Seoul’s Copyright Ruling Poses Major Hurdles for AI Firms Relying on News and Textbooks (Yonhap)
SEOUL, Dec. 4 (Korea Bizwire) — South Korean authorities have concluded that artificial intelligence systems that train on entire news articles or textbooks without permission are infringing copyright, marking the government’s strongest stance yet on data use in the AI era.
At a public briefing on Thursday, the Ministry of Culture, Sports and Tourism and the Korea Copyright Commission released new guidelines on what constitutes “fair use” in AI training.
The document, titled Guidance on Fair Use of Works for Generative AI Training, clarifies that large-scale ingestion of copyrighted material—including news stories, textbooks, commercial images and music—is not permissible under current law.
Under South Korean copyright rules, fair use is allowed only if it does not conflict with “normal exploitation” of the work or “unreasonably prejudice” the rights holder. But rapid advances in generative AI have sparked disputes about whether mass data scraping and model training fall within that exemption.
The new guidance lists four factors that weigh against fair use: harm to the copyright owner’s legitimate interests, lack of transformative purpose, absence of social or public benefit, and commercial intent.
Regulators cited several examples.
Training AI on full news articles to generate commercial summaries does not qualify as fair use because it is not meaningfully transformative and could inflict economic harm on publishers, the commission said.
Under South Korean copyright rules, fair use is allowed only if it does not conflict with “normal exploitation” of the work or “unreasonably prejudice” the rights holder. (Image courtesy of Pix4free/CCL)
Similarly, feeding legitimately purchased textbooks into an AI model to produce new textbooks or workbooks was deemed infringing, as it risks weakening the market position of publishers.
The use of paid commercial images as training data to generate derivative images was also seen as violating copyright—particularly when technological protection measures are bypassed.
And AI systems trained on purchased music files to produce “AI cover songs” were deemed likely to directly substitute the originals, causing clear economic harm.
The guidance also outlines scenarios where AI training can qualify as fair use, including the use of public datasets for natural-language processing research, training on publicly accessible scientific papers to produce summaries, the use of openly available numerical data and graphs from STEM research, and analyzing lawfully acquired video for criminal-behavior studies.
Kang Seok-won, chair of the Copyright Commission, said the document reflects extensive consultation with rights holders and AI developers. “We aimed to present an interpretation standard that fits domestic conditions,” he said. “We hope the guidance reduces uncertainty surrounding AI training.”
Jerry M. Kim (jerry_kim@koreabizwire.com)
