生成式人工智能训练数据的治理路径  

A Governance Path for Generative AI Training Data

作　　者：陆瑶 LU Yao(School of Law,Renmin University of China,Beijing 100872,Beijing,China)

机构地区：[1]中国人民大学法学院,北京100872

出　　处：《昆明理工大学学报（社会科学版）》2024年第4期21-29,共9页Journal of Kunming University of Science and Technology(Social Sciences)

摘　　要：机器理解具有不透明性,以治理为进路方可兼顾生成式人工智能发展与安全的两大目标,并以安全为重。现行立法对这一新兴行业的治理策略主要围绕生成式人工智能服务提供者展开。有必要逐个厘清提供者与搜索链接服务提供者等六类法律责任主体类型之间关系,以保持体系化的规范逻辑。就“公私兼重”的双重立法目的而言,训练数据合法性问题的私人执行机制严重不足。对此,应当引入特征库和事中标识两大技术性规则,将安全港规则予以改造并另行规定举证责任倒置,并鼓励行业组织围绕自律公约、科技伦理、信用评价这三大支柱实现硬法与软法相协同的治理。Machine understanding is opaque,and governance is the way forward to balance the dual goals of development and safety of generative AI with a focus on safety.The governance strategy of the current legislation for this emerging industry mainly centers on generative AI service providers.It is necessary to clarify the relationship between the providers and the six types of legal liability subjects,including search link service providers,in order to maintain a systematic logic of regulation.In terms of the dual legislative goal of“the public and the private being equally important”,private enforcement mechanisms for the legality of training data are grossly inadequate.In this regard,two technical rules,namely,the feature library and the identification-in-action rule should be introduced,the safe harbor rule should be reformed and the reversal of the burden of proof should be separately stipulated,and industry organizations should be encouraged to realize the synergistic governance of hard law and soft law around the three pillars of self-regulatory conventions,scientific and technological ethics,and credit evaluation.

关 键 词：生成式人工智能服务提供者 技术性规则 安全港规则 行业自律 数据治理 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] D922.17[自动化与计算机技术—控制科学与工程]