Artificial intelligence companies training large language models are purchasing and scanning thousands of rare, historic books to expand their training data [1].

This practice raises critical questions about the intersection of technological progress and cultural preservation. While digitizing texts can preserve knowledge, the reported destruction of original physical copies threatens the permanence of human history.

Investigations into events occurring in 2026 indicate that AI firms are seeking massive text corpora to improve the performance of their models [1, 2]. These companies have targeted antiquarian sources to acquire high-quality, niche data that is not available in common digital archives.

Specific incidents were cited at the shop of Pieter de Vries, an antiquarian map and book dealer in Haarlem, Netherlands [2]. The process involves the acquisition of thousands of books [1], which are then scanned into digital formats. In some instances, the original physical copies are destroyed after the data has been extracted [1, 2].

Supporters of these practices said that digitization ensures the survival of the information contained within the books. They said that the digital record is more accessible and durable than decaying paper.

Critics, however, described the practice as cultural barbarism and vandalism [2]. They said that the physical artifact of a rare book holds historical value beyond the text itself, such as the paper, ink, and provenance, which cannot be replicated by a digital scan.

These activities highlight a growing tension between the data requirements of generative AI and the ethics of archival science. As models require more diverse and complex datasets to reduce errors, the hunt for rare physical documents has intensified [1].

“AI companies are buying and digitising thousands of rare books, sometimes destroying the originals.”

This conflict illustrates the 'data hunger' of large language models, where the pursuit of marginal performance gains may lead to the irreversible loss of physical cultural heritage. It suggests a future where the value of a historical object is measured solely by its utility as training data, potentially shifting the responsibility of cultural guardianship from librarians and historians to private tech corporations.