以实质性相似标准为视角论AI侵权认定中的“过程审查”

“Process Review” in the Determination of AI Infringement from the Perspective of Substantive Similarity Standards

摘要: 生成式人工智能输出内容与在先作品构成实质性相似时侵权责任的认定，已经是著作权法领域十分明确、紧迫的前沿课题。传统“接触 + 实质性相似”判定规则是以人类创作为基本前提构建的，故自然将侵权认定集中于结果层面的表达比对，但毋庸讳言，AI生成内容的技术逻辑与传统创作有根本性不同：大模型把训练数据分解为令牌，以统计相关性为基础做随机重组，因此其输出有概率生成而非复制再现的本质特征。正因如此，传统规则在面对AI生成内容时必然遭遇三重困境：技术本质与法律定性的错位、判断标准适用的困境、责任主体分配的模糊。因此本文提出了以下解决方案：AI生成内容的侵权认定宜从“结果比对”转向“过程审查”，系统地将技术过程的拆解、人类贡献的评估、模型状态的识别都纳入考查范围。具体而言，应区分“模型记忆型输出”与“概率生成型输出”，前者适用严格的实质性相似标准，后者以人类贡献度为责任分配的核心依据。由此导出著作权保护与技术创新之间的动态平衡。

Abstract: When the output content of generative artificial intelligence is substantially similar to an existing work, the determination of infringement liability has become a very clear and urgent frontier issue in the field of copyright law. The traditional “contact + substantial similarity” determination rule is constructed based on human creation as the fundamental premise, so it naturally focuses on the expression comparison at the result level of infringement determination. However, it is undeniable that the technical logic of AI-generated content is fundamentally different from traditional creation: large models decompose the training data into tokens and randomly reorganize them based on statistical correlations. Therefore, their output has the essential characteristic of probabilistic generation rather than replication and reproduction. As a result, traditional rules inevitably encounter three dilemmas when dealing with AI-generated content: the mismatch between technical essence and legal determination, the dilemma of applying judgment standards, and the ambiguity of liability subject allocation. Therefore, this article proposes the following solutions: The infringement determination of AI-generated content should shift from “result comparison” to “process review”, systematically including the decomposition of the technical process, the assessment of human contribution, and the identification of model status in the examination scope. Specifically, “model memory-based output” and “probabilistic generation-based output” should be distinguished. The former should apply the strict standard of substantial similarity, while the latter should use the degree of human contribution as the core basis for liability allocation. This leads to a dynamic balance between copyright protection and technological innovation.