AI Firms Digitize Books for Training Data
Tech companies are buying and scanning bulk physical book collections to train AI models without licensing agreements.
Updated on Oct. 20, 2026 in Artificial Intelligence
Live Poll
Should companies be allowed to use copyrighted books for AI training without explicit publisher permission?
AI developers including Last Token and BookData.ai are digitizing physical books in bulk to create specialized training datasets. The process bypasses traditional publisher licensing by treating purchased volumes as disposable training inputs.
Why it matters
Books contain dense, high-quality long-form writing that is difficult to replicate through internet scraping alone. By avoiding licensing, developers claim they can train models faster than current publisher negotiations allow.
Last Token maintains a monthly capacity to digitize up to 400,000 books. AI developers utilize these physical volumes to extract specialized knowledge and long-form prose that is largely unavailable in standard web-based datasets.
The players
Last Token
This AI data company was founded in 2026 by a former Meta employee and processes up to 400,000 books per month.
Anthropic
This AI research company utilized books supplied by the Woongjin Book Center to assist in its machine learning development.
Korean Publishers Association
This organization leads a petition against the unauthorized use of books for AI training with 158 organizational and 321 individual signers.
ROSS
This legal research company was involved in a September 2026 court ruling that restricted its use of proprietary data for training.
The details
Tech firms purchase physical books in large quantities, scan them to create digital files, and subsequently discard the physical copies. This method has triggered pushback from the Korean Publishers Association, which has gathered 158 organizations and 321 individuals for a petition against the practice.
Timeline
In 2025, Woongjin Book Center provided book supplies to Anthropic.
A US federal district court ruled in June 2025 that Anthropic's use of books for training was fair use.
In August 2026, Amazon reportedly purchased used books in large quantities for scanning.
A US appeals court ruled in September 2026 that ROSS's use of Thomson Reuters data was not fair use.
As of October 16, 2026, the Korean Publishers Association petition reached 479 total signers.
The Tech Race
The digitization of books follows the precedent set by the 2026 US appeals court ruling against ROSS regarding Thomson Reuters' headnotes. This highlights a mounting legal challenge for AI companies relying on proprietary content to fuel model performance.
Readers may notice increased demand for second-hand books, which has driven sales up by 5 times at some retailers. This trend could eventually affect the availability and cost of physical books for casual readers and collectors.
The takeaway
The race to capture high-quality training data is transforming the second-hand book market into an industrial-scale resource. Consumers should be aware that the intellectual property within books they donate or sell may end up powering future artificial intelligence models.
Further reading
Learn more about the ongoing legal landscape in our Artificial Intelligence section.
Live Poll
Should companies be allowed to use copyrighted books for AI training without explicit publisher permission?
