The largest copyright settlement in history was approved this summer, and almost everyone drew the wrong conclusion from it. A federal judge signed off on Anthropic paying $1.5 billion to authors — roughly $3,000 per work across about 500,000 books. Read the headline and you would assume the courts had ruled that training AI on copyrighted books is illegal, and that the industry had just been handed an enormous bill for it.
The opposite happened. Anthropic won the argument that mattered most to the industry. It paid anyway. Understanding why is the key to the entire legal landscape now forming around AI — and it turns on a distinction almost nobody outside the courtroom is making.
The ruling that split the question in half
In June 2025, Judge William Alsup issued a decision that did something unusual: it separated two things everyone had been treating as one. The first question was whether training a large language model on copyrighted books is fair use. Alsup said yes — training on books acquired legally is fair use, a finding widely read as a landmark win for AI developers.
The second question was where the books came from. And here Anthropic lost badly, because a large portion of its library had been downloaded from pirate sites. The court found that acquiring and storing over seven million pirated books was not protected, because obtaining them that way was not necessary to train the model — and deciding later not to train on the pirated copies did not erase the liability for having assembled them. Alsup put it bluntly: “Anthropic had no entitlement to use pirated copies for a central library.”
So the billion-dollar liability did not come from building an AI. It came from the shortcut taken to feed it. That distinction — provenance rather than purpose — is now the fault line running through nearly every case in this space.
Why this is a strange kind of victory for everyone
Look at what the outcome actually rewards and it starts to feel less like justice and more like a toll booth. The legal system has arrived, roughly, at this position: you may train on the entire corpus of human writing without asking permission, provided you buy your copy first. The intellectual act — ingesting a book, extracting its patterns, building a commercial product on what it taught you — is fine. The procurement is what gets you.
For authors, that is a deeply ambivalent result. They received real money, but on a theory that implicitly blesses the practice most of them actually object to. A writer who is upset that a model absorbed their life’s work is not consoled by a ruling that says it would all have been fine had the company spent $18 on a paperback. The payout was for the piracy, not the training — which is why many creators viewed the settlement as a loss dressed as a win.
For AI companies, the message is narrower than the headline suggests but expensive all the same: the models are probably legal, and the data-acquisition practices of the 2021-to-2023 land grab probably are not. Every lab that scraped shadow libraries in the early scramble is carrying an unpriced liability from a period when nobody thought anyone would check.
The $1.5 billion that settled nothing
Here is the part that deserves far more attention than it has received, because it is genuinely strange. The largest copyright payout ever recorded establishes almost no law.
Alsup’s fair-use ruling was a single district court decision, and by settling, Anthropic ensured the case will never reach an appeals court to become binding precedent. Other judges remain entirely free to reach different conclusions on their own facts. And at least one already has: in Thomson Reuters v. Ross Intelligence, a Delaware court found that training on copyrighted material was not fair use where the resulting tool competed with the source — a case now on appeal before the Third Circuit.
This is what makes the settlement so revealing as a strategic move rather than a legal one. A billion and a half dollars bought certainty for one company and preserved uncertainty for everyone else. The favorable ruling stays on the books as persuasive authority; the risk of an appellate court overturning it disappears. Whether or not anyone intended it that way, the effect is that the most consequential fair-use question of the decade remains formally unanswered — and the litigation continues, with Hachette, Elsevier, and Cengage suing Google over books used to train Gemini in the same month the settlement was approved.
What actually decides this
If you want to predict where this lands, watch the licensing market rather than the courtrooms. Fair use analysis weighs the effect of the use on the market for the original work — and that factor is not fixed. It moves as the market moves. When no legitimate way existed to license books for AI training, a court could reasonably find that training displaced nothing. But as a real licensing market emerges, the argument inverts: if you could have paid for the data, taking it for free starts to look like lost revenue rather than a novel use. Analysts tracking these cases expect exactly this dynamic to make fair use harder to argue over time.
That is the quiet mechanism by which this resolves, and it will not look like a dramatic ruling. Each licensing deal signed makes the next unlicensed use slightly less defensible. The frontier labs, which can afford to license, are already moving that way — which conveniently raises the cost of entry for everyone behind them. The likely endgame is not that AI training gets banned or blessed. It is that it becomes a normal, expensive input cost, available to whoever can pay for it, with the early free-scraping era treated as an embarrassing adolescence that a few companies paid handsomely to put behind them.
For anyone actually building with this technology, the practical takeaway is unglamorous and immediate: provenance is now the compliance question. Not “is training legal” — courts have mostly said it is — but “can you prove where every piece of this came from.” That is a documentation problem, not a philosophical one, and it is the part that turns into money in a courtroom.
Related reading: Is AI a Bubble or a Revolution? The Honest Answer Is Both
