发，通过"自下而上"的 方式延伸至模型本身。理解以上路径的共性与差异，有助于厘清模型治理在人工智能治理 ...

Core Viewpoint - The article discusses the evolving landscape of artificial intelligence governance, particularly focusing on the governance of general-purpose and frontier models in the US, EU, and China, highlighting their distinct approaches and regulatory frameworks [2][10]. Group 1: EU Governance Approach - The EU has established a complex risk governance framework categorizing AI systems into four risk levels: prohibited, high-risk, limited-risk, and minimal-risk, with stricter regulations for higher-risk categories [4]. - The EU's governance mechanism for general models distinguishes between those with and without "systemic risk," requiring all providers to disclose technical documentation and training summaries, while those with systemic risk must undergo model assessments and report significant incidents [5]. - The EU's framework is characterized by overlapping standards for models and applications, leading to a burdensome regulatory environment that may hinder innovation, prompting the EU Commission to push for simplification of related regulations [6]. Group 2: US Governance Approach - California has adopted a lighter regulatory approach with the signing of the "Frontier AI Transparency Act" (SB 53), focusing on self-regulation and limiting the scope of obligations for model developers [6]. - SB 53 targets "frontier developers" using models with over 10^26 FLOPs, with additional criteria for larger developers, thus narrowing the regulatory scope compared to the EU's broader approach [6]. - The obligations under SB 53 are minimal, primarily requiring basic transparency regarding website information and intended use, contrasting sharply with the EU's extensive documentation requirements [6]. Group 3: China's Governance Approach - China's governance strategy is application-driven, focusing on real-world issues and extending regulations from application services to model governance [7][8]. - The country has established a regulatory framework for algorithm governance, which has laid the groundwork for model governance, addressing risks associated with algorithmic recommendations and deep synthesis technologies [8]. - China's governance framework emphasizes practical measures for risk identification and management, categorizing risks into endogenous, application, and derivative risks, thus providing a clear delineation of responsibilities [9]. Group 4: Commonalities and Future Directions - Despite differing backgrounds and regulatory obligations, the US, EU, and China share a tendency towards "flexible governance" and industry-led initiatives, allowing for greater compliance autonomy [11]. - All three regions are exploring the establishment of assessment ecosystems to address uncertainties in model capabilities, with suggestions for community-driven evaluation mechanisms [11]. - Transparency has emerged as a core governance tool across the three regions, facilitating maximum control with minimal constraints, thereby fostering innovation while ensuring accountability [12].