Explainer: LLM training-data compliance risks in copyright, personal data and trade secrets
36Kr republished an article by Sui Jingxuan, originally from the WeChat account Zhichanli, that maps large-model training-data risk across three dimensions — third-party copyright, personal information, and patents and trade secrets — and proposes a framework of classifying data by type and controlling it across the full data lifecycle. The piece cites 2025–2026 cases including US authors collectively suing six AI vendors, a Munich court finding OpenAI infringed by training on lyrics without authorization, and the Shanghai Intellectual Property Court concluding China's first AI large-model copyright infringement case. The author argues that China's Article 24 of the Copyright Law uses a closed list of fair-use exceptions that model training does not readily fit, and cites industry views that the practice falls under neither fair use nor an expanded statutory licence, leaving its legal status highly uncertain. Recommended paths include licensing or prioritizing open-source and public-domain data, anonymizing personal information as the first option, and screening data sources for patent and trade-secret content.
The article argues that the basic step of copying massive data onto training servers itself falls within the scope of the reproduction right, so training on protected works without permission carries direct infringement risk — a regulatory threshold facing every large-model developer.
生成式AI的能力跃迁，本质上是海量训练数据“喂”出来的。但现行知识产权与数据保护规则，并非为大规模机器学习场景设计，制度与技术的落差，让训练数据合规成为悬在所有大模型开发者头上的达摩克利斯之剑。
2025年至2026年期间，美国、德国、中国等地相继出现涉及AI训练数据版权侵权的标志性案件，从美国作家集体起诉六大AI厂商、德国慕尼黑法院认定OpenAI未经授权使用歌词训练构成侵权，到上海知识产权法院审结国内首例AI大模型著作权侵权案，全球司法实践正在一步步勾勒训练数据的法律边界。
本文从第三方版权、个人信息、专利与商业秘密三个核心维度，分析训练数据的侵权风险与合规路径，最终构建体系化的合规框架。
第三方版权：侵权风险与合理使用的争议边界
（一）训练行为本身即可能触发复制权侵权风险
大模型训练的基础环节，是将海量文本、图像、音视频等数据复制到训练服务器，开展预处理与模型迭代。这一“复制”行为，本身就落入著作权法第十条规定的复制权控制范围。如果训练数据中包含受著作权保护的作品，而开发者未取得权利人许可，就存在直接的侵权风险。
#OpenAI#Shanghai Intellectual Property Court#Munich Regional Court#36Kr