Skip to content
MarketsIndicesCommoditiesFXRates
Top News

'Project Panama' Reveals AI Firms Destroying Old Books

Booksellers worldwide report bulk book orders tied to Anthropic's Project Panama, a scheme to destructively scan books for AI training.

'Project Panama' Reveals AI Firms Destroying Old Books

Marçal Font, owner of Librería Fénix, a well-known secondhand bookshop in the Catalan city of Badalona, received an order on April 30 that caught his attention. He told BBC Mundo he has worked as a secondhand bookseller for 20 years and has seen millions of strange purchases, so little alarms him anymore.

Font, who is also a poet and a professor of Comparative Literature at the University of Barcelona, said the order came from a company based in Canada. His shop had shipped Catalan-language books abroad before, even to Japan, but he said the buying pattern this time was unusual.

Suspicious order patterns

When a new round of orders arrived in early May, Font decided to investigate for fear of fraud. He said it was rare to look into who was buying from the shop, but in this case several emails arrived within a window of 40 minutes to an hour, each requesting a different number of books, with shipping costs he described as absurd.

Font contacted his friend Xavier Vinaixa, an artificial intelligence researcher and technical director at the tech company Sorensen.ai. Vinaixa recalled Font calling him about the strange pattern, the same buyer, shipments abroad, and shipping fees paid three times over, and that his first thought was that it might be a scam.

Vinaixa said that as they looked further, they found other people with the same concerns in Germany and New Zealand, which made the picture clear. A search for "antique books" and "artificial intelligence" turns up similar accounts. The British newspaper The Telegraph reported that a bookseller in Haarlem, Netherlands, Pieter de Vries, said it was very unusual in the antiquarian book trade for someone to want more than one book, so an order for a couple hundred copies stood out. The Guardian quoted a bookseller in Victoria, Australia, Delfina Manor, who received an order from the same company that bought from Font; she said the buyer placed the order, paid in advance and raised no objection to the shipping costs.

The Anthropic settlement and a leaked plan

Both reports, along with coverage of the Catalan bookseller's case, referenced an investigation that uncovered what is known as "Project Panama." In September 2025 the AI company Anthropic agreed to pay $1.5 billion to settle a class action lawsuit brought by authors who alleged the company used their work without permission to train its AI models.

Four months later, The Washington Post published a document from the court proceedings describing a plan to digitize every book ever written. The leaked text stated that Project Panama was the company's effort to destructively scan all the books in the world, and that it did not want its work on this to become known.

Maximiliano Firtman, an Argentine professor and author of books on programming and artificial intelligence, was one of the plaintiffs in the class action. He told BBC Mundo the judge ruled that if the company purchased a copy of a book, it could use it to train its AI. He said testimony then began emerging from booksellers worldwide who received orders for books no one normally touched, alongside companies that scan such books and hand the resulting text to AI firms, which is how the destruction of books entered the picture.

Two ways to digitize a book

Rodrigo Álvarez, a technology journalist at the Argentine outlet Todo Noticias, told BBC Mundo there are two digitization methods. One does not destroy the book, using flatbed scanners or V-shaped devices that scan pages without taking apart the binding.

The other, the "destructive" method, begins by cutting the spine so the pages come loose. The separated pages then go through high-speed industrial scanners that digitize them automatically and much faster. Once digitized, the book can no longer be reassembled, and the remains are usually sent for recycling. Álvarez said companies choose this method for scale, since it allows millions of copies to be digitized faster and more cheaply, and confirmed this was the procedure used for Project Panama.

Feeding the AI's need for data

Vinaixa explained that when people talk about AI, they usually mean a large language model, or LLM, a program that learns to understand, summarize, translate and generate text by predicting the next word or fragment. Firtman compared it to predictive text on a phone keyboard, but multiplied by a million.

To train these models, developers first used everything available online, including Wikipedia, blogs and forums, and when that began to run out, they moved to newspaper articles and even YouTube video subtitles. Firtman said that once that material was exhausted, companies turned to already-digitized printed books, then libraries, and now to obscure, non-commercial books found only in secondhand shops and antiquarian dealers. Vinaixa said Font had mentioned orders for material such as a history of Catalan cured meats, the kind of text that has likely never been used to train AI before.

Concerns over rare and out-of-print books

Miguel Ángel Ortega, president of Spain's Professional Association of Antiquarian Books and Collecting, told BBC Mundo a distinction should be made between mass print runs and limited editions. He said dealers who preserve cultural heritage believe a very limited, scarce or unique edition should not end in destruction.

Álvarez agreed, saying the biggest risk is that indiscriminate buying destroys rare or out-of-print copies without first checking how many remain in stock at distributors, bookshops, or in private hands.

Font said he is not especially concerned about the destruction itself, comparing it to veterinarians who train to save animals but end up euthanizing the most. He said no one destroys more books than a secondhand bookseller, since they buy entire libraries and save what they can, though much cannot be saved. He added that if a book has a legal deposit copy guaranteeing at least five or ten preserved copies, he sees no problem with one copy being destroyed so its content moves into a new format.

What digitization leaves out

Font's real concern is that mass digitization overlooks material without an ISBN, the International Standard Book Number assigned to published works. He said political dissent literature, LGBTI activist writing, rebellious fanzines, counterculture and underground publications have historically been produced without ISBNs and could be lost if not digitized as well.

Both Álvarez and Font stressed that digitization is not the same as preservation. Álvarez said digital files typically end up in private databases, out of reach of readers, libraries and researchers, and that extreme digitization can also lose features beyond the text itself, such as binding, annotations, illustrations, print quality and provenance.

Font said the absence of this kind of metadata undermines proper interpretation of a text, comparing it to the shift from stone inscriptions to paper, where transcribing the text was not enough; the stone's materials and location also had to be recorded. Ortega added that two books with identical text can carry very different cultural value if one has an ex libris, artisanal binding or handwritten notes that cannot be transferred to a digital file.

Proposed alternative and what comes next

To avoid destructive scanning by intermediaries, Font and Vinaixa proposed that institutions digitize their own heritage and license the resulting data while keeping the physical books. Vinaixa said demand for training data will keep growing until there is nothing left to feed AI systems, noting that AI's limits turned out to be data availability rather than processing power.

Firtman said the next step is synthetic data, text generated by AI itself, and that the biggest risk in coming decades is that AI could degrade if it is trained on content it produced itself. Álvarez likened this to a snake eating its own tail, warning that companies feeding models with AI-generated content could multiply errors and progressively simplify language, producing a summary of a summary of a summary that drifts further from the original information.

At Librería Fénix, the situation has kept evolving since the unusual orders of April and May. Font said the buyer has changed logistics and companies, and has begun asking that books be sent not only to North America but also to Germany. After a month without new orders, Font received another order this week, which he said can only mean one thing: Project Panama is still underway.

Related

Leave a comment

Your email address will not be published. Required fields are marked *