欧盟生成式人工智能立法实践及镜鉴

Legislative Practice of Generative Artificial Intelligence in EU and Its Implication

生成式人工智能利用海量未标记数据和合成数据进行持续训练,依赖深度神经网络等机器学习技术逐渐形成自主的行为能力,输出新颖结果、应用日趋广泛,正深刻改变着人际间的互动方式,其模型开发的资源密集型特性也促使复杂价值链条形成。生成式人工智能在运行节点的技术跃迁,引发了版权侵权、数据偏见、能耗过大、风险难测、虚假信息传播以及损害认定困难等监管挑战。欧盟人工智能法作出紧急回应,以“通用人工智能模型”为概念中枢,经由“通用人工智能系统”过渡,将生成式人工智能纳入“人工智能系统”范畴;输入端从数据数量和数据质量双管齐下设置合规义务,处理端引入“高影响能力”的自主性程度判断标准,并将“具有系统性风险的人工智能”嵌入风险分类分级制度,输出端则设计“检测、披露和透明度”等义务来规制虚假信息传播,部署端也专门设计价值链上的责任分配专条。虽然欧盟立法为应对生成式人工智能风险作出了努力,但在“抽象定义的确定性”“衡量数据训练效果的方法”“高级模型与小型模型之区分”“系统性损害的确定”以及“API接口和开源模式对价值分配的影响”等方面仍有继续完善的空间。

Generative AI uses massive unlabeled data and synthetic data for continuous training,relies on machine learning technologies such as deep neural networks to gradually form autonomous behavioral capabilities,output novel results,and become more and more widely used,which is profoundly changing the way people interact with each other,and the resource-intensive nature of model development is also promoting the formation of complex value chains.The technological leap of generative AI in the operation node has raised regulatory challenges such as copyright infringement,data bias,excessive energy consumption,unpredictable risks,disinformation dissemination,and difficulty in determining damages.The EU AIA responds urgently by bringing generative AI into the category of'AI systems'through the transition of'AI general systems',with'general AI models'as the conceptual center;the inputting side sets compliance obligations based on both data quantity and data quality;the processing side introduces the criterion for judging the degree of autonomy of'high impact ability',and embeds'artificial intelligence with systemic risk'into the risk classification and grading system;obligations such as'detection,disclosure and transparency'are designed to regulate the spread of disinformation on the output side;the deployment side also specifically designs a dedicated article for the allocation of responsibilities along the value chain.Although EU legislation has made efforts to address the risks of generative AI,there is still room for further improvement in such areas as'certainty of abstract definition','methods for measuring the effectiveness of data training','distinction between advanced and small models','determination of systemic damage',and'impact of API interfaces and open source models on value distribution'.