I am standing in Stanford University's Green Library with a large box of books at my feet. These are community cookbooks, rare in the sense that all are from small, local organizations and are part of short print runs. I have been here for about an hour in front of the Book Eye, a high-quality book scanner that nonetheless requires me to manually flip a page and push a button every few seconds to get my scan. The process is tedious, my feet are going numb, and I have only scanned two books. The numbers game is not looking good for me. I have about fifty more books to scan and, I hope, eventually more, for the project I am working on, digitizing and parsing these hyper-local cookbooks to preserve them and make them accessible for anyone who may want to use them.[1]  With no grant or fellowship money to support my time standing in front of the book scanner with the laser light flashing across my vision, I grow increasingly demotivated. 

Later, I complain to a friend who has more experience in the Digital Humanities. She mentions something called destructive scanning, a process where the spine of a book is removed, and its individual pages are fed through a high-speed scanner. We decide that it won’t work for me. These books are too rare. I want to keep them intact. Nevertheless, the prospect lingers in my mind as a possibility. 

On January 27th of this year, The Washington Post released an article entitled, “Inside an AI start-up’s plan to scan and dispose of millions of books.” The catchiness of this title is paralleled only by the opening quotation, an excerpt from a leaked internal document at Anthropic: “Project Panama is our effort to destructively scan all the books in the world” (Schafer et al.). 

The turn of phrase is villainous to an absurd degree. What could it mean that Anthropic wants to destructively scan all of the books in the world? If the backlash to this leak suggests anything, it is that the public is interpreting the project as a kind of symbolic, if not literal, book burning. 

From the perspective of a humanities scholar, the backlash, outrage, and public ire directed at this research method are concerning and a distraction from much more pressing issues at play in the Project Panama case. The negative scrutiny turned upon destructive scanning and, relatedly, copyright laws that benefit libraries and archives poses a threat to literary preservation and literacy, in that literary research and the public's access to literature can depend, at times, on these scanning methods. 

In what follows, I make a case that, however controversial, the destruction of books is a valuable means of preservation of our literary history and a tool for literary research—if used correctly. I would encourage turning attention away from the destructive process itself and toward  the unethical hoarding of information, the concerning usage of the books for the purpose of feeding Anthropic’s large language models (LLMs), and the abuse of copyright laws put in place to protect the public's access to knowledge. 

Misinformation and Public Backlash Around Destructive Book Scanning

At some point, during the tedious process of manually scanning my cookbooks, it dawns on me that most of them are bound in the cheapest way possible: with a comb binding. This means I can easily remove the binding, feed the book through the scanner, and reattach the binding later. I tell my friend I’ve decided to give destructive scanning a shot. In preparation for putting the books through the scanner, I remove the bindings carefully. Some of them are brittle from old age; others snap painfully across my knuckles as I extract them. When I’m done, I’m left with a neat pile of loose papers, ready for scanning. 

On Instagram, one user, @abbyfromcle, with tears in her eyes, comments on a video of the spines being chopped off of what appears to be an antique book. “Books being destroyed in any way is horrifying to me. Like, truly, this video makes me genuinely physically ill,” she says. “I think that books are some of the most sacred things that we have as humans, and this is equivalent to book burning for me. This is 1984. This is Fahrenheit 451” (@abbyfromcle). 

This video of a “hydraulic press” cutting the spines off of old books has accompanied many of the critiques across social platforms of Anthropic's Project Panama. The claims, specifically, that Anthropic is destructively scanning “rare” books, have been frequently paired with this particular video, perhaps understandably so. At a surface level, the video is alarming. 

The original video, however, has no connection to Anthropic and was, in actuality, first posted by The Book Rescuers, an antiquarian bookseller that restores books bound for the landfill and resells them at reduced prices. Their hydraulic press video, which garnered criticism when they originally posted it in May 2026 (long before The Post's article on Anthropic) depicts them cutting the spines off of modly books that are, as they put it, unfit for rehoming. They save what scraps of the books they can and recycle the rest. 

This video began to circulate as evidence of Anthropic's destruction of old books after an article  from news.com.au posted the clip as an example of a hydraulic press book cutter.[2] Though they misattribute the video to Canada's ABTec Solutions, a company that sells non-destructive book scanners, they do specify that the video does not come from Anthropic.

This instance of telephone, in which a video posted by a news outlet becomes misattributed by means of being reposted so many times, is not an uncommon phenomenon. But even if the video of the book guillotine does not belong to Anthropic, the question remains as to whether cutting the bindings off of books is necessary at all, either for scanning them or, in the case of The Book Rescuers’ video, for recycling them. Couldn’t, for example, The Book Rescuers have saved these books, as their name implies? 

The Book Rescuers explain in the follow-up post to their spine-cutting video that the books they cut and recycled were molded and beyond repair. Though commenters still protest, preservation practices at the Library of Congress support this decision. Library of Congress specifies that “Recovering collections from a mold outbreak is time-consuming and costly. Mold can literally destroy a collection or item,” and that “If the mold-affected collection item is considered replaceable, replacement should be considered.” (Library of Congress).

While it may be tempting to view disposing of books as sacrilegious in all instances, reputable booksellers often do not sell books that have mold, mildew, or severe damage to the spine.[3] Mold rehabilitation, in extreme cases, requires the skills of an experienced conservator, skills best reserved for rare or out-of-print books. 

Which Books Are Worth Preserving? 

As to whether this video is representative of what Anthropic is doing with the books they have gathered, it is unclear, based on the existing documents, whether and what percentage of the books that they have purchased (an amount purported to be upwards of 2 million) have been rare, out of print, or antiquarian (Milmo). It is less likely that they have scanned antiquarian books (books over a hundred years old) as these are typically expensive and would likely not have been sold cheaply by used booksellers. There is a whole, separate market for these specialty books. An unnamed spokesperson for Anthropic claimed via an email to The Guardian that "None of our data acquisition programs buy and destroy 'rare' or 'antiquarian' books" (Milmo). Nonetheless, given the volume of the books destroyed, it is likely that they have destroyed books that might be designated rare, or at least books that are out of print. 

But the designation for the category "rare" can, as the Library of Congress reports, be a slippery one. For the purposes of the Library of Congress, an item is considered rare if it was published before 1801, the reason being that the introduction of new print technologies in the 19th century rapidly increased the quantity of books being printed. Ever since this revolution in printing technologies, the number of books being printed every year has been steadily rising. For example, the number of books being printed in the US grew from 2,600 in 1869 to 4,500 in 1890, according to the Smithsonian (Boudreau). In 1911, Publishers Weekly reported that in the year 1910, 11,671 new titles were printed in the US (112). Now, Publishers Weekly reports that about four million traditionally published new titles are released in the US per year (Millot). For this reason, rarity is not the only metric for whether a book should or shouldn’t be destroyed in the process of scanning. For example, a book may not be rare in that it exists in fewer numbers, but may be out of print, meaning that those copies in circulation are more endangered than a book that doesn’t have many copies in circulation, but that is still in print. The matter of issues and editions also factors into these designations. 

The photo that The Post released of the Anthropic warehouse—a real photo of the books they have been destroying—depicts books that appear to be relatively new, mass market paperbacks, and is relatively devoid of books that seem, at the surface level, rare. But whether or not they’re common or unique, the question remains: need they have been destructively scanned?

Image

What is Destructive Scanning and Why Do It? 

My friend helps me use the scanner at Stanford’s Center for Spatial and Textual Analysis (CESTA). Now that I’ve carefully removed the comb bindings from my books, I can feed the book into the top of the scanner, which flips through its pages so rapidly that they blur together into an arch of whirling paper and land, if a bit askew, back in the rightful order in the tray below. The scanner then creates a PDF file and uses optical character recognition (OCR) to generate a text file. This is ideal for me—I need both file types for my research, and Book Eye only outputs PDFs. Destructive scanning, I realize, not only saves me scanning time, but also time processing and cleaning my data. 

Another Instagram user, @kaburbankechoing, echoing the concerns of many, asks, “Having these pages cut with a hydraulic press is not necessary to scan the pages of the book, right? I mean, with modern technology we could find a way to do this without ruining the books, but they’re doing it because it's faster. They’re destroying the books because of a profit motive, right?” (@kaburbank). Taking a step back, for a moment, from the fact that this particular instance of destructive scanning was done by Anthropic, I would like to attend to the question of whether it is worth it to “ruin” the books because it is faster to run them through a document scanner (a scanner that scans loose pages) rather than with something like a Book Eye. The answer, in the case of mass market paperbacks, is, for a myriad of reasons, yes. Yes, at least, for the purposes of research or preservation. 

To understand why, it is worth spending some time going over book scanning methods. The method that most are familiar with, when it comes to scanning books, is likely using a copy machine, a process which, I’m sure no one would disagree, is tedious. Libraries and archives tend to use something a little more specialized, but which still requires human labor, something like the Book Eye, or like the slightly nicer book scanner that the Internet Archive uses to scan their books while preserving their integrity. These book scanners use a book cradle to hold the book open without putting too much stress on the spine.[4] Specialized foam blocks can also be used to provide further support to especially fragile books or manuscripts.

On the other hand, destructive scanning typically uses what might most accurately be called document scanners, scanners which are meant to process loose-leaf documents. For this type of scanning, the spine is cut off with something called a guillotine paper cutter and the covers removed. The pages are then fed through the scanner and come out in the bin below in the same order. The process of scanning itself does not destroy the pages. It does not shred or incinerate them, as some posts have implied. 

Notably, document scanners are often much cheaper than specialized book scanners like the Book Eye and can make creating scans cheaper and more accessible for smaller organizations with less funding. Additionally, the scans can come out at a higher quality and with better OCR, whereas overhead scanning is prone to messier scans and missing or wrong information in the OCR. 

Destructive scanning is most useful for libraries and archives that want to provide digital access to copies of their books, but that do not have the funding or staffing to scan each of the books' pages by hand. In this instance, it may be cheaper and less labor-intensive for a library to buy two copies of the mass-market paperback, making one their physical copy and digitizing the other, than it would be to spend the time and labor required to scan each individual page. In such instances, the library would have to destroy the scanned copy of the work, rather than allow both the physical copy and the digital copies to be checked out at once. 

Destructive scanning is also useful for Digital Humanities scholars who, for the sake of their research, need to analyze huge corpora of texts, corpora that, as an individual researcher, it would take far too long to scan each individual book. I talked to one such researcher, Carmen Thong, Assistant Professor in Digital Humanities at the National University of Singapore. Thong received her PhD at Stanford and was an active member of the English Department's Literary Lab. Her work involves “literary text mining” in order to analyze and interpret literary texts using computational methods.[5] For Thong, whose research focuses on contemporary novels in English from postcolonial countries, destructive scanning is the only way for her to perform her research without breaking the law. She explains that most of the books she studies “are not digitized/not available for text-mining due to copyright.” This is because publishing companies protect their ebooks from data mining through a variety of legal and technical means. 

As opposed to physical books, which are only protected by copyright, “ebooks are subject to license agreements which restrict the ways in which they may be used, and often specifically disallow text-extraction” (Warner et al.). Since Thong works in the contemporary (the late 20th century to the present) her research often depends on her ability to scan physical books. “Destructive scanning (of non-rare books)” she explains, “is the only way I can do so [conduct research] legally without having to spend months turning pages.” 

The Copyright Concern With the Anthropic Case 

Thong's comments pick up on the lesser-attended-to, but more concerning, feature of the Anthropic case: its intersection with copyright law, particularly fair use laws as they relate to books and book distribution. The primary reason that Anthropic opted to begin scanning physical books was to supplant the need for pirating books from shadow libraries such as LibGen. After settling a  lawsuit leveled by authors for illegally downloading titles, Anthropic has switched to scanning physical books in order to train its models. As The Washington Post reports, Judge William Alsup “found that Anthropic was within its rights to use books for training AI models because they process the material in a ‘transformative’ way” (Schafer et al.). In this way, Anthropic is leveraging fair use copyright laws, the same set of laws that protect libraries' and researchers' abilities to access these materials. For example, researchers similarly recycle the pages of the books after a destructive scan. Thong explains, “[t]he loose stack of pages then has to be recycled, because the researcher has technically bought only rights for one copy.” This is the same right to an individual copy that allows Anthropic to scan and utilize the contents of a book, granted that they only keep the physical copy. 

For Thong, the public outrage and backlash are not surprising: “The fact that they are used to feed AI exacerbates the sense that human culture is being physically and manifestly destroyed… deleted... to feed a machine that supposedly seeks to automate our self-expression.” She remains concerned, however, about ire being turned toward the method, rather than the bad actors utilizing it:  

[T]hose who have not heard of destructive scanning might be aghast at the idea of books being cut-up. These are all valid reactions. However, my first thought was worry about how destructive scanning will be removed as an avenue of fair use for researchers because of the actions of for-profit AI companies. I would hope that any policies that might be put in place would create a distinction between fair use for researchers and for-profit companies. (Thong)

If public opinion turns against the same laws that protect researchers' ability to do their work and libraries' ability to offer access to books, it could bode poorly for that public's access to the very books and knowledge that they hold so dear. 

In actuality, destructive scanning comprised only a small part of the copyright lawsuit leveled against Anthropic. Perhaps the more striking, though less attended to, feature of the case, is the fact that the judge ruled training the models "transformative." 

In the court documents, the language that is used both to justify the model as "transformative" and to argue against the idea that it would supplant the place of the original works it was trained on, is explicitly humanizing. The language likens the LLM to an author:

In short, the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative. Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different. (United States District Court, 13-14) 

In this section, the LLM is a "reader aspiring to be a writer," as though its processing abilities function in the same way as a learning human mind. Perhaps more damning, however, is the section that deems the output "original works of authorship."

But Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works. This is not the kind of competitive or creative displacement that concerns the Copyright Act. The Act seeks to advance original works of authorship, not to protect authors against competition. (United States District Court, 28) 

Such an understanding of both the training and the output of an LLM—as though it were a human child or an aspiring author—raises a myriad of questions around the definition of authorship as it relates to copyright and fair use. These passages imply that LLMs both learn and create in the same way that people do, and that their relationship to the material on which they are trained is comparable to that of a person. In actuality, to put it simply, LLMs are probability machines, relying on statistical likelihood of the next word occurring in a sequence in order to "write." They also write exponentially faster than humans can. 

Rather than focus in on those fair use laws which protect researchers, educational institutions, and libraries' rights to digitize one copy of a book, it may be more useful to question how fair use law defines "authorship" and "autonomous creation"—statuses it was meant to maintain the sovereignty of—in an age when these words are becoming increasingly detached from the human. 

Though I do not destroy any books in the process, destructive scanning is what allows me to digitize most of the books in my collection and has proven to be a saving grace for my research. The steps of removing the binding and feeding the pages through a scanner are essentially the same, except that the spines of my books can be reattached. Luckily, I will not need to destroy these originals to do my work. Ironically, it will be “destructively scanning” these books that saves them. 

Notes

[1] For more on why scanning community cookbooks is an important act of preservation see Choice Receipts from American Housekeepers: A Collection of Digitized Community Cookbooks from the Library of Congress.

[2] This is evident as many of the subsequent posts which misatribute the video to anthropic bear the news.com.au watermark. 

[3] Both Abe Books and Half Price Books specify that their lowest grade (poor) includes books with physical damage or water damage. They do not sell books with mold damage. 

[4]A book cradle is a piece of supportive furniture used to hold a book at an angle so as not to put pressure on its spine. Used to read, display, or, in this case, scan books. Particularly useful for protecting delicate books. 

[5] Literary text mining, text mining, or text data mining (TDM) involves mining the text for data. Here, data can refer to a myriad of things such as individual words and phrases or punctuation marks, to name only a few. 

Works Cited 

@abbyfromcle. Instagram, 28 July, 2026, https://www.instagram.com/reel/DbWt0XfCYIW/?utm_source=ig_web_copy_link&igsi=MzRlODBiNWFlZA==.

AbeBooks. "Guide to Book Conditions." AbeBooks, Accessed 24 Aug. 2026, https://www.abebooks.com/books/rarebooks/collecting-guide/understanding-rare-books/guide-book-conditions.shtml. 

Alison P. Kelly. “Choice Receipts from American Housekeepers: A Collection of Digitized Community Cookbooks from the Library of Congress.” The Public Historian, vol. 34, no. 2, 2012, pp. 30–52. JSTOR, https://doi.org/10.1525/tph.2012.34.2.30. Accessed 24 Aug. 2026.

The Book Rescuers. TikTok, 6 Apr. 2024, https://www.tiktok.com/@thebookrescuers/video/7353761958010866990.

The Book Rescuers. TikTok, 22 Apr. 2024, https://www.tiktok.com/@thebookrescuers/video/7361127641128144171.

Boudreau, Joan. The book boom: “Early bookbinding inventions.” National Museum of American History, 1 Oct. 2015, https://americanhistory.si.edu/explore/stories/book-boom-early-bookbinding-inventions 

Half Price Books. "Condition Guide." Half Price Books, Accessed 24 Aug. 2026, https://www.hpb.com/condition-guide.html.

@kaburbank.. Instagram, 28 July 2026, https://www.instagram.com/reel/DbVs_7XN5w_/?utm_source=ig_web_copy_link&igsi=MzRlODBiNWFlZA==.

Library of Congress. “If Mold is Present.” Library of Congress, Accessed 24 Aug. 2026, https://guides.loc.gov/wet-collections/mold 

Millot, Jim. "Book Output Topped Four Million in 2025." Publishers Weekly. Mar 17. 2026, https://www.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/99943-book-output-topped-4-million-in-2025.html

Milmo, Dan. “Secondhand booksellers in UK and Ireland suspect AI firms behind ‘strange’ bulk orders” The Guardian. 15 Aug. 2026, https://www.theguardian.com/technology/2026/aug/15/uk-ireland-booksellers-suspect-ai-companies-bulk-orders-data-acquisition 

Publishers Weekly. “Books of 1910.” Publishers Weekly, 28 Jan. 1911, pp. 112–114, babel.hathitrust.org/cgi/pt?id=uc1.b4172267&seq=121.

Schafer, Anjelica, et al. “Inside an AI start-up’s plan to scan and dispose of millions of books.” The Washington Post. Jan 27. 2026, https://www.washingtonpost.com/technology/2026/01/27/anthropic-ai-scan-destroy-books/

Schaffer, Aaron & Oremus, Will & Tiku Nitasha. "Inside an AI start-up’s plan to scan and dispose of millions of books." The Washington Post, 27 Jan. 2026, https://www.washingtonpost.com/technology/2026/01/27/anthropic-ai-scan-destroy-books/ 

Thong, Carmen. Personal interview. 2026.

United States District Court, Northern District of California. Bartz et al. v. Anthropic PBC. June 27 2025. No. C 24-05417 WHA. https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_2.pdf 

Warner, Matt & Nomura, Nichole & Thong, Carmen & Keener, Alix & Sherman, Alexander & Birch, Gabi & Kurzynski, Maciej & Algee-Hewitt, Mark. (2025). “Slow, Painful and Expensive: Current Challenges in Text-Mining Corpus Construction for the Digital Humanities.” Digital Humanities Quarterly. 19. 10.63744/g9ey5yeptah4.