AI Companies Massively Buy and Destroy Second-Hand Books
Artificial intelligence companies are increasingly acquiring large quantities of used books to train their large language models, a practice that often results in the physical destruction of the volumes. The process involves digitizing the text to feed data into neural networks, after which the original hardbound or paperback copies are frequently discarded or shredded to ensure they cannot be resold or reused.
The Mechanism of Data Acquisition
Tech firms and data brokers source these materials through various channels, including secondary market book retailers and bulk donation centers. By purchasing inventory in massive volumes, these companies secure high-quality, long-form text that is essential for improving the reasoning and linguistic capabilities of generative AI models. Unlike digital archives, physical books often provide a cleaner, more structured dataset that is less prone to the noise found in web-scraped content.
Once a book is processed, the physical copy is often treated as a liability rather than an asset. According to reports from the Dutch technology news outlet Bright, companies utilize industrial scanning equipment to convert pages into machine-readable text. Following the digitization process, the books are systematically destroyed. This step is often attributed to internal logistical policies aimed at preventing the re-entry of these volumes into the consumer market, which could complicate potential copyright or licensing arrangements.
Legal and Ethical Implications
The practice has drawn scrutiny from authors, publishers, and intellectual property advocates who argue that the unauthorized use of copyrighted works for AI training constitutes an infringement. While some companies claim these actions fall under "fair use" exemptions—arguing that the data is transformed into a new, non-expressive format—legal challenges are currently moving through courts in the United States and Europe.
The destruction of the physical copies adds a layer of complexity to these disputes. Critics contend that the act of destroying the books highlights a disregard for the intrinsic value of the literature, focusing solely on the extraction of data. Furthermore, because these models require an ever-increasing volume of diverse text, the demand for physical books has created a specialized supply chain that bridges the gap between traditional publishing and high-tech data ingestion.
Industry Supply Chain Dynamics
The secondary book market has seen a shift in demand as these companies compete for specific genres and high-quality prose. Large-scale buyers are now prioritizing collections that offer linguistic variety, which has occasionally led to price increases for bulk used books. Despite the growth of this niche industry, many entities involved in the collection process maintain strict confidentiality regarding their partnerships with AI developers.
Major AI laboratories, including OpenAI and Google, have faced persistent questions regarding the specific datasets used to train models like GPT-4 and Gemini. While these organizations have acknowledged using public internet data, they have provided limited disclosure regarding the use of digitized physical books obtained through third-party liquidators. As of the current reporting cycle, no major AI company has publicly committed to halting the practice of purchasing or destroying physical books for data training purposes.