AI companies are shredding rare books
TL;DR Highlight
AI firms bulk-buy books, scan and shred them; a judge ruled it fair use.
Who Should Read
A developer working with LLM APIs or fine-tuning who worries about where training data comes from and its copyright risk
Core Mechanics
- Books are cut, scanned, then shredded
- AI companies bulk-buy rare books and run them through high-speed machines that cut the spines off.
- The originals are shredded after scanning.
- ISBNdb brokers orders up to a million books
- The service keeps buyers anonymous.
- The post claims it offers NDAs as a feature and coaches clients to call it "digital preservation".
- Pre-2022 books carry a premium
- They are valued because they are free of AI-generated text.
- The post presents this as the reason for the demand.
- A judge ruled it fair use
- The reasoning cited is that destroying the original leaves only one copy at a time.
- The author expects the practice to accelerate because of the ruling.
- Anthropic hired a Google Books veteran
- The hire is the former head of Google Books partnerships.
Community Reactions
- Doubts that the books are truly rare
- Commenters said the targets are mostly recent books still under copyright, not medieval manuscripts.
- Several asked for evidence or titles of rare books actually destroyed.
- It's just the legal process
- One view was that US copyright law produced this and the judge followed the law.
- Another view was that the law should allow scanning owned books without shredding.
- Cost structure comments
- One commenter said buying an old book for $5 and scanning it for $25 beats paying licensing fees.
- They added that libraries also discard old books with no demand, which keeps costs low.
- A reprint business's practice
- A commenter running a reprint business said they cut bindings and scan.
- They keep the cut pages sealed and the scans on magnetic disks, and said no AI company has asked to train on them.
How to Apply
- List the sources of your training and RAG data
- Tabulate each corpus with its source and license basis today.
- Flag entries whose origin you don't know.
- Check the legal basis of book-derived datasets
- If you use book data, ask the supplier in writing how it was acquired and under what terms.
- If the supplier is anonymous or NDA-only, talk to legal first.
- Tag pre-2022 data separately
- Add collection-date metadata to your own data.
- Keep pre-2022 material as a separate set for evals where AI-text contamination matters.
Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.