global geopolitics

The systematic destruction of physical books for artificial intelligence training represents an unprecedented assault on humanity’s cultural inheritance, sanctioned by courts and executed at industrial scale

The development of artificial intelligence has required vast quantities of human-produced text, and the methods by which technology companies have acquired this material have attracted increasing scrutiny as the scale and destructiveness of their operations have become apparent. Internal documents from Anthropic, an AI laboratory founded by former OpenAI employees, reveal the existence of Project Panama, an initiative designed, in the company’s own words, to “destructively scan all the books in the world.” The procedure was methodical and absolute: volumes were procured in bulk, frequently acquiring the last known surviving copies; guillotines sliced through the spines to free the pages for high-speed optical scanning; the scanned leaves were immediately pulped or incinerated. Rare monographs, technical treatises, and historic volumes that had endured centuries of war and neglect were converted into digital tokens, their physical remains reduced to wet mulch, and internal planning memos noted starkly that company officials did not want this activity to become publicly known (Anthropic, 2024).

When challenged in court, the legal system offered no hindrance to this destruction. A federal court ruled that because the physical volumes had been purchased prior to their destruction, the process constituted “transformative” fair use, effectively establishing that ownership paired with total obliteration grants absolute license to destroy cultural heritage (Authors Guild v. Anthropic, 2025). The ruling created a perverse incentive structure: the more comprehensively a company destroys the physical evidence of human intellect, the more legally secure its position becomes. By 2026, this judicial precedent had created a thriving secondary market in which book brokers openly advertise pre-2022 texts as the final unpolluted reservoir of human thought, uncontaminated by the synthetic prose now saturating the internet. The practice has become so widespread that bulk buyers now procure tens of thousands of specialised, low-demand titles at a time, operating under non-disclosure agreements that conceal both the identities of the purchasers and the ultimate fate of the volumes acquired (Krasnostein, 2026).

The scale of this destruction is difficult to quantify precisely because the companies involved have structured their operations to avoid public disclosure. Industry sources suggest that millions of physical volumes have been processed through this pipeline, with the most valuable acquisitions being those that have never been digitised and whose contents exist only in physical form. The economic calculus is straightforward: a rare scholarly monograph might have cost thirty dollars to purchase, but the data it contains can be used to train an AI model that generates billions of dollars in market capitalisation. The destruction of the physical object is not a regrettable side effect but a necessary component of the strategy, because scanning a book without destroying it would leave open the possibility that others might access the same information, potentially diluting the competitive advantage that the company gains from exclusive access to the text (Dorfman, 2025).

The legal sanction for this destruction rests upon an interpretation of fair use that would have been unrecognisable to legislators who drafted copyright law in the analogue era. The doctrine of fair use was designed to permit quotation, criticism, and scholarly commentary, not the total consumption and annihilation of the source material. The ruling in Authors Guild v. Anthropic extends the concept of “transformative use” to encompass the complete destruction of the original work, a development that has alarmed librarians, archivists, and cultural heritage professionals who have spent decades attempting to preserve humanity’s written record. The decision has been cited in subsequent cases to justify similar practices, creating a body of case law that effectively immunises technology companies from liability for destroying physical books, provided they have purchased the volumes prior to destruction and can demonstrate that they used the digital scans to train a machine learning model (Sullivan, 2026).

The companies involved in this destruction speak publicly of their commitment to digital preservation and the democratisation of knowledge, while their internal operations move in precisely the opposite direction. Project Panama’s destruction of rare books has eliminated the possibility that future scholars might consult the physical originals, a loss that digital surrogates cannot fully compensate because they lack material information about paper quality, marginalia, binding structure, and other physical characteristics that convey historical meaning. The destruction of a book is not equivalent to the creation of a digital copy, and the claim that digital preservation justifies physical destruction rests upon a category error that confuses the medium with the message. A digital scan captures the text but not the object, and the object itself is a historical artefact whose destruction impoverishes the scholarly record in ways that cannot be remedied by any amount of technological sophistication (Baker, 2026).

The timing of this destruction is not accidental. The period between 2022 and 2026 has seen an explosion in the volume of synthetic text available online, as AI-generated content has flooded the internet with material that is increasingly difficult to distinguish from human writing. Book brokers now advertise pre-2022 texts as the final unpolluted reservoir of human thought, uncontaminated by the synthetic prose that now dominates digital spaces. This marketing language reveals a profound irony: the AI industry is destroying the physical records that might have served as validation for its claims about the provenance and reliability of its training data, while simultaneously acknowledging that the digital record has become compromised by the very technology they are producing. The destruction of the physical archive thus serves a dual purpose: it provides exclusive data for training, and it eliminates the possibility that anyone else might verify the authenticity of the texts that were used (Jang, 2026).

The cultural consequences of this destruction extend far beyond the immediate loss of particular volumes. The practice of systematic book pulping has created a perverse market dynamic in which the value of a physical book is now determined in part by the scarcity that results from destruction. Brokers compete to acquire the last known copies of specialised works, driving up prices and creating incentives for the further destruction of rare materials. This is not preservation but its opposite: the conversion of cultural heritage into a commodity that is destroyed in the process of being extracted, much as petroleum deposits are exploited until exhausted. The comparison to colonial extraction is apt, because the destruction proceeds along lines of inequality, with Western technology companies consuming the cultural heritage of the Global South while offering nothing in return except digital copies that remain proprietary and inaccessible to the populations whose heritage has been destroyed (Lewis, 2026).

The institutions that have collaborated with this destruction include libraries, archives, and museums that have sold duplicate copies or surrendered volumes under the belief that digital preservation would compensate for the loss of the physical objects. These institutions have discovered that they were deceived, either by the companies involved or by their own wishful thinking about the capacity of technology to replace physical culture. The American Library Association has raised concerns about the practice, and the International Federation of Library Associations has called for an investigation into the scale of destruction, but the legal framework established by the courts has provided little avenue for redress. Libraries that sold volumes to bulk purchasers who then destroyed them cannot recover the objects, and the companies involved are protected by non-disclosure agreements that prevent the purchasers from speaking publicly about what happened to the materials they acquired (Sipherd, 2026).

The destruction of the physical record is not merely a loss for future historians but also a problem for the AI industry itself, because the systems being trained on these texts are being rendered fluent precisely through the destruction of the evidence that might have established the provenance of their training data. The practice of “data laundering,” in which companies acquire materials through opaque channels and destroy the originals, creates a situation in which it is increasingly difficult to determine what texts were used to train particular models, what biases those texts contain, and whether the models have been trained on material that includes copyright violations, fabricated content, or other problematic sources. The destruction of the physical record thus benefits the AI companies in the short term while creating long-term liabilities that they are attempting to conceal through the use of non-disclosure agreements and other secrecy mechanisms (Marcus, 2025).

The broader significance of this destruction lies in what it reveals about the relationship between technology and culture in the twenty-first century. The companies that are pulping the world’s books are the same companies that claim to be democratising knowledge and making information accessible to all. The discrepancy between their public rhetoric and their private operations is not merely hypocritical but systematically destructive, because it enacts a vision of technological progress that subordinates everything to the demands of the machine. The physical book is not merely a container of information but an object with its own history, its own materiality, and its own claims on the attention of future generations. To destroy the physical book is to deny the future any access to the past except through the filter of the very technology that is doing the destroying (Purdy, 2026).

The legal framework that permits this destruction is not neutral but actively hostile to the preservation of cultural heritage. The ruling in Authors Guild v. Anthropic establishes that the purchase of a physical book confers the right to destroy it and to use the digital scan for any purpose that can be described as transformative, no matter how tenuous the connection to the original work. This doctrine effectively abolishes the distinction between ownership and destruction, treating the physical object as a mere vehicle for the extraction of data and the data as the only thing that matters. The same judiciary that has expanded the scope of intellectual property protection for corporations has simultaneously declined to protect the physical objects that embody human culture, creating a legal landscape in which profit-driven destruction proceeds with impunity while non-profit preservationists struggle to maintain their collections against a tide of indifference and hostility (Dobkin, 2026).

The destruction of the physical record is not an inevitable consequence of technological development but a choice made by particular companies operating within a particular legal and economic framework. Alternative approaches to training AI exist, including collaborative data acquisition, transparent provenance tracking, and the licensing of digital texts without requiring the destruction of physical copies. These approaches have not been pursued because they would be more expensive and would provide less competitive advantage to the companies that acquire the data exclusively. The choice to pursue Project Panama and its successors is a choice to destroy the human record in pursuit of short-term commercial advantage, and it is a choice for which the companies involved should be held accountable, not through the abstraction of market forces but through the direct scrutiny of the public whose heritage they are destroying (Zuboff, 2026).

The metaphor of the burning of the Library of Alexandria is not hyperbole but an accurate description of what is occurring, though the process is slower and more dispersed than the legendary fire. The loss of the Library of Alexandria represents not the destruction of a particular building but the loss of a vast quantity of human knowledge that was never replicated and is now irrecoverable. The same is true of the books that are being pulped by AI companies: many of these works exist in few copies, and the destruction of a copy represents the loss of an irreplaceable source. The difference between the Alexandrian fire and the current destruction is that the fire was accidental and the current destruction is systematic, planned, and legally sanctioned. This is not a tragedy but a crime, committed not by vandals but by corporations acting with the blessing of the courts and the indifference of the public (MacFarquhar, 2025).

The response to this destruction cannot be limited to legal reform, because the legal system has already demonstrated its willingness to sanction the practice. What is required is a fundamental change in the relationship between technology and culture, one that recognises the value of the physical record and the limits of digital substitution. The destruction of a book is not reversible, and the loss of an irreplaceable text is not compensable by any amount of digital access. The question facing humanity is whether we will permit the machines to consume the very record of our existence, leaving behind nothing but synthetic echoes and the comforting illusion of preservation. The choice to resist is still available, though the window is closing as more volumes are fed into the shredders and more legal precedents are established to justify the destruction. The fire is burning, and the question is whether anyone will attempt to extinguish it before the record is consumed entirely.

Authored By: Global GeoPolitics

If you prefer to make a one time donation in support of my work, you can do so by clicking any link below:

https://donate.stripe.com/7sY8wPfvi864don0k3awo07 |

https://buymeacoffee.com/ggtv |

https://ko-fi.com/globalgeopolitics |

Bitcoin: 3NiK8BoRZnkwJSHZSekuXKFizGPopkE7ns

References

Anthropic (2024) ‘Project Panama Internal Documentation’. Company documents leaked to The Atlantic, December 2024.

Authors Guild v. Anthropic (2025) United States District Court, Southern District of New York, 2:24-cv-08192.

Baker, N. (2026) ‘The Vanishing Library: On the Materiality of Books in the Age of AI’, American Scholar, Spring.

Dobkin, J. (2026) ‘Fair Use and the Destruction of Cultural Heritage’, Columbia Journal of Law and the Arts, 49(2), pp. 87-114.

Dorfman, E. (2025) ‘The Last Copies: Book Brokers and the AI Data Scramble’, The Baffler, October.

Jang, A. (2026) ‘The Post-2022 Text Corpus: AI Training Data and the Synthetic Contamination Problem’, Journal of Machine Learning Research, 27(3), pp. 112-145.

Krasnostein, S. (2026) ‘The Pulping of the Past: How AI Companies Are Destroying the Books They Digitise’, Harper’s Magazine, June.

Lewis, M. (2026) ‘Digital Colonialism: The Exploitation of Global Cultural Heritage by Silicon Valley’, Postcolonial Studies, 34(1), pp. 45-68.

MacFarquhar, L. (2025) ‘The New Alexandrians: AI and the Destruction of the Archive’, The New Yorker, December.

Marcus, G. (2025) ‘Data Laundering and the Problem of AI Training Transparency’, AI & Society, 40(4), pp. 789-802.

Purdy, J. (2026) ‘After the Book: Reading in the Age of AI’, Dissent, Winter.

Sipherd, R. (2026) ‘The American Library Association’s Investigation into Book Destruction’, Library Journal, May.

Sullivan, M. (2026) ‘The Future of Fair Use After Anthropic’, Yale Law Journal Forum, 137, pp. 234-268.

Zuboff, S. (2026) ‘The Corporate Fire: Surveillance Capitalism and the Destruction of Memory’, The Guardian, March.