In late July 2026, news that simultaneously rocked the tech and book-collecting worlds: multiple US AI companies have been revealed to be bulk-purchasing rare, out-of-print books through middlemen — guillotining off the spines before feeding them into industrial high-speed scanners, then shredding or pulping the originals. Anthropic (the developer of Claude) is documented in court filings as having destroyed an estimated 2 million physical books in its internal "Project Panama."
This is not an isolated case of corporate extremism — it reveals a structural incentive system already in operation.
The Reverse Incentive Hidden in Legal Text
US federal courts, in related copyright litigation, have handed down a ruling that is logically coherent within legal reasoning but culturally disturbing in its consequences: Anthropic legally purchased the physical books and destroyed the originals (keeping only a digital copy), and this has been classified as "fair use" — not infringement.
The logic chain runs like this: if you buy a book and scan it, but keep the original, you may need to prove your use is "transformative" and does not substitute for the original book's market. But if you destroy the original, then the only version you hold is a digital copy, and the purpose of that digital copy (training an AI model) is entirely different from the original book's purpose (the reading experience) — making it easier to satisfy the standard for transformative use.
In other words: keeping the original book carries higher legal risk; destroying the original actually makes your conduct legally safer.
US copyright law, in the new battleground of AI training data, has created a perverse incentive that "rewards destruction and punishes preservation." This was never the intention of the law — it is the lag of a legal system struggling to keep pace with technological acceleration.
This is not an isolated legal interpretation. It exposes a deeper problem: when a system of rules is designed to solve the problems of one era, it can hardly foresee where technological progress will push it decades later.
A Real Supply Chain of Collectibles
Multiple second-hand book dealers in Europe and the US have reported receiving a surge of unusual orders — placed late at night, picking random obscure academic titles, asking no questions about condition, and haggling over nothing. Those nearly-out-of-print local historical gazetteers, indigenous-language documents, and obscure scientific monographs are being bought in bulk and then disappearing into hydraulic cutters and shredders.
"Sometimes there are only a few copies left of local histories, indigenous-language documents, obscure scientific monographs — and they may just be gone."
— A second-hand book dealer's account
This is not ordinary acquisition. AI labs are not chasing bestsellers or classic literature — they are targeting texts published before 2022. The logic is simple: these books contain almost no AI-generated "garbage text," effectively preventing synthetic data contamination during model training, and thereby warding off the so-called "model collapse."
To AI companies, these books are clean training fuel. To the book-collecting world, they are irreplaceable cultural heritage. The collision of these two worlds is unfolding in the gray zone of the law.
Who Pays for the Survival of Knowledge
The operations of Anthropic's Project Panama were no secret. Court filings confirm the company legally purchased and destroyed roughly 2 million physical books. Other AI labs are reported to have conducted similar operations through intermediary platforms like ISBNdb. Buyer identities are typically kept confidential under non-disclosure agreements.
Although this conduct has caused widespread moral discomfort, it does appear to be legal under current US law. US federal courts have characterized it as "fair use" — if you buy a book, you can decide its fate, including its destruction.
This raises a much deeper question: when AI companies consume humanity's intellectual heritage as training fuel, who ensures that the original carriers of that knowledge do not disappear in the process? Can a digital copy bear the full weight of cultural transmission? If the company that scanned an out-of-print book today goes bankrupt tomorrow and deletes its server data — then a book that no longer physically exists will have truly vanished.
The Time Gap Between Law and Technology
At the root of the entire episode lies a classic temporal mismatch: the speed at which legal systems catch up with technological change is far slower than the speed at which technology itself evolves.
The current copyright framework was cemented before the digital age. Its core assumption is that copying requires authorization, and unauthorized copying constitutes infringement. But the method of acquiring AI training data — bulk purchase, scanning, destruction — creates a scenario the law never anticipated. Courts are trying to fit new problems into an old framework, and the result is a ruling that is legally coherent but culturally absurd.
This is not a call for the law to grant a "green light" to new technology, nor a blanket ban on AI training data acquisition. The issue is that when a legal system faces a scenario it never considered, the answer it produces can have irreversible consequences — and the destruction of books is precisely that: irreversible.
The essence of this "book burning" is not the moral failure of individual companies, but the structural misalignment between technological acceleration and legal inertia. Throughout human history, large-scale destruction of books has always accompanied violence or censorship. In the age of AI, it can happen quietly through lawful commercial activity. That contrast itself is the new challenge our technological era presents.
Shen Yi's Question — Destroying Correct Answers to Monopolize Knowledge
▸ Incremental addition 2026-07-29
On July 29, Shen Yi (a prominent Chinese international relations scholar and commentator) reposted coverage of the same incident, but his angle was neither legal nor cultural — it was about the logic of power revealed by an X post he had seen.
The post described how US AI companies are bulk-purchasing rare out-of-print books, scanning them at high speed to obtain clean (AI-uncontaminated) training data — then destroying the originals. A bookseller reported that those nearly-vanished local historical gazetteers, indigenous-language documents, and obscure scientific monographs were disappearing into hydraulic cutters and shredders. Anthropic, according to court filings, had destroyed roughly 2 million physical books under its internal "Project Panama."
Shen Yi's comment was just a single sentence, but the angle it chose was different from everyone else's: "Are they destroying the correct answers to seek absolute monopoly over this knowledge?"
That sentence bears unpacking. "Destroying the correct answers" — if knowledge is seen as the data required to train an AI, then destroying physical books reduces the possibility of others obtaining the same training data. "Seeking absolute monopoly" — an AI model's capabilities depend on the quantity and quality of its training data, and whoever controls the sole digital copy of high-quality data like out-of-print books possesses an irreproducible competitive advantage. The US copyright "fair use" ruling allows AI companies to destroy the original carrier while retaining the digital copy; in effect, the law has endorsed an extreme data-exclusion strategy — you can physically eliminate the data sources that your competitors might legitimately access.
This forms an interesting contrast with Anthropic's public accusations that Chinese AI labs are "distilling" its models. On one hand, accusing others of stealing your data; on the other hand, systematically reducing the data that others can legitimately obtain. Placed side by side, the two actions are not about morality — they are about power: whoever gets to define what constitutes "clean training data" also gets to decide which knowledge should be preserved and which should be destroyed.
Lingshi Xiantan's analysis points to the institutional dimension (the structural misalignment by which law rewards destruction); Shen Yi's question points to the power dimension (destroying knowledge carriers equals monopolizing knowledge itself). Together, the two threads expand the significance of this "book burning" from a question of institutional lag to a question of knowledge-power distribution in the AI age — whoever controls the scarce supply of uncontaminated text holds the key to training the next generation of models.
Musk's Choice — Speaking Up Itself Is the Industry's Signal
▸ Incremental addition 2026-07-29
On the same day Shen Yi raised his question, Elon Musk also took note of the same story. He reposted the coverage on X and stated: he had instructed the SpaceX AI team to "preserve any rare books in the library in the traditional way and scan them, rather than cutting off the spines and scanning."
When the Chinese Weibo account Baorong Wanwu Henheng Shuǐ (包容万物恒河水) relayed this news, it added: "Musk is of course no saint — but when even Musk feels the need to step on Anthropic to show his own company does it better, you know just how far gone Anthropic really is."
Musk's response, theatrical as it may be, carries a concrete requirement — "don't cut the spines" — and his decision to use Anthropic as a negative example is itself a signal. In the race for AI training data, preserving original books versus destroying them is becoming a dividing line that separates two approaches.
The comment captures another dimension of the issue: when an industry reaches a point where even Musk feels compelled to publicly say "we don't do that," it means the line has already been widely recognized. This is not a moral judgment but the formation of an industry consensus — what kinds of data acquisition are acceptable and what are not.
Just days earlier, US federal courts had ruled that Anthropic's destruction of original books was "fair use"; yet players within the same industry were already drawing a different line — with actions and public statements. The law says it's permissible, but the industry says it isn't. That dissonance, more than either position taken alone, reveals the depth of the problem.
Three angles, taken together: Lingshi Xiantan saw the institutional loophole (the law rewards destruction), Shen Yi saw the power logic (destroying knowledge equals monopolizing knowledge), and Musk's choice hints at an emerging industry norm (even a commercial competitor considers this a step too far). Three threads converging on the same conclusion: this "book burning" is not a moral question, nor a legal question — it is what happens when a fledgling industry, operating without clear rules, takes the cheapest path to fuel, until the public and the industry itself belatedly call a halt.
Shen Yi's question — "destroying correct answers to monopolize knowledge" — and Musk's public decision to preserve original books — these two threads appeared on the same day, each pointing from a different direction to the same conclusion: AI companies' destruction of physical books has moved from an internal efficiency choice to a public issue that demands a response.