Singularity.Kiwi AI progress, with honesty.
A dark library with books on shelves, one book lying open on a table with a bright scanner light passing over it, the book slowly dissolving into particles, somber and ominous lighting, no text no logos no letters no signs
News 27 July 2026 · 18:38 NZST

AI Companies Are Shredding Rare Books to Train Their Models — and a Judge Said It's Legal

Anthropic's Project Panama shredded millions of books. ISBNdb now brokers orders of up to a million books with NDAs. A judge said destroying the original is 'transformative.' But what happens when three AI companies each buy one of three surviving copies of a rare work?

AnthropicCopyrightAI Training DataCultural HeritageFair Use

The books that survived wars, fires, and centuries of handling are being fed into machines. A hydraulic cutting machine neatly removes the spines. Industrial scanners capture every page. The originals are then scheduled for recycling. An AI model learns to write a better marketing email, and a book that existed for 200 years no longer does.

This is not a hypothetical scenario. It is happening right now, at industrial scale, and a federal judge said it’s legal.

The Book Shredder

In January 2026, the Washington Post revealed that Anthropic — the company behind Claude, one of the world’s most advanced AI models — had been running a project internally called “Project Panama.” Its goal, in the words of Anthropic executives, was to “destructively scan all the books in the world.”

The process was efficient and deliberately quiet. Books were purchased in bulk from resellers. A hydraulic cutting machine removed the spines. Pages were fed through high-speed, production-grade scanners. The physical books were then collected for recycling. Anthropic hired the former head of Google Books partnerships to help source the books. They also, according to court documents, didn’t want it to be known that they were doing this.

In June 2025, Judge William Alsup ruled that the practice was fair use. His reasoning was simple: the print original was destroyed, the digital copy replaced it, and the digital copy was never shown, shared, or sold outside the company. He called the process “exceedingly transformative.” Anthropic was ordered to pay $1.5 billion for separately downloading pirated digital books — but the physical book-shredding was given a green light.

That ruling didn’t slow anything down. It accelerated it.

The Industry That Grew From a Loophole

A company called ISBNdb, which describes itself as the “world’s largest book database,” now facilitates bulk orders of between 1,000 and one million books for AI training. According to reporting by 404 Media’s Emanuel Maiberg in July 2026, ISBNdb keeps buyers anonymous and offers NDAs as a service.

ISBNdb’s website is candid about the optics. “AI company destroys two million books is not a headline that generates sympathy,” it acknowledges. It coaches clients to call the practice “digital preservation.” It markets pre-2022 books as “structurally guaranteed to be free” of AI-generated text contamination — clean human writing from before large language models flooded the internet with synthetic prose.

“A physical book is a delivery mechanism for information,” ISBNdb explains on its site. “Once that information has been extracted and encoded into an AI model, the delivery mechanism has served its purpose.”

A bookseller told 404 Media that in April, he went from selling no more than 20 books a week to hundreds. His inventory includes rare and out-of-print books — meaning AI companies may be destroying some of the few remaining copies of certain titles. In the Netherlands, rare booksellers are being inundated with bulk purchases they believe are coming from AI companies, though the NDAs make it impossible to confirm.

The Cumulative Problem

Judge Alsup’s fair use reasoning was built on a single-transaction model. One book is bought. One book is scanned. One book is destroyed. One digital copy exists. Therefore, only one copy exists at a time. The reasoning is internally consistent for a single transaction.

But ISBNdb is facilitating an industry, not a single transaction. Multiple AI companies are buying books independently. They are not sharing their scans — competitive advantage means hoarding training data. Nobody is coordinating across companies to track how many physical copies of a given title survive.

Consider a book where three copies remain in existence. Three different AI companies each buy one copy. Each company scans its copy and destroys the original. Each individual transaction is legal under Judge Alsup’s reasoning. The cumulative result is that zero physical copies survive, and three private digital copies exist in three separate corporate silos, none of which the public can ever access.

The law has no mechanism to see the whole picture. Each transaction is evaluated in isolation. The destruction of a cultural artifact happens one purchase at a time, each one individually legal, the final outcome being total loss.

The Goldie Question

This is where the principle being established starts to matter beyond books.

Books work for this legal strategy because multiple copies usually exist, and the first-sale doctrine allows the buyer to do what they want with a physical purchase. But not all cultural artifacts are books.

Charles Goldie’s portraits of Māori chiefs are one-of-a-kind oil paintings. They cannot be bought, scanned, and destroyed under any legal framework. They are unique. They hold cultural and spiritual significance that goes beyond their informational content. They are taonga.

But art is not always unique. Limited edition prints, hand-pulled lithographs, woodcuts, and early photography — these exist in small runs. Some older NZ prints may have only a handful of surviving copies. These sit closer to the “rare book with few remaining copies” situation than to a Goldie painting.

If an AI company bought a limited edition print, scanned it, and destroyed the original, the legal argument would be structurally similar to the book case. Buy a physical copy. Scan it. Destroy the original so only one copy remains. Call it transformative. The fact that only 20 copies ever existed, and now 19 remain, would be legally irrelevant — because the law evaluates each transaction individually.

Does New Zealand Need to Legislate?

New Zealand is in a different legal position from the United States. NZ has no “fair use” doctrine — it has “fair dealing,” which is narrower and doesn’t include the flexible, case-by-case balancing test that US courts use. The Copyright Act 1994 is currently under review, and no specific AI training exemption exists in NZ law yet.

This means the specific legal strategy that made book-shredding legal in the US — the transformative use argument under fair use — would not automatically apply in New Zealand. NZ courts would need to interpret fair dealing, which has historically been more restrictive.

But that also means there is no clear framework for what AI companies can do with physical works in NZ. The absence of a ruling is not the same as protection. And the Copyright Act review is an opportunity — either to explicitly ban destructive scanning for AI training, or to create a framework that allows training while requiring the physical original to survive.

The question is whether NZ should wait for a test case or legislate proactively. The US example shows what happens when the legal system reacts rather than anticipates: a single judge’s ruling created a precedent that built an industry. By the time appeals could correct it, millions of books were already gone.

The Protected Objects Act 1975 covers taonga tūturu — tangible Māori cultural objects — but does not extend to Pākehā art or to books. A Colin McCahon painting, a rare first edition of Katherine Mansfield’s Bliss, or a limited run of historical photographs would all fall outside the Protected Objects Act and rely solely on copyright law.

The Principle

The core issue is not about books, or art, or any single medium. It is about the principle being established: that a physical cultural artifact is merely a “delivery mechanism for information,” and that destroying it after digitising it is acceptable because the information is preserved.

But information is not the same as a cultural artifact. A scan of a book captures the text. It does not capture the binding, the paper, the marginalia, the provenance, or the physical presence of an object that existed for centuries. A scan of a painting captures the image. It does not capture the brushwork, the texture, the scale, or the fact that a human being stood before a canvas and made something that did not exist before.

The “digital preservation” framing assumes that the information content is all that matters. If that principle becomes embedded in law — in any country — then the physical destruction of cultural artifacts becomes a licensing issue, not a preservation issue. The question of whether both a physical and digital copy should be required to survive is not just a policy detail. It is the difference between preservation and extraction.

For New Zealand, the question is whether a small country with a relatively small body of cultural artifacts should wait for an AI company to test the boundaries of its copyright law — or whether it should decide now, while the physical copies still exist, what it wants to protect.

Sources:

Related:

— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.

Related Articles