生成式人工智能训练数据集的法律风险与包容审慎规制

Legal Risks of Generative AI Training Datasets and Inclusive Prudential Regulation

训练数据集是生成式人工智能的基础,亦是一种聚合型权益客体,可以成为规制生成式人工智能的重要切入点。训练数据集的开发存在未经许可使用作品、违法处理个人信息等法律风险。然而,现有规制框架却呈现出风险防范形式化的现象,许可使用、合理使用等著作权法律规则难以因应规模化的训练数据集,而个人同意、目的限制与数据最小化等个人信息保护机制及原则亦存在效能不彰的问题。为此,有必要重塑现有的规制框架,将规制目标从权利保护进阶为公平使用,将规制路径由命令控制转变为包容审慎。为了实现生成式人工智能训练数据集的包容审慎规制,一是要构建包容审慎的著作权法规制,包括重构合理使用的标准与范围、构建著作权临时许可裁定制度、发布著作权合规最佳实践指南;二是要迈向包容审慎的个人信息保护法规制,包括拓展个人信息处理的合法性基础、改造告知同意机制、实施个人信息保护影响评估。

Training datasets serve as the backbone of generative artificial intelligence and a subject of aggregated rights,representing a crucial focal point for regulating its development.However,the creation and utilization of these datasets pose legal challenges,such as potential copyright infringement and improper handling of personal information.Nevertheless,the current regulatory framework demonstrates a tendency towards formalization in risk prevention,posing challenges for copyright laws to adequately address in terms of fair use and reasonable use the scale of training datasets.Furthermore,mechanisms and principles for protecting personal information,such as individual consent,purpose limitation,and data minimization,also exhibit limited efficacy.To enhance the regulation of generative AI training datasets,a shift is needed from a focus on rights protection to embracing fair use and from a rigid command-and-control approach to a more inclusive and cautious strategy.To achieve this,it is essential to reform existing legal frameworks by promoting fair use principles in copyright laws,establishing a system for temporary copyright licenses,and providing guidance on copyright security practices.Additionally,a more inclusive and cautious approach to personal information protection law is necessary,including expanding the legal basis for personal data processing,updating notification and consent mechanism,and implementing impact assessment for personal information protection.