AI Model Training Often Relies on Pirated Book Data
Research suggests major AI models are trained on unauthorized literature, signaling ongoing copyright conflicts.
Updated on Oct. 26, 2026 in Artificial Intelligence
Live Poll
Should technology companies be required to obtain explicit permission to use copyrighted books for AI training?
Research findings from March 2023 indicate that Huawei and DeepSeek integrated copyrighted, pirated books into the training datasets for their respective AI models. These collections, which included works by Korean authors, were utilized to improve model performance during competitive development cycles.
Why it matters
The use of unlicensed literature highlights a pervasive industry practice of relying on unverified scrapings to train models for long-context coherence. This approach creates significant legal and ethical risks as copyright holders continue to challenge the unauthorized use of their intellectual property in machine learning.
DeepSeek leveraged over 1 million total e-books from Anna's Archive for its DeepSeek-VL model, while Huawei utilized the 800-gigabyte Pile dataset. Analysis shows the DeepSeek model successfully reproduced 19 consecutive words from Han Kang's 'The Vegetarian', which was first published in 2007.
The players
Huawei
A major provider of telecommunications equipment and AI computing hardware that utilized The Pile dataset for its Pangu Sigma model.
DeepSeek
An AI research firm based in China that integrated over 1 million pirated e-books to train its DeepSeek-VL language model.
Anthropic
An AI safety and research company that established a legal precedent by paying $3,000 per book to settle copyright litigation.
The details
AI developers assemble massive collections of books and academic papers to improve a model's ability to maintain narrative structure and interpret long-form inputs. Researchers identified these specific training sources by cross-referencing technical reports with known pirated content repositories like Books3, a sub-collection of the 800-gigabyte Pile dataset. By training on these texts, models gain statistical proficiency in complex language patterns, though the inclusion of unauthorized material remains a central point of tension.
Timeline
1996: Kim Young-ha's I Have the Right to Destroy Myself was published.
2007: Han Kang's The Vegetarian was published.
2008: Shin Kyung-sook's Please Look After Mom was published.
2014: Han Kang's Human Acts was published.
March 2023: Researchers documented the use of pirated datasets in AI training.
The Tech Race
The use of pirated literary collections signals an industry-wide push to prioritize data volume and model depth over licensing compliance. This trajectory sits in contrast to the path taken by firms like Anthropic, which have begun settling copyright lawsuits at $3,000 per book.
Authors and rights holders are increasingly identifying their intellectual property within AI training logs, potentially leading to future litigation similar to recent industry settlements. For users, this reliance on pirated data means that current AI models may inadvertently regurgitate copyrighted passages in their outputs.
The takeaway
The unauthorized use of copyrighted literature has become a standard, if legally contested, method for enhancing AI performance. Stakeholders should monitor upcoming copyright litigation and potential industry-wide licensing standards that may force companies to purge pirated data from future models.
Further reading
Learn more about the infrastructure and methods behind modern model training in our Artificial Intelligence section.
Source note: This article includes information reported by Koreajoongangdaily.
Live Poll
Should technology companies be required to obtain explicit permission to use copyrighted books for AI training?
