Acquisition Versus Training

Most coverage of Bartz v. Anthropic collapsed two different questions into one: may you train a model on a book? and may you get the book from a pirate library? The case answered them separately, and the $1.5 billion attaches to only one of them (Authors Guild summary).

• Training on a book you bought. Judge Alsup held that using legally acquired books to train a model was fair use.
• Getting the book from a pirate library. Anthropic was denied summary judgment on the piracy claims. This is what the settlement bought peace on.

So the settlement prices acquisition, not training. For a reader, that is the teachable part: where the data came from carried the cost; what was done with it did not, in this case. It means provenance — keeping a record of how each piece of training data was obtained — is where the money risk sits.

There is a second limit that is easy to miss. The Authors Guild notes that the fair-use holding, as decided, applies to the three named plaintiffs, not to the whole class. It is not a general licence, and reading it as "training is fair use, settled" overstates it.

Why it matters for the token price

If training itself were the costly act, every model trained on books would carry a liability spread across its tokens. If acquisition is the costly act, the cost depends on how the data was obtained — bought, licensed or taken — and a model built on bought or licensed data carries none of this particular penalty. The same book can be free to train on and $3,000 to have pirated. That difference is set on this rung, before any token is served.

Key Claims

• The court treated training on lawfully acquired books as fair use, and did not grant Anthropic summary judgment on the piracy claims. Evidence: moderate (a single secondary summary of the rulings by an advocacy organisation; the rulings themselves are not held in this KB) (Authors Guild)
• The $1.5bn priced the acquisition, not the training. Follows from the structure above. Evidence: moderate (Authors Guild)
• The fair-use holding is not a general licence. As decided it applies to the three named plaintiffs. Evidence: weak (the Authors Guild's reading; a party-adjacent commentator, not the court) (Authors Guild)

Open Questions

• Does the acquisition/training split hold in other courts and other jurisdictions? This is one case, one defendant.
• Does it apply to data that is neither bought nor pirated — scraped web text, for instance? The source does not say.

Related Concepts

• The price of training data — the $3,000-per-title figure this split explains

Backlinks

Pages that reference this concept:

• The price of training data
• Anthropic
• The Authors Guild