论生成式人工智能版权侵权“双阶”避风港规则的构建

Establishing a"Two-Tier"Safe Harbor Rule for Copyright Infringement in Generative AI

生成式人工智能的作品使用具有海量化及算法化特征,面临侵权责任认定与分配难题。避风港规则以满足特定条件给予免责为构造,可以回应机器训练的行为转变,契合风险分配的规制目标,具有事前预防效果。基于生成式人工智能服务提供者的技术能力、大模型版权侵权的规制需求和版权人的获益需求,为其新设“双阶”避风港规则具有必要性。在训练阶段,可以通过设置信息披露、权利保留的识别尊重、非直接获得经济利益和整体性补偿义务,使生成式人工智能服务提供者无须经事先许可使用作品,且不必承担解除学习等责任;在输出阶段,可以为生成式人工智能服务提供者配置建立投诉处理机制、消除重复作品数据、优化模型过度拟合、干扰用户恶意引导、基于请求的版权过滤等义务,使其免受抽象侵权标准影响。

The utilization of works by generative artificial intelligence(AI)has two distinctive features:massive scale and algorithmic processes.These features present substantial challenges in analyzing and assigning copyright infringement liability.The safe harbor rules,based on meeting certain conditions,offer immunity that can address the structural shift to machine training,align with risk allocation goals,and have preventive effects.Given the technical capabilities of service providers,the regulatory needs of large AI models,and the benefits sought by copyright holders,a"two-tier"safe harbor rule for service providers is essential.During the training stage,safe harbor would allow service providers to use works without prior permission,provided they satisfy the obligation of disclosure,identify and respect the reserved rights,obtain indirect economic benefits and provide comprehensive compensation.During the output stage,their obligations should include establishing a complaint mechanism,eliminating redundant data of works,optimizing model performance to prevent overfitting,mitigating malicious user conducts,and implementing request-based copyright filtering.