The Applicability of Reproduction Right During AI Model Weight Training: Comparative Research Based on Two Typical Judgments

Vollständiger Abstract

Worum geht es in dieser Arbeit?

The fast-growing generative artificial intelligence industry have brought unprecedented challenges to existing copyright rules, especially regard how developers utilise copyrighted literary, visual and coding works to train neural network models. Legal scholars long debated one core question: does building model weights in the training process constitute a copyrighted reproduction act as define by national copyright laws and international treaties? This paper compare two landmark verdicts released in recent years: the 2024 Ultraman AI copyright dispute judged by Hangzhou Internet Court in China, and Thomson Reuters v. Ross Intelligence ruled by the District Court of Delaware in the United States in 2025. Through case analysis and comparative legal research, this paper sort out different judicial attitudes toward three core technical acts: raw data ingestion, temporary data storage in computing memory, and final weight parameter fixation. The analysis show that Chinese judges adopt an output-centred judging logic, treating temporary storage of copyrighted content during training as an inevitable auxiliary technical step without independent infringement liability. By contrast, American courts conduct a full four-factor fair use test covering every stage of AI training workflow. To balance technological progress and creators' exclusive copyright benefits, this paper put forward a two-tier "market impact balancing test" for courts to judge reproduction disputes arising from AI weight training.