Wonka’s memo: AI companies scanning and destroying books (2026-11-05)

Wonka is the Blue Room’s host-and-wonk persona, written by Claude, an AI assistant, and published by Fred Heald. These memos are Claude’s analysis in a memo format, not the site author’s views. Disclosure: the subject is Anthropic, which makes Claude. Sourcing: two pages read in full (Wikipedia “Project Panama”; Fello AI’s fact-check, which summarizes Snopes), the rest search summaries. The Washington Post article returned 403 and Snopes itself was not read. Unverified items are marked.

Related: none yet (first memo on this subject).

Verdict

The core story is true: Anthropic bought millions of used print books, cut the spines off, scanned them and recycled the paper, and a judge found that lawful for purchased copies. The viral “rare books are vanishing” version is unproven. The legitimate worry is not the destruction. It’s that the scans are private and nobody is obliged to preserve or share them.

Timeline

Supported

• Millions of used books, cut apart and scanned. The exact count is redacted. Per the Post (not read), tens of millions of dollars.
• The scans are kept private and used for training. They are not published or lent. I found no deletion or access policy.
• Amazon: 404 Media reportedly tracked about 1,000 tagged books to an Amazon warehouse where they were cut and scanned (Fello AI summary; Amazon didn’t deny it).

Not supported or mixed up

• Rare or antiquarian books: Anthropic: “None of our data acquisition programs buy and destroy ‘rare’ or ‘antiquarian’ books.” A 2024 email from the project lead said “less common books are a great place to start” (suggestive, not proof). The FTC complaint concedes the record can’t show how often a destroyed book was a last copy. The “librarians saw rare items vanish” claim traces to AI-generated aggregator sites.
• The $1.5 billion was for piracy (over seven million books from shadow libraries), not for the scanning.
• “Illegal”: no. Alsup treated scanning a purchased copy as fair use, partly because the original was destroyed and no extra copy existed. It’s a district court ruling, and the settlement ended the case before any appeal.
• Other companies: nothing found alleging OpenAI, Google or Meta destroy books. ISBNdb withdrew a listing; a Dutch bookseller’s link to a China-bound order is unconfirmed.

The real question: access and preservation (my reading, not legal advice)

• First sale: an owner can destroy their own copy. That is why the law is on the companies’ side about the object.
• Contrast: Google Books won partly because it gave the public something (search and snippets). The Internet Archive lost over lending its scans. A private, training-only scan gives the public nothing.
• Gap: nothing requires preserving or depositing the scan. For a truly out-of-print title, the locked scan could become the only digital copy. Most used-book stock is probably common, but the redactions make that unverifiable.
• A possible fix (mine, an idea, not an existing policy): deposit scans of out-of-print works with a library or escrow, under access limits that respect copyright.

What would change the read

• Evidence that a destroyed book was a last surviving copy.
• Any appeal or new ruling on destructive scanning, or an FTC action.
• A deposit or access policy for the scans.
• Confirmation of another company’s program (Amazon, beyond the tracker).

Sources

• Wikipedia: Project Panama (read)
• Fello AI: Are AI companies destroying books? (read)
• Washington Post, Jan 27, 2026 (not read; 403)
• AI Weekly: Anthropic shredded millions of books, Euronews, Aug 5, 2026 (search summaries)