AI Companies Destroying Rare Books: A New Ethical Crisis in Training Data Sourcing

In a development that has sent shockwaves through the literary and tech communities, AI companies are systematically purchasing rare books, scanning them by cutting off their spines, and then shredding the physical originals. This practice, facilitated by a service called ISBNdb, allows companies to order up to a million books at a time while keeping buyers anonymous. The rationale is chillingly pragmatic: pre-2022 books are considered premium training data because they are free of AI-generated text, making them a goldmine for improving language model quality.

The ISBNdb Model: Anonymity and Destruction

ISBNdb's website explicitly acknowledges the PR challenge, stating, "'AI company destroys two million books' is not a headline that generates sympathy." Yet, the company has built an entire business around making this happen quietly. They offer NDAs as a feature and coach clients to refer to the process as "digital preservation." A federal judge has ruled the practice legal under fair use, arguing that eliminating the original means only one copy exists at a time, thereby reducing copyright infringement risk. This legal green light is accelerating the trend.

The scale is staggering. A bookseller told 404 Media that rare books with almost no surviving copies are being fed into this pipeline. These are volumes that survived wars, fires, and centuries of handling—now being shredded so an AI can learn to write a better marketing email. The irreversible nature of this destruction is what sets it apart from other data scraping controversies. As one commentator noted, "You can re-upload a website. You can reprint a bestseller. You can't replace the last three copies of an 18th-century botanical text once someone shreds them for training data."

Anthropic's Aggressive Sourcing Strategy

Anthropic, the company behind the Claude chatbot, has been particularly aggressive in this area. They hired the former head of Google Books partnerships with the explicit goal of obtaining "all the books in the world." This move comes on the heels of a $1.5 billion settlement approved by a judge in a copyright lawsuit brought by authors. The settlement, the largest ever paid in a U.S. copyright case, covered approximately 500,000 books, with compensation running at roughly $3,000 per title. Among the works involved were over 14,000 titles from Bloomsbury, the publisher of Harry Potter.

The court in that case introduced a critical nuance. Judge William Alsup had ruled in 2025 that training an AI on books could fall under fair use, but that reasoning fell apart when the files were sourced from pirate sites like LibGen and PiLiMi. The distinction is clear: training on books is fine; stealing them is not. This ruling effectively creates a market for legally acquired books, which is where the shredding of rare books comes into play. By physically destroying the originals, companies argue they are not creating multiple copies, thus sidestepping copyright concerns.

Broader Implications for AI and Culture

The ethical implications are profound. AI companies are effectively engaging in a form of cultural destruction to fuel their models. This is not about scanning library books or using public domain works; it's about systematically eliminating rare, often irreplaceable artifacts. The practice has drawn comparisons to the burning of the Library of Alexandria, as one user on X noted, calling it "Burning of the library of Alexandria 2.0." The destruction of physical copies means that the only surviving version of these works is now a digital scan controlled by a private company, which may or may not preserve it indefinitely.

This trend is unfolding against a backdrop of intense competition in the AI industry. Anthropic and OpenAI are both preparing for potentially trillion-dollar IPOs, valuations that depend on expectations of global market dominance. However, Chinese AI companies like Moonshot AI are producing models that can rival those from U.S. labs, with the White House accusing Moonshot of distilling Anthropic's Fable model. The pressure to maintain a competitive edge is driving companies to seek ever-larger and more unique training datasets, making rare books a prime target.

The Market and Legal Landscape

The economic incentives are clear. Tech stocks make up an outsized share of U.S. markets, and much of that growth has been tied to expectations that AI demand will continue to soar. Companies have piled hundreds of billions of dollars into data centers, chips, and energy. The cost of training data, however, is becoming a significant line item. The $1.5 billion settlement paid by Anthropic signals that the era of free data from pirate sites is over. Now, companies must pay for their training material, and rare books represent a high-value, low-supply resource.

The legal landscape is still evolving. While the fair use ruling provides some cover, it is likely to be challenged. Critics argue that the destruction of physical copies is not a legitimate part of fair use, as it eliminates the possibility of future transformative uses by others. The practice also raises questions about the preservation of cultural heritage. Libraries and archives are increasingly concerned that AI companies are stripping the market of rare books, driving up prices and making it harder for institutions to acquire them for legitimate preservation purposes.

What Comes Next

The acceleration of this practice seems inevitable unless there is a legal or regulatory intervention. The judge's ruling has, in effect, created a business model built on the destruction of cultural artifacts. AI companies are now openly seeking rare books, and services like ISBNdb are facilitating the transactions with anonymity and NDAs. The tech industry's response has been muted, with most companies declining to comment on their sourcing practices. However, the growing protest movement among Silicon Valley workers, who are concerned about AI's threat to jobs and humanity, may eventually turn its attention to this issue.

For now, the message from the AI industry is clear: the end justifies the means. The quest for better models is leading to the destruction of books that have survived for centuries. Whether this is a price worth paying for AI progress is a question that society must urgently address. As one observer put it, "We shred rare books and offer NDAs so nobody finds out" is a legitimate business model in 2026. What a timeline indeed.