I ranked seven accounts of the same event by one test: how much of each survives being read next to the docket in Bartz v. Anthropic(N.D. Cal., 3:24-cv-05417). The window runs from Judge William Alsup's summary judgment order(23rd June, 2025) to the last week of July 2026, when a thread about shredded books and a reply calling that thread fabricated both showed up in my feed on the same day.
What I refused to count: aggregators recycling "Tom Harvey" for Tom Turvey, stories illustrated with AI-generated images, and any claim about rare or last-surviving books that does not name a title. That last rule strips most of the emotional weight out of the story, which is exactly why I applied it. I buy secondhand paperbacks and I ship software on top of Claude, so I have a foot on both sides of this and no interest in either side's tidy version.
Bartz v. Anthropic, 3:24-cv-05417
- Order: partial summary judgment, 23rd June, 2025, Judge William Alsup
- Held: training on lawfully purchased books is fair use, retaining a pirated library is not
- Piracy record in the order: Books3(196,640 books, early 2021), LibGen(at least 5 million, June 2021), PiLiMi(at least 2 million, July 2022)
- Settlement: 1.5 billion dollars, final approval 20th July, 2026, Judge Araceli Martínez-Olguín, roughly 500,000 works at 3,000 dollars gross per work
Alsup's own words on the purchased books: "every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy. The print original was destroyed. One replaced the other." Destruction was not a magic rite, it held the copy count at one, on top of first sale. He never held that destruction was required, and he denied summary judgment on the retained library copies because Anthropic, in his description, dodged discovery. This is one district order that was never appealed, so it binds no other court. To move up it needs the unredacted final count of books destroyed and the total spend, both still blacked out.
ANT_BARTZ_000004034
- Vendor: Datamation Imaging Services, 7700 Griffin Way Suite B, Willowbrook, IL, SOC 2 Type II facility
- Scope line: "Anthropic is seeking an experienced document scanning services vendor to convert from 500,000 to two million books over a six-month period"
- Window: January to June 2024
- Process: bindings removed, "utilizing a hydraulic powered cutting machine," barcode separator sheet per book, high-speed feeder, auto de-skew
- Output: 200 dpi greyscale, multi-page PDF, indexed by ISBN, author, title, stored on customer private cloud
The most useful number on this page is 200 dpi greyscale. Library preservation scanning runs 400 to 600 dpi in colour. What replaced the book is OCR fodder, worse than the object it displaced, and it sits on a private cloud where nobody outside the company will read it. The page is a proposal answering an RFP, not a receipt, and the range floor works out to ~2,750 books a day rather than 11,000. Two redactions matter: where books were delivered from, and where they were "returned to" after cutting. Pulling the surrounding correspondence off the docket would settle whether Datamation was engaged and at what volume.
The Washington Post
- Story: 27th January, 2026, based on more than 4,000 pages of unsealed filings
- Quotes: "Project Panama is our effort to destructively scan all the books in the world" and, on the soft codename, "we don't want it to be known that we are working on this"
- Hire: Tom Turvey, former head of partnerships for Google Books, February 2024, tasked with obtaining all the books in the world
- Named suppliers: Better World Books, World of Books, batches in the tens of thousands
Every other account in this list is downstream of this reporting, including the ones that misspell Turvey's name and thereby announce they are copies of copies. The secrecy line is the strongest fact anyone has produced, because it is a company arguing internally about optics rather than about law. Alsup's order quotes Dario Amodei preferring to avoid the "legal/practice/business slog" of buying. The Post would move up by publishing the unsealed exhibits themselves alongside the story, or by getting Anthropic on the record about whether Panama is still running.
404 Media and ISBNdb
- Story: 21st July, 2026, on ISBNdb, which sources printed books in bulk for AI labs
- Marketing language: pre-2022 print is "structurally clean of modern poisoning tools," and "'AI company destroys two million books' is not a headline that generates sympathy"
- Order sizes offered: 1,000 to 1,000,000 books per engagement, under NDA
- Bookseller signal: weekly sales up from fewer than 20 books to several hundred beginning in April, no pattern by author or subject
- Page offering sourcing "tailored to your LLM training needs": removed from the site
The incentive here is real and easy to state. The open web is contaminated with synthetic text, pre-2022 paper is the last clean reservoir, and a broker with an NDA lets the buyer stay unnamed. What the reporting does not establish is that any given buyer destroys what it receives, and CryptoSlate's review found no named rare, unique, or last-surviving title destroyed by Anthropic or an ISBNdb client anywhere in the public record. One buyer's name, or one invoice tied to a cutting vendor, would move this from an incentive story to a documented one.
Evan Luthra's thread
- Right: both Panama quotes, the Google Books hire, the hydraulic blade, the 1.5 billion dollar figure, the "slog" attribution
- Wrong denominator: 3,000 dollars per work applies to roughly 500,000 class works, not the 7 million pirated files, and attorneys' fees above 100 million dollars come off the top
- Left out: Books3 authors recover nothing, Alsup found its metadata too poor to identify titles
- Overstated: "every book on the planet" is internal corpus language about the secondhand market, and "11,000 books a day" uses the ceiling of a 500,000 to two million range
- Legally off: "if you scan a book and keep it, that's a crime" is not the holding, keeping both is the question Alsup expressly left open
The thread's best line, that a shelved book lasts centuries without power while the model it fed is discarded, is rhetoric rather than a sourced claim, and it is still the part I keep turning over. Growing up in Howrah I watched paper survive damp, blackouts and moving house, which is more than I can say for any drive I have owned. To move up, the thread needs the exhibit's Bates number and a docket link under it, the range restored, and the video swapped out.
"this would never pass PR"
- Claim: the behaviour may be real, the document was not used, and the post is engagement bait
- Test applied: a real company's communications team would never approve that wording
- Artifact in question: a vendor proposal on Datamation letterhead, marked Confidential, with redaction bars and a Bates stamp
The heuristic is backwards for anything that comes out of discovery. Unsealed internal material is blunt precisely because PR never saw it, which is why "we don't want it to be known that we are working on this" survives as a quote at all. The instinct behind the reply is not silly: a source did feed journalists an AI-generated fake internal document, DramaAlert circulated a misattributed scanner clip on 29th July, 2026, and X pays creators on impressions from verified users, so bait has a price. As such, the charge is reasonable in general and fails against this particular page. I have not matched the screenshot to the docket myself, and that is the check that closes it: post the CourtListener entry for ANT_BARTZ_000004034, or a reverse-image hit showing the page predates the case, and the accusation stands up.
Treventus Mechatronics
- Machine: ScanRobot 2.0 MDS, Treventus Mechatronics GmbH, Austria
- Method: 60 to 100 degree V-shaped cradle, prism capture, automatic page turning
- Rate: up to 2,500 pages per hour, marketed for historic books with no overstretching
- Relevance: the machine shown in the thread's own video destroys NOTHING
The outsider in this list is a vendor whose product is the road not taken. Google Books scanned without cutting and gave the public snippet search, Turvey came from that programme, and Anthropic's chosen path gave the public a private PDF nobody will ever see. Elon Musk said he had asked his AI team to scan rare books "the hard way vs just cutting off the spine," which is a claim, not a practice anyone can audit yet. Treventus would move up in this ranking the moment it publishes whether Anthropic or any lab ever requested a quote, and at what per-page cost against Datamation's cutting line, because that number is the whole argument about why the blade won.