The learning was legal.
The shortcut cost $1.5 billion.
A federal judge just gave final approval to the largest copyright recovery in American history. Read the ruling carefully, because the headline hides the lesson: the court said training AI on books is fair use. What cost $1.5 billion was the pirate library used to get them. The bill wasn't for the learning. It was for the sourcing.
A ruling that split the baby
The court drew one line with two sides. Training an AI model on books the company had lawfully acquired: fair use, transformative, protected. Downloading and warehousing seven million pirated books to skip the acquisition step: infringement, no shelter, headed for a jury trial on damages that could have reached hundreds of billions of dollars. Anthropic settled rather than roll those dice.
Fair use
Training on lawfully obtained works was ruled transformative and protected. The capability itself was never the crime.
Infringement
Building a permanent library from pirate sites fell outside fair use entirely. The datasets must now be destroyed.
One more thing the headline hides: this was a settlement, not a verdict. It sets no binding precedent, and the same question is still grinding through the courts against Meta, Google, OpenAI, and others. This was the first bill. It will not be the last word.
What actually got priced
Provenance. For years, "we'll sort out the rights later" was the industry's unwritten data strategy. This settlement wrote the price on it: three thousand dollars a book, half a million times over, plus destruction of the data. Where your vendor's model came from, what it was trained on, and under what terms is no longer a philosophical question. It is a number on a court docket.
For the businesses buying AI, that moves one conversation from paranoid to prudent: the supply chain question. Serious providers can now answer it in writing, and the biggest ones back their commercial terms with copyright indemnification. The vendors who wave it off are telling you where the risk lives: with you.
The questions to ask before you sign
You do not need a legal team to protect yourself. You need a vendor who will answer four questions in writing:
- Which AI models process our data, and who operates them?
- What are those models' commercial terms, and do they include copyright indemnification for our use?
- What data was our vendor's own tooling trained or built on, and can they state its provenance?
- If a court order forced a model or dataset out of service tomorrow, what is our continuity plan?
If a vendor answers all four without flinching, keep talking. If they can't, you just learned something worth more than the demo.
Full disclosure: we build on frontier models ourselves, including Anthropic's. This is our supply chain, which is exactly why we read the filings. We use these tools under commercial terms, with named providers, on licensed data, and we put those commitments in our contracts, not just our marketing. When the ground shifts in this industry, we want to be the first to know, because our clients are standing on it with us.
The 5 percent who win with AI know their supply chain. A Recon maps yours: every tool, every vendor, every question they should have already answered.
Start with a Recon