Anthropic agreed to pay $1.5 billion, about $3,000 for every book in its training pile. The reason shows where the law on AI training data has landed.
A rule is forming from the first rulings and settlements, and it’s more specific than either side predicted. Training a model on copyrighted material can be fair use. But how you got the material, and whether the model’s output competes with it, can turn the same act into huge liability.
The ruling that drew the line
In Bartz v. Anthropic, Judge William Alsup issued the split decision that’s now the template: training an AI on copyrighted books can be transformative fair use, but downloading and storing pirated copies to build the training library is not. The case then settled for $1.5 billion, one of the largest copyright settlements in US history.
The same line runs through Kadrey v. Meta. There the court accepted fair use for the training but kept alive the claims about pirated works obtained by torrent.
Why courts are sceptical
The labs leaned on one argument: training is transformative because the model learns patterns and doesn’t store works. Courts have been sceptical for three reasons.
Global, and heading towards licences
In November 2025 the Munich Regional Court held OpenAI liable in GEMA v. OpenAI for memorising and reproducing copyrighted song lyrics. It was among the first major European rulings against an AI developer on training data. Meanwhile England’s High Court rejected a secondary copyright claim in Getty v. Stability.
NYT v. OpenAI continues. The plaintiffs are building evidence that pirated books were downloaded from the LibGen shadow library and stored, the Bartz theory again. It’s widely expected to settle.
There are more than 70 copyright suits filed against AI companies, and the industry is moving from courtrooms to contracts. Warner Music’s suit against Udio ended in a partnership rather than a verdict.
What this means for you
A few years ago, pulling a corpus of books off a torrent was a quick, free way to get training data. Nobody priced the risk. Today it’s a class action with statutory damages. In Anthropic’s case it settled at about $3,000 a book, a price high enough to change how the next corpus gets built.
Build models? “It was publicly available” is closing as a defence. Licensed or lawfully acquired data is the only kind that survives a challenge over verbatim copying.
Buy or fine-tune someone else’s model? Put training-data provenance on your due-diligence list. It sits beside security, uptime and the political risk every AI roadmap now carries, because liability can travel downstream. Ask the vendor where the data came from, and treat “the open internet” as a risk.
Frequently asked questions
Is it legal to train AI on copyrighted material?
It can be. In Bartz v. Anthropic, a court found training on copyrighted books can be transformative fair use. Getting and storing pirated copies to build the training set is not. It depends heavily on how the data was obtained and whether the output reproduces or competes with the original.
Why did Anthropic pay $1.5 billion?
The settlement, about $3,000 per work and among the largest copyright settlements in US history, resolved claims that Anthropic downloaded and stored pirated books as training data. The court had found that piracy fell outside fair use, even though training itself could qualify.
What is the test courts are applying?
They weigh four things. Is the use truly transformative? Does the model reproduce large parts of the work? Is the use commercial? Does it compete with the original's market? A US Copyright Office report notes that competing with existing licensing markets tilts the analysis against the AI company.
What does this mean for companies using AI?
Where training data came from is now a due-diligence question. Builders should use licensed or lawfully acquired data. Buyers of AI models should ask vendors where the training data came from, because liability for unlicensed data can carry downstream.
Three-minute lesson on this What AI actually is Marketing course, basic. Free, no sign-up.