AI Should Read The Classics. We Should Keep Them Too.

AI Companies Scan Used Books for Training Data, Sparking Preservation Debate

Reports indicate that several artificial intelligence developers have purchased large quantities of used books, scanned their contents for model training, and disposed of the physical copies afterward. Booksellers and collectors have expressed concerns, particularly regarding rare or older editions whose physical forms hold historical value beyond their text.

The practice stems from the need for high-quality textual data to train large language models. Proponents argue that exposure to established literary works benefits system development by providing structured language, narrative patterns, and historical context. Critics counter that discarding originals risks losing irreplaceable artifacts when digital versions serve as the sole remaining record.

Legal frameworks governing such uses remain under discussion. Courts have examined whether training on copyrighted material qualifies as fair use, with some rulings supporting transformative applications while others require licensing or permission. These decisions affect the scale at which companies can acquire and process physical books without retaining them.

Historical examples illustrate risks associated with post-publication alterations. In 2002, director Steven Spielberg modified the re-release of the film E.T. the Extra-Terrestrial by replacing firearms with communication devices; he later stated that the change was an error and restored the original elements in subsequent versions. Publishers of Roald Dahl’s children’s books issued revised editions that adjusted phrasing and descriptions, citing alignment with current market expectations; the changes were implemented by the rights holders after the author’s death.

Such cases demonstrate that digital files can be updated or replaced by their controllers, whereas physical copies remain fixed once produced. Ownership of originals therefore provides an independent reference point that cannot be remotely modified.

Discussions of book availability often center on institutional policies. Restrictions in schools or libraries limit access within those settings but do not prevent purchase or possession elsewhere. In contrast, exclusive reliance on centralized digital platforms introduces the possibility of removal, access revocation, or content revision by the platform operator.

Preservation strategies vary. The Internet Archive has documented scanning protocols that prioritize retention of fragile volumes, using slower methods to avoid damage during digitization. This approach aims to produce digital surrogates while keeping physical items intact. Commercial scanning operations focused on volume may apply different standards driven by cost and speed.

Private collections, used bookstores, university libraries, and individual owners currently hold multiple copies of many titles. Redundant physical distribution reduces the impact of any single loss or alteration. Ownership by individuals rather than solely institutions or corporations adds further distribution of control.

Advocates for broader access note that digitization can surface lesser-known works through search and translation tools. At the same time, they acknowledge that a digital file requires ongoing maintenance by whoever hosts it, creating dependencies not present with physical objects.

The underlying tension involves balancing efficiency in data acquisition against the durability of physical records. A 500-year-old volume functions both as text and as a material artifact whose production methods and condition supply information separate from its words. Decisions to destroy such objects during scanning treat the content as extractable without residual value in the original form.

Industry representatives have stated that the volume of text required for current training runs makes exhaustive retention impractical under existing timelines. Preservation groups respond that certain categories of material warrant slower handling regardless of efficiency metrics.

No single policy resolves these trade-offs. Continued acquisition of physical copies by private parties, combined with selective digitization that retains originals where feasible, represents one documented approach. Legal precedents on fair use and copyright will likely shape future acquisition methods.

The article structure follows the original sequence of topics while substituting reported facts and attributed statements for evaluative phrasing. Claims about cultural continuity or technological progress are presented as positions held by participants rather than settled conclusions.