[Paper Review] Generative AI Copyright: Transformative Use and Enjoyment of Emotions

Paper: "First Rulings on Fair Use for Generative AI Training — Comparative Analysis of Anthropic and Meta Cases" (Lee Dae-hee, 2025). Building AI datasets from vast copyrighted works can in principle constitute copyright infringement (reproduction rights violations). But obtaining permission from every rights holder is practically impossible and would stifle the AI industry. Japan''s 2018 copyright law amendment: added provision (Article 30-4) permitting "use not aimed at enjoyment of thoughts or emotions expressed in the work." The provision''s logic: copyright is designed to ensure creators can receive compensation for the "enjoyment" of their works — if content is being processed mechanically for pattern recognition rather than being experienced/enjoyed culturally, it doesn''t compete with the rights holder''s market. Therefore AI training (mechanical processing, not appreciation) is permissible without consent. US fair use analysis in Anthropic and Meta cases: Anthropic case — court ruled AI training on licensed/legally obtained data qualifies as transformative fair use (transforms expressive works into statistical patterns for different purposes); Meta case — more complex where pirated data (Library Genesis) was used; court declined to extend fair use protection to illegally obtained training data, finding the piracy itself removes the transformative use defense. The key distinction emerging: the legality of how training data was OBTAINED matters independently of how it was USED — transformative use arguments cannot sanitize the means of acquisition. Korea''s position: between Japan (permissive statute) and US (case-by-case fair use analysis); current legal framework unclear on AI training; regulatory discussion ongoing but legislation still early stage. Practical implication: AI companies training on public web data face different legal risk profiles depending on whether that data was licensed, public domain, or scraped without authorization — requiring careful audit of training data provenance.