Training AI on Copyrighted Books: It's the Download That Bites
The Anthropic ruling said AI training on books was lawful — and still cost $1.5 billion. The liability was in how the files were obtained. Here's what that means for small teams.
Is it legal to train AI models on copyrighted books? After the Anthropic case, the honest answer is that the training was ruled lawful and the downloading was what cost $1.5 billion. That split is the entire story, and most coverage flattened it into "AI companies lose."
TechCrunch's piece Is it legal to train AI models on copyrighted books? It's complicated lays out the cases. I want to talk about what they mean if you are building something small here with a laptop, a free tier, and a scraper.
🔍 The ruling everyone read backwards
Judge William Alsup ordered Anthropic to pay a $1.5 billion copyright settlement to writers whose books were used in training. The part that got lost: Alsup ruled the training itself was lawful. He compared an LLM learning from books to a writer studying literature. The penalty attached to something else entirely — Anthropic sourced books from illegal shadow libraries.
| Case | Decided by | What actually happened |
|---|---|---|
| Anthropic copyright settlement | Judge William Alsup | $1.5B settlement, but training ruled lawful; the liability came from pirated sourcing |
| Thomson Reuters v. Ross Intelligence | Judge Stephanos Bibas | Not fair use — Ross trained on Reuters content to build a directly competing legal platform |
| Thaler v. Perlmutter | — | Works that are 100% AI-generated are not copyrightable |
Key takeaway: Your legal exposure is not in the model architecture. It's in your data supply chain — where the bytes came from, and whether you can prove it.
⚖️ What a court is actually weighing
Copyright law in the US hasn't been meaningfully updated since 1976, which is why every one of these questions ends up decided case by case. The factors judges weigh, per the TechCrunch piece, include the purpose and nature of the work, the amount used, and the effect on the market — with market impact carrying the most weight.
| Factor | The question it asks | Where small teams get caught |
|---|---|---|
| Purpose and nature | Is the use transformative, or just a repackage? | Wrapping someone's content in a thin UI is not transformative |
| Amount used | How much of the original was taken? | Full-corpus scrapes are hard to argue down |
| Market impact | Does it substitute for the original? | Building the thing your data source sells |
IP attorney Cathy Gellis put the mechanism plainly:
"Copyright law hinges on copying, but it doesn't hinge on using the work or experiencing the work, consuming the work, reading the work."
That is why Alsup could bless the training and still punish the acquisition. Reading is not the infringement. Obtaining an illegal copy is.
🛠️ The provenance habit that costs you nothing today
The practical lesson for anyone fine-tuning a model, building a RAG index, or shipping a dataset: log where every file came from, at the moment you get it. Reconstructing that two years later, under pressure, is impossible.
A manifest entry per source costs about thirty seconds:
{
"source": "https://example.gov.lk/reports/2025-annual.pdf",
"obtained": "2026-08-24",
"method": "direct download, public URL",
"licence": "Sri Lanka government publication, no stated restriction",
"sha256": "a3f1...",
"robots_txt_checked": true
}
Four rules I'd hold anyone to:
- Never pull from a shadow library. This is the single fact pattern that produced a nine-figure bill. Free is not the same as legal.
- Record a checksum per file so you can prove which version you trained on. Our hash generator does SHA-256 in the browser — nothing gets uploaded.
- Prefer licensed and public-domain corpora. Project Gutenberg, government publications, Creative Commons, and datasets with an explicit licence field.
- Check
robots.txtand terms before scraping, and write down the date you checked.
Warning: "I got it from a HuggingFace mirror" is not provenance. If the upstream dataset was assembled from pirated books, inheriting it does not launder it.
📄 The other half nobody plans for: what you can sell
Thaler v. Perlmutter held that fully AI-generated works are not copyrightable. If you're freelancing — logos, marketing copy, landing pages, boilerplate code — that's not academic. It's the question of what your client is actually paying for.
The awkward follow-on, which the TechCrunch piece flags directly: nobody has settled how you'd prove how much AI was involved, or what percentage of human input makes a work protectable.
| Deliverable | Human contribution | Practical risk |
|---|---|---|
| Prompt in, image out, shipped as-is | Near zero | Client may own nothing enforceable |
| AI draft, substantially edited and art-directed by you | Meaningful | Much stronger position |
| AI-assisted code you architected, reviewed, and tested | Meaningful | Standard practice, low concern |
My read: keep your working files. Drafts, revision history, the notes where you rejected three versions. That record is the evidence of human authorship, and it's free to keep.
🌐 Why US rulings matter from Colombo
None of these decisions bind a Sri Lankan court. I'm an engineer, not a lawyer, and I'm not going to pretend otherwise. But they matter here anyway, for reasons that have nothing to do with jurisdiction:
- Your clients are often abroad. A US or EU company commissioning work will push provenance and IP-warranty terms down to you in the contract.
- The vendors you build on are US companies. What Alsup and Bibas decide shapes what OpenAI, Anthropic and Google will and won't ship, and what their terms of service allow you to do downstream.
- Licensed data is becoming a paid market. Attorney Jason Henderson framed it as: "Copyright is always about protecting and growing the market." Expect more datasets to carry a price tag rather than a takedown notice.
- Competing with your source is the risky shape. Ross lost because it trained on Reuters content to build a Reuters competitor. If you fine-tune on a company's data to sell against that company, fair use gets much harder to argue.
💡 What this means for you
If you're building on a learning budget, the good news is that the expensive mistake in this story is also the easiest one to avoid. Alsup didn't say training on books is illegal. He said helping yourself to pirated copies is.
So: keep a manifest, checksum your sources, prefer licensed and public-domain corpora, and don't fine-tune a model to compete head-on with the people whose data you used. Keep your drafts so you can show human authorship in what you deliver.
Bottom line: Train on what you're allowed to have, and keep the receipt.
None of that requires a lawyer or a budget. It requires a habit, and the cheapest time to start it is on the project you haven't collected data for yet.