In April 2025, internal documents obtained by 404 Media revealed that Anthropic had allocated $12 million to purchase over one million physical books—many of them rare, out-of-print editions—and subsequently destroyed them by cutting their spines and feeding them into high-speed scanners. The goal was to generate a pristine digital corpus for training their Claude series of language models. The program, codenamed 'Project Panama,' was kept under a strict non-disclosure agreement specifically to hide the buyer’s identity from publishers and the public.
This is not a story about a startup's desperation for data. It is a case study in the breakdown of verifiability in the AI supply chain—a breakdown that the blockchain community has been warning about for years. As an on-chain detective who has spent the past seven years tracing transaction patterns and auditing smart contracts, I recognize the same pattern of selective transparency and unaccounted-for externalities that I saw during the DeFi Summer liquidity stress tests, the NFT wash-trading bubble, and the Terra/Luna collapse. The details differ, but the structural flaw is identical: a party with asymmetric information and execution power makes irreversible decisions that affect a distributed network of stakeholders, while claiming the moral high ground.
Code speaks louder than promises. Anthropic claims to build 'safe, beneficial AI' and has publicly criticized competitors for training on its own model outputs without permission. Yet Project Panama shows that when it comes to acquiring training data, the company is willing to destroy physical artifacts of human knowledge—irreplaceable first editions, regional histories, and academic monographs—to gain a competitive edge. The scanners processed books at a rate of 80 pages per second, and the destruction was absolute: books were cut, scanned, and the paper remnants were incinerated. The digital copies were stored in a closed data lake with no external audit trail.
The Forensic Lens
Let me apply the same method I used in 2021 when I exposed the wash-trading bot network behind the top ten NFT collections. Back then, I linked 40% of trading volume to four wallet clusters controlled by a single entity, using on-chain analysis of gas consumption and temporal patterns. The problem was not that the trades were illegal—they were on-chain, after all—but that the narrative of organic demand was fabricated. Similarly, Project Panama is not necessarily illegal under current copyright law; the 'purchase and destroy' strategy may fall within the 'fair use' framework established by the Authors Guild v. Google case. But the ethical and structural implications are just as damaging.
Follow the gas, not the narrative. In blockchain, we track the gas spent on each transaction to identify anomalous behavior. In the AI training data ecosystem, the comparable metric is the cost of data acquisition relative to its volume and quality. Anthropic spent roughly $12 per book, including shipping and scanning labor. That is significantly cheaper than licensing the same content from academic publishers, which often demand $50–100 per article. By destroying the physical copies, the company not only avoided licensing fees but also eliminated the possibility of future copyright claims—physical evidence gone. The digital files were stored in a proprietary format without any public hash or timestamp. There is no way for external auditors to verify what was scanned, how it was processed, or whether the training run actually used that data.
This is the equivalent of a crypto project claiming a billion-dollar TVL but refusing to release its smart contract addresses for verification. The market takes it on faith. And in crypto, faith without verification is a liquidation event waiting to happen.
The Actuarial Cost of Destruction
Based on my experience auditing the 0x Protocol v2 smart contracts in 2018, where I discovered seven critical vulnerabilities in the order routing logic, I have learned to focus on the 'invisible liabilities' in any system. In that audit, the reentrancy flaw was hidden not in the public-facing functions but in the internal bookkeeping of fill orders. Similarly, the invisible liability in Project Panama is the loss of cultural and informational diversity.
Anthropic’s data team reportedly targeted rare books from private collections and university library discards. These are not the digitized bestsellers available on Amazon Kindle. They are niche academic works, regional poetry, technical manuals from the 1980s, and self-published memoirs. The network of rare-book dealers who supplied the volumes was instructed to remove any provenance markings—ownership stamps, library barcodes, annotations—before shipping. Once the books were scanned and destroyed, the derivative digital text became the only record of that knowledge. But because the digital record is proprietary, locked inside Anthropic’s data lake, the broader ecosystem loses access to that knowledge forever.
This is the opposite of what we, as the blockchain community, stand for. The core value proposition of decentralized storage networks like IPFS and Arweave is that data, once stored, becomes immutable and accessible to anyone. Anthropic’s approach is the antithesis: data is extracted, concentrated, and made inaccessible. The company effectively burned a library of Alexandria for the sake of a model benchmark improvement.
Trust is verified, not given. During the 2022 Terra/Luna collapse, my mathematical model demonstrated that the death spiral was not a black swan but a deterministic outcome of the peg maintenance logic. I published the post-mortem, and regulators cited it. Why? Because the data was on-chain, and anyone could replay the logic. Project Panama has no such verifiability. The only internal documents leaked to 404 Media describe the scanning pipeline but do not include cryptographic proofs of data integrity. If Anthropic truly believed its methods were ethical, why not publish a hash of the scanned dataset on a public blockchain, allowing researchers to verify later that the destruction was necessary and that the digitization was accurate?
The Contrarian Angle
Now, let me address what the bulls got right. There is a legitimate argument that scanning and destroying physical books reduces long-term storage costs and prevents unauthorized redistribution of the digital copies. If Anthropic had simply digitized the books and returned them to the sellers, the physical copies would remain, and the digital copies could be leaked or pirated, causing real financial harm to authors and publishers. The destruction, in this view, is a form of data security: it ensures that the only digital copies are those controlled by Anthropic, which can license them out later or keep them under strict access control.
Furthermore, the scale of the project suggests that Anthropic is aiming for a level of training data quality that goes beyond web-crawled text. The Claude models, particularly the long-context versions, require coherent narratives spanning hundreds of thousands of tokens. Physical books—with their chapter structures, character arcs, and logical argumentation—are superior to random web pages for training such models. If the destruction of 1 million books leads to a breakthrough in AI reasoning that benefits humanity, the trade-off might be defensible from a utilitarian perspective.
Elon Musk’s public criticism—announcing that SpaceXAI would use 'non-destructive scanning' and donate the rare books to digital libraries—also carries an element of competitive positioning. Musk’s xAI is a direct competitor to Anthropic, and his moral posturing is as much a marketing move as it is an ethical stance. The market should treat such criticism with the same skepticism we apply to any rival’s claims.
Logic outlives the hype cycle. The contrarian argument does not hold up under scrutiny. First, the 'data security' defense is hollow because the scanned data itself could be leaked. Anthropic has no bulletproof internal security; a disgruntled employee could copy the data lake to a USB drive. The destruction of physical copies only guarantees that outsiders cannot reconstruct the digital corpus independently. It does not prevent Anthropic’s own internal leak. Second, the utilitarian calculus ignores the irreversibility of cultural loss. A rare book destroyed cannot be re-scanned later with better technology or shared with other researchers. The decision to destroy is a permanent act of centralization. Third, Musk’s promise, while self-serving, highlights a feasible alternative: non-destructive scanning with digital preservation. If SpaceXAI or any other company can achieve the same training data quality without destroying physical books, then Anthropic’s destruction was unnecessary.
From my experience as a senior quant analyst during the DeFi Summer, I learned that the most dangerous narratives are those that wrap a plausible technical benefit around a fundamentally unsound ethical core. Compound’s yield farming incentives were mathematically unsustainable, but the market ignored the actuarial models because the short-term APYs were seductive. Similarly, Anthropic’s improvements in model performance might be real, but the cost—irreversible loss of cultural data—is an externality that the market is not pricing in.
The Regulatory and Market Implications
This incident will accelerate the demand for verifiable data provenance in AI. Just as the SEC’s enforcement actions against unregistered securities have forced crypto projects to implement KYC/AML compliance, Project Panama will push AI companies to adopt on-chain attestations of training data origin. I expect to see a new category of 'data notarization' services, where AI companies publish cryptographic commitments (hashes) of their training datasets on Ethereum or a dedicated L2, allowing independent auditors to verify the source and integrity of the data without revealing the full dataset.
This is a direct extension of the forensic wallet clustering techniques I developed. Instead of tracking token flows, we will track data flows: which publisher contributed which books, which scanner hardware was used, and whether the digitization process was non-destructive. Blockchain provides the only immutable, public ledger suitable for such attestations.
Every error has a signature. The signature of Project Panama is the absence of a public hash. Anthropic’s silence on the matter—no official statement, no apology, no corrective action—is itself a data point. The company is betting that the public’s attention span is short and that the next AI breakthrough will overshadow the controversy. History suggests otherwise. The Terra/Luna collapse was ignored by mainstream media for three months until the death spiral became undeniable. By then, the damage was done. I am tracking this situation with the same on-chain tools I used to monitor the Luna foundation wallets. The difference is that here, the 'chain' is not a blockchain but a supply chain of physical-to-digital transformation. The principles of verification, however, remain the same.
Takeaway
The Anthropic book-burning scandal is not an isolated incident of corporate overreach. It is a stress test of the industry’s ability to self-regulate data provenance. If the AI community continues to tolerate opaque, destructive data acquisition methods, it will invite the same type of heavy-handed regulation that crypto faced after the 2021 bubble. The blockchain toolkit—cryptographic hashes, decentralized storage, immutable timestamps—offers a path forward. The question is whether any major AI company has the courage to adopt it before another library burns.
