Research on Copyright Governance and Security Challenges of Training Data for Generative Artificial Intelligence
TL;DR
This paper compares the reasoning of Chinese and American courts in copyright disputes over training data, traces the concentration of infringement risk at the data-input stage, and considers why the fair use doctrine fits these cases poorly and why licensing and compensation mechanisms no longer work.
Abstract
Generative AI depends on training with very large volumes of data, and the acquisition and use of that data has become a focal point for copyright infringement and for privacy and security risks. This paper examines how generative AI training data is governed and what security problems it raises. It compares the reasoning of Chinese and American courts in copyright disputes over training data, traces the concentration of infringement risk at the data-input stage, and considers why the fair use doctrine fits these cases poorly and why licensing and compensation mechanisms no longer work. From there it turns to security, mapping risks across the data life cycle and showing, with reference to membership inference attacks and privacy extraction in federated learning, how model memorization allows training data to leak. Technical responses such as verifiable privacy-preserving inference are assessed alongside these risks. The paper concludes by arguing for a governance framework that pairs procedural proof with substantive rules, adapts fair use to local conditions, uses standardization as a lever, and relies on agile multi-party co-governance. Its purpose is to clarify how innovation and the protection of data interests can be balanced.
