Anthropic’s secret Project Panama to buy millions of books for AI training could have consequences for South African authors, after a US court sanctioned the company’s practice of shredding and scanning lawfully purchased books.
Some local commentators warn it could strip older, scarcer works from circulation; others see it as a lawful step in AI’s evolution, provided companies play by the rules.
World Wide Worx MD and author of 21 books, Arthur Goldstuck, says: “My first reaction was a wry acknowledgement of the audacity of it, followed by astonishment at the sheer disrespect for books. The most advanced AI companies in the world have discovered that one of their most valuable raw materials is an old-fashioned printed book.”
AI companies are buying books by the pallet, cutting off the spines, scanning every page and destroying what remains, Goldstuck explains. “It sounds like post-apocalyptic science fiction. Perhaps a sequel to “Fahrenheit 451” in which books are all burned and people memorise them.”
T4i director Mark Walker, formerly with research company IDC, says: “As a bibliophile, I felt outrage and disappointment; as a rational technologist/scientist, I wasn’t surprised and I am hopeful that the original scans will be preserved digitally for future retrieval, analysis and enjoyment.”
A rarity
The irony, says Goldstuck, is that printed books now offer something increasingly hard to find online. “I love that books written before the generative AI boom offer something increasingly scarce online: words known to have been written by people.”
South Africa has an immediate reason to pay attention, says Goldstuck, as the country is moving towards an American-style fair-use system through the Copyright Amendment Bill, back with Parliament, just as US courts apply that doctrine to AI training.
The Bill would replace SA’s “fair dealing” system with a broader “fair use” exception, allowing copyrighted material to be used without permission for education, research, libraries, archives and people with disabilities, provided the use is considered fair.
The Constitutional Court upheld most of the Bill in June but sent parts of the education provisions back to Parliament as unconstitutional.
“Yet our Bill says nothing explicit about generative AI, text and data mining, or payment for the use of creative work in commercial models,” says Goldstuck. “Parliament therefore has the opportunity to address the biggest copyright dispute of the AI era.”
Hush-hush
The case was brought by New York Times best-selling author Andrea Bartz and others, alleging Anthropic infringed their copyrights by training AI on their books without permission.
It exposed Anthropic’s secret “Project Panama”, under which the company spent tens of millions of dollars buying printed books, stripping their bindings and scanning them into a digital library before discarding the originals.
Federal Judge William Alsup ruled that training AI on books, and creating a single internal digital copy of lawfully purchased works, qualifies as fair use. Anthropic must still face trial over its alleged use of millions of pirated books.
Writing for book database ISBNdb, Val Giordano says Anthropic kept the programme confidential “not because it was illegal, but because it looked bad. Destroying millions of books evokes images of burning libraries, even if the law was on their side.”
Legal and fair aren’t the same thing, argues Goldstuck, who says the ruling leaves authors without a clear share of the value created from their work. “An author can lose a fair-use case and still have a legitimate grievance about receiving nothing. That exposes a gap in the law, even if the judge applied existing law correctly.”
Walker says fairness depends on what happens to the digital copies. “If this is just a mass copy and destruction only to feed the AI maw, then no I am not happy as the terms are too narrow. If this includes digital capture, attribution and coupled to full free public access then it will be fairer.”
ICT veteran commentator Adrian Schofield disagrees. “Millions of books are already available via the internet for nothing or for a few cents. Equally, this will not change the value of older books. They have either exhausted their market, or a relatively few adherents will continue to seek out hard copies. This process does not abolish libraries.”
Goldstuck agrees that cleaner training data could improve AI models, but warns the books being gathered will also shape the worldview built into them. “AI could therefore become better trained on a narrower slice of humanity, while much of Africa remains outside the library.”
The Book Thief
Vanessa Rudman, owner of specialist bookstore Books & Collectables, says the debate goes beyond copyright or AI. “Most books being bought for AI training won’t be museum pieces, but they’re still part of a finite pool of printed works. Every time a physical copy is cut apart and destroyed, that’s one less copy available to readers, researchers, collectors and libraries.”
Rudman doesn’t think AI companies are deliberately targeting rare or culturally significant books, but the cumulative effect could still matter. “If older books become a raw material for AI, we’re changing what those books are. Instead of being preserved as cultural objects or sources of knowledge, they become consumables. That has implications for collectors, second-hand booksellers, researchers and anyone trying to source older works.”
Goldstuck wants greater protection for culturally significant works. “Once millions of books are removed from second-hand circulation, this is no longer merely a private owner disposing of private property. It becomes industrial-scale extraction of cultural material. The law may allow a company to destroy a book it has bought, but rare, annotated and culturally significant books need additional protection.”
Schofield rejects the book-burning comparison as legally purchased books belong to their owners, and genuinely significant works are likely to be protected by libraries, museums and cultural organisations anyway. “Who decides that a book is culturally significant? The Bible, the Koran, Mein Kampf, Project 2025,” he asks.
Rudman says it’s worth asking whether safeguards are needed for works of historical, cultural or research value before they’re permanently removed from circulation.
“Books are more than containers for text. They are artefacts in their own right. Once a physical copy is destroyed, future generations lose the opportunity to experience that edition exactly as it was published.”
Share
