Over the past week, a viral post claimed that Anthropic had “secretly destroyed all the books in the world” to train Claude.
The post spread from English into Spanish and other communities. According to a snapshot cited in the source draft, the original X post had reached roughly 32,000 likes, 14,500 reposts, and more than 1,000 comments when captured. The replies split into two predictable camps: one side called it a dark moment for big tech, while the other dismissed it as another outrage loop driven by missing context.
This article does not defend Anthropic, and it does not join the pile-on.
Instead, it separates the viral claims into three categories:
- confirmed facts
- exaggerated or simplified claims
- unresolved questions
The goal is practical: after reading this, you should not need to guess which part of the thread was real and which part was compressed for outrage.
What the viral post claimed
The viral thread centered on several claims:
- Anthropic internal documents used the phrase “Project Panama is our effort to destructively scan all the books in the world.”
- Another internal line reportedly said Anthropic did not want outsiders to know it was working on the project.
- Anthropic previously downloaded around 7 million books from piracy-linked sources to train Claude.
- After legal risk became harder to ignore, Anthropic hired a former Google Books executive to seek a lawful way to acquire books.
- Publishers did not provide the kind of licensing path Anthropic wanted.
- Anthropic then bought physical books, cut off their spines with hydraulic cutters, scanned the pages, and recycled or discarded the paper.
- One vendor quote allegedly described capacity to process up to 2 million books in six months, or roughly 11,000 books per day.
- Anthropic later agreed to a $1.5 billion settlement, often summarized online as “about $3,000 per book.”
- Some posts framed the story as authors receiving nothing except a destroyed copy of their book.
Some of this is real. Some is missing context. Some remains unverified.
Confirmed facts
There is a real case: Bartz v. Anthropic
The relevant lawsuit is Bartz v. Anthropic, filed in the U.S. District Court for the Northern District of California. The case was brought by authors including Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson.
In June 2025, Judge William Alsup issued a mixed ruling.
The court held that Anthropic’s use of books to train Claude and related LLMs was highly transformative and could qualify as fair use under Section 107 of U.S. copyright law. The court also held that converting lawfully purchased print books into digital files was fair use because Anthropic was replacing purchased print copies with digital copies for its central library.
But the court drew a different line around pirated books.
Anthropic had acquired millions of book copies from sources including Library Genesis, Books3, and Pirate Library Mirror / PiLiMi. The ruling did not protect that permanent central library of pirated copies as fair use.
That distinction is the core of the case:
Lawfully purchased print book -> destructive scan -> one digital replacement copy
is not the same legal category as
pirated digital copy -> permanent central library
Lawfully purchased print book -> destructive scan -> one digital replacement copy
is not the same legal category as
pirated digital copy -> permanent central library
The destructive scanning workflow is supported by court reporting
Multiple reports based on court filings describe a physical workflow:
- buy physical books
- remove bindings or spines
- scan pages with high-speed equipment
- discard or recycle the remaining paper
- keep digital files for the central library
Publishers Weekly summarized the court’s findings as Anthropic spending millions purchasing physical books that were “destructively scanned.” The Washington Post also reported on Project Panama based on unsealed court documents, though its full article may be paywalled for some readers.
Tom Turvey was part of the sourcing effort
Court reporting also says Anthropic hired Tom Turvey, formerly associated with Google Books partnerships, in February 2024. His role was tied to finding lawful ways to acquire books after legal risk around piracy-linked datasets became more visible.
This part of the viral story is not invented.
The piracy-linked settlement is real
Anthropic agreed to a $1.5 billion settlement over pirated book copies.
AP reported that a federal judge approved the settlement and that it covers more than 482,000 books, with roughly $3,000 per book expected before fees, costs, and allocation between rightsholders. The Authors Guild similarly explains that around 500,000 eligible titles were identified from roughly 7 million downloaded copies after accounting for duplicates and eligibility requirements.
This settlement is tied to pirated digital copies, not simply to the act of buying and scanning physical books.
Exaggerated or oversimplified claims
“They planned to secretly destroy every book on Earth”
The phrase “all the books in the world” appears in reporting about internal Project Panama documents. It is a real phrase reported from the documents.
But it should not be read literally as proof that Anthropic destroyed, or even could destroy, every physical book on Earth.
The known numbers do not support that interpretation.
The viral posts conflate:
- a broad internal project ambition
- a vendor capacity quote
- actual completed scanning volume
- the global number of existing book titles
A vendor capacity estimate of up to 2 million books in six months is not the same as proof that 2 million were processed, and it is far from “all books in the world.”
“Authors got nothing except shredded books”
This line mixes two different categories of books.
The books that were destructively scanned were physical books Anthropic reportedly purchased through sellers or distributors. The settlement, by contrast, relates to pirated digital book copies from unauthorized sources.
Rightsholders in the settlement are expected to receive money for eligible works associated with the pirated-copy claims. AP and the Authors Guild report figures around $3,000 per eligible title, subject to fees, claims, ownership splits, and other settlement details.
That does not resolve every ethical concern about destructive scanning. But the statement “authors got nothing” is not accurate as a summary of the settlement.
“Destroying the book made everything legal”
This is the most common simplification.
The court did not say “if you shred a book, copyright disappears.”
The court’s logic was narrower: if Anthropic lawfully purchased a print copy, destroyed that copy, and retained one digital replacement copy, the total number of library copies did not increase. The court treated that differently from keeping pirated copies that were never lawfully acquired.
You may disagree with that legal reasoning. Many authors and publishers do. But the reasoning is not simply “the shredder created a loophole.”
“A vendor destroyed 11,000 books per day”
The math is real:
2,000,000 books / 180 days = about 11,111 books per day
2,000,000 books / 180 days = about 11,111 books per day
But the source draft describes this as a vendor quote or capacity figure.
That is not the same as verified evidence that the vendor processed that many books every day in practice.
Still unresolved
Several important questions remain unanswered in public sources.
How many physical books were actually bought and destroyed?
Reports describe millions of physical books, but the exact total remains unclear in public summaries. Some court filings and vendor details may be redacted, sealed, or not easily accessible.
Were rare, out-of-print, or one-of-a-kind books destroyed?
This is the most emotionally charged question.
Recent reporting has raised concern from booksellers and rare book experts about AI companies buying and destroying physical books. But a verified list of rare, unique, or irreplaceable titles destroyed by Anthropic has not been established in the sources reviewed for this article.
That should be marked as unresolved, not assumed.
Does the settlement fully end the legal risk?
The approved $1.5 billion settlement is a major event in Bartz v. Anthropic, but it does not answer every AI copyright question.
It does not settle every lawsuit in the broader AI industry. It also does not create a final appellate precedent on whether all AI training uses of copyrighted books are lawful. Other courts may reach different outcomes on different facts.
Did Anthropic respond to every public criticism?
Public statements and court filings cover parts of the dispute. But many social-media criticisms, especially around rare books and physical cultural loss, do not yet have detailed official responses tied to specific book lists.
Fact-check table
| Viral claim | Classification | What the available evidence supports |
|---|---|---|
| Internal documents referred to destructively scanning “all the books in the world” | Partly confirmed, often amplified | The phrase is reported from internal documents, but it is a project framing, not proof of literal global destruction |
| Anthropic downloaded around 7 million pirated books | Confirmed | Court reporting and legal summaries describe millions of books from LibGen, Books3, and PiLiMi / Pirate Library Mirror |
| Destructive scanning made everything legal | Oversimplified | The court’s logic focused on lawfully purchased print copies being replaced by digital copies, not shredding as a universal defense |
| Anthropic paid $1.5 billion and authors got nothing | Misleading | The settlement covers piracy-linked claims and provides around $3,000 per eligible title before fees and allocation |
| A vendor destroyed 11,000 books per day | Unverified as actual performance | The number follows from a reported capacity quote, but capacity is not the same as confirmed daily output |
| Anthropic destroyed rare or one-of-a-kind books | Unresolved | Public concern exists, but no authoritative list was found in the reviewed sources |
A five-step method for checking viral AI posts
Use this checklist when an AI controversy spreads faster than the underlying documents.
1. Find the original source
Ask whether the post links to:
- a court filing
- an official statement
- a settlement website
- a primary document
- an article that reviewed primary documents
If it only links to another viral post, treat it as a lead, not evidence.
2. Circle every number
For each number, ask:
Is this an actual result, a plan, a quote, a capacity estimate, a legal class size, or a social-media metric?
Is this an actual result, a plan, a quote, a capacity estimate, a legal class size, or a social-media metric?
In this story, “7 million downloaded copies,” “500,000 eligible titles,” “2 million books in six months,” and “$1.5 billion” refer to different things.
3. Separate categories that the post merges
The most common misinformation pattern is not a totally fake claim. It is merging two real facts into one misleading story.
Here, the key distinction is:
- pirated digital books involved in the settlement
- lawfully purchased physical books involved in destructive scanning
Those are related, but not identical.
4. Use at least one independent source
For legal stories, use court documents, settlement websites, legal organizations, or credible news reporting. Social metrics do not prove factual accuracy.
5. Strip out emotional adjectives
Words like “disturbing,” “horrifying,” “secret,” or “dystopian” may describe a writer’s reaction. They do not prove the underlying claim.
Remove the adjectives and check what factual claim remains.
What this means for Claude API users
This case is about how Anthropic historically acquired training data and how courts treat copyright, piracy, fair use, and destructive scanning.
It does not change your current Claude API configuration.
It does not change your Base URL, model ID, API key handling, SDK setup, pricing logic, or day-to-day Claude API usage.
The practical lesson is about information hygiene.
AI-related posts travel faster than verification. A single viral thread can merge two separate facts into one stronger emotional story before you forward it to a customer, teammate, or investor.
ClaudeAPI’s role is not to defend any company or attack any company. The more useful role is to unpack these stories, separate what is known from what is exaggerated, and show what remains unresolved.
If you see another AI story that feels too clean, too angry, or too perfect for a hot take, use the five-step checklist before reposting it.
Sources and data notes
- U.S. District Court order in Bartz v. Anthropic, June 23, 2025, via Copyright Alliance PDF mirror: Bartz v. Anthropic Order
- Publishers Weekly: Federal Judge Rules AI Training Is Fair Use in Anthropic Copyright Case
- AP News: Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
- The Authors Guild: Bartz v. Anthropic Settlement: What Authors Need to Know
- The Washington Post: Inside an AI start-up’s plan to scan and dispose of millions of books
- The Guardian: Australian booksellers raise alarm over destruction of rare titles to feed AI
Social engagement numbers cited in the introduction come from the Chinese source draft’s X snapshot. They were not independently verified here and may have changed.
Disclosure
ClaudeAPI is an independent third-party technical service and is not affiliated with Anthropic, AP, The Authors Guild, Publishers Weekly, The Washington Post, or The Guardian.
