In a warehouse outside San Francisco, a contractor sliced the spines off 100,000 physical books, fed the pages through a high-speed scanner, and then incinerated the remains. The buyer was Anthropic. The project was called Panama. And the goal was not to destroy knowledge, but to possess it—without a trace.
This is not a dystopian novel. It is the leaked truth of how one of the most hyped AI safety companies acquires its training data. And it reveals a contradiction so deep that it threatens the very moral foundation of the entire industry.
We built the temple, but forgot who the god is.
Context: The Hunger for Clean Data
Artificial intelligence, at its core, is a data refinery. The quality of output depends almost entirely on the quality of input. For large language models like Anthropic’s Claude, training data must be vast, diverse, and—critically—free from digital watermarks or copyright traps that could trigger legal liability.
Most AI companies rely on web scraping, partnerships with publishers, or public domain archives. But web data is noisy, filled with SEO spam, and lacks the deep, structured knowledge found in physical books. Publishers, wary of losing revenue, have increasingly locked down digital editions. So Anthropic took a different route: they bought physical copies—second-hand, remaindered, even rare editions—and physically destroyed them to extract the pages.
The scale is staggering. Internal documents seen by 404 Media show that Anthropic ordered up to one million books in a single purchase. The books were shipped to a dedicated scanning facility where contractors cut off the spines, fed pages through industrial scanners, and then disposed of the husks. The entire operation was hidden behind NDAs. No one was supposed to know.
Core: The Ethics of Extraction
Let me be clear: as someone who has spent years auditing tokenomics and governance models, I see a pattern here that is all too familiar. It is the same pattern that led to the collapse of Terra-LUNA, the same pattern that turned NFTs into a casino. The pattern is this: a group of insiders decides that the ends justify the means, and that the rules—whether legal or ethical—are for other people.
Anthropic has positioned itself as the ethical alternative to OpenAI. Its mission is “AI safety.” Its CEO has testified before Congress about responsible AI development. Yet behind the curtain, it is literally burning books to feed its models. The hypocrisy is not just ironic; it is structural.
Code is law, until the law breaks the code.
From a technical perspective, the approach is highly efficient. Physical books contain no digital watermarks, no DRM, no encrypted metadata. They are analog gold. By acquiring them through third-party sellers, Anthropic avoids any direct licensing agreement with publishers. The legal argument would likely rest on “fair use”—the same doctrine Google used to digitize books for search snippets in Authors Guild v. Google (2015). But that case involved indexing for search, not training commercial AI models. And Google did not destroy the books.
From an ethical standpoint, the destruction is unforgivable. Rare books, out-of-print editions, and culturally significant works are being incinerated. Once scanned, the physical copy is gone forever. Libraries, collectors, and future researchers lose access to that artifact. The data is now locked inside a private corporate model, accessible only via API—if Anthropic chooses to share it.
This is not open science. This is extraction colonialism, applied to knowledge.
The Double Standard
The most damning critique comes from David Sacks, the venture capitalist and co-host of the All-In Podcast. He pointed out that Anthropic has publicly demanded that other AI companies stop training on its Claude outputs without permission. Yet here they are, taking the entire contents of physical books without any consent from the authors or publishers. “They want to protect their own IP while freely taking everyone else’s,” Sacks wrote. “That’s not ethical. That’s rent-seeking.”
Elon Musk, never one to miss a chance to attack a competitor, also weighed in. He announced that his new AI company, xAI, would use “non-destructive scanning methods” for any physical books it digitizes, and would preserve the originals. Whether this is a genuine commitment or a PR stunt is irrelevant—it forces Anthropic into a defensive posture. The message is clear: you can compete on model quality without destroying cultural heritage.
Authenticity is a signal lost in the noise.
Contrarian: A Necessary Evil?
Let me pause and offer a contrarian view. The hunger for high-quality, low-noise training data is not going away. Models are plateauing on public web data. If we want AI that can understand nuance, history, and rare languages, we must find ways to digitize books that have never been digitized. Is it better that these books rot in a warehouse or remain locked in a private collection, only to be scanned piecemeal by Google Books under restrictive terms?
Perhaps the real crime is not the scanning, but the secrecy and the destruction. If Anthropic had partnered with libraries, preserved the physical copies after scanning, and released the resulting dataset under an open license for academic research, the narrative would be radically different. They could have been heroes. Instead, they chose to hide, and that choice reveals their true priority: not knowledge, but competitive advantage.
Furthermore, the “destroyed” books are mostly mass-market paperbacks and overstock remainders. The number of genuinely rare editions is small. Anthropic likely avoided damaging truly irreplaceable items—though the documents do not prove this. But even one lost rare book is one too many.
The Unanswered Questions
Several critical questions remain. First, does Anthropic still possess the scanned digital copies? If so, are they planning to keep them indefinitely? The destruction of the physical medium may have been a one-time security measure to prevent leaks, but the digital files are the real asset. If Anthropic stores them on a centralized server, they become a single point of failure—and a single point of control.
Second, is this practice unique to Anthropic? Internal sources suggest that at least two other major AI labs have considered or piloted similar physical-digitization projects. OpenAI has denied it. But without external audits, we cannot know. The industry operates in a fog of secrecy that is fundamentally incompatible with the transparency that blockchain advocates (like myself) believe is essential for trust.
Third, what will be the legal fallout? The U.S. Copyright Office is currently preparing a report on AI and fair use. This scandal will land squarely on their desks. A ruling that destroys physical books for commercial AI training is not fair use could set a precedent that forces every company to reveal its data sources. That would be a good thing.
Takeaway: The Ledger Remembers, but the Heart Forgets
This is not just a story about Anthropic. It is a story about the entire philosophy behind closed-source, centralized AI development. Blockchain advocates have long argued that transparency is a precondition for trust. The same principle applies here. If we are to build AI that serves humanity, we must know what it was trained on. We must be able to audit the data, verify the consent, and ensure that no cultural heritage was destroyed in the process.
Truth is not a token you can trade.
The book burning at Anthropic is a symptom of a deeper disease: the belief that intelligence can be extracted from the world without giving back. That is not a sustainable model. It is not ethical. And it is not, ultimately, safe.
We traded soul for speed, and called it progress.
As the dust settles, I hope this scandal sparks a broader conversation about data sovereignty, digital provenance, and the role of decentralized ledgers in tracking the origin of training data. Projects like OriginTrail, which use blockchain to create verifiable supply chains for data, could become the standard for ethical AI. If every book scan were timestamped, hashed, and linked to a public ledger, we could prove that the digitization was done with consent and without destruction. That is the future I want to see.
For now, the temple has been built on a foundation of ash. We must decide whether to rebuild on a foundation of trust.
Faith in the protocol is not faith in the people.
Based on my experience auditing DeFi protocols, I have seen what happens when transparency is optional. Shortcuts become habits. Habits become culture. And culture, once corrupted, is nearly impossible to repair. Anthropic still has time to come clean, to open-source their scanned datasets, and to donate to libraries. But the window is closing. The public is watching. And the ledgers, unlike the books, cannot be burned.
