Legal Challenges and Evidentiary Barriers in Determining Copyright Infringement by Generative AI: A Taiwan-Centred Analysis

This study examines the evidentiary challenges in copyright infringement litigation involving generative AI technologies under Taiwan's legal framework. Drawing from the operational mechanisms of large language and diffusion models, it explores the difficulty copyright holders face in proving substantial similarity or unauthorised reproduction when AI developers refuse to disclose training datasets. The paper analyzes two high-profile US cases: Andersen v Stability AI, where the court partially dismissed the plaintiffs' claims due to insufficient factual allegations regarding compressed copies and third-party use, and The New York Times v OpenAI and Microsoft, in which plaintiffs submitted outputs from ChatGPT and Browse with Bing that closely resembled original copyrighted articles, suggesting potential infringement of reproduction and derivative rights. These cases illustrate both the legal uncertainty and the potential for novel evidentiary strategies. The paper argues that prompt engineering--crafting input commands to provoke infringing outputs--may assist plaintiffs in building stronger prima facie cases. Finally, the paper proposes legislative reform by introducing a statutory licensing scheme specifically tailored to AI-related uses in Taiwan, aiming to reduce the evidentiary burden on authors and ensure fair compensation.