Fair use rulings shift AI copyright risk to data sourcing

Two federal judges found that training on books can be fair use, but pirated copies and market harm stay open. Data provenance is now a balance-sheet item.
A federal judge has ruled that training on books is fair use and still sent Anthropic to trial over more than seven million pirated copies. In Bartz v. Anthropic on 23 June, Judge William Alsup held that training was exceedingly transformative and that scanning print books Anthropic had bought was fair use too, but he refused that protection for the pirated central library, where willful infringement carries statutory damages of up to $150,000 per work.
The legal risk of training, first tested in the artists' lawsuit, has moved from the model to the dataset. Training itself now has a ruling in its favor, which lowers the tail risk that hung over every lab valuation. What replaces it is a visible price for shortcuts. How a company obtained its data can decide liability, and statutory damages scale with the number of works.
Meta won its own case on 25 June, when Judge Vince Chhabria granted it summary judgment against thirteen authors over Llama, writing that this does not make the training lawful, only that the authors made the wrong arguments. Within three days, two federal judges in California had issued the first US rulings on training language models with copyrighted books. Chhabria also opened a second front, market dilution, the claim that a model flooding a market with similar works harms the originals. Together the rulings reward labs that bought or licensed their data and leave exposure with those that took it from shadow libraries, so diligence on any model company now covers data provenance alongside compute and talent.
Teams that train or fine-tune on lawfully acquired material stand on firmer ground than a week ago, provided they can document where every dataset came from and prove it. That requirement is a market in itself: licensed data with a clean chain of title, provenance tracking, and audit tools that let a buyer or an insurer verify a training set. Chhabria further held that authors are not entitled to a market for licensing their works as training data, so data businesses should sell quality, coverage and convenience rather than legal permission alone.
District rulings set the direction, and three open matters will set how far it reaches. The trial over Anthropic's pirated library, and any damages figure, will put a price on past shortcuts. Whether either ruling goes to the Ninth Circuit will show how the appeals court weighs market dilution against transformation. The cases against other labs over news, music and images weigh most, because there the outputs compete more directly with the originals and the dilution argument is stronger.
Copyright has turned from an existential question for AI into a cost that depends on discipline, and a court has now drawn the line between fair use and liability at the way data was obtained. Data that is bought, licensed or a company's own is what Fanaura wants to see behind a model, and it asks founders to show that chain of title early. A clean dataset is now an asset, and a pirated one is a liability with a price per work.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building data licensing or provenance tools for teams that train models? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.