Part 1 - Too Small to Sue, Too Small to License, Ingested all the Same
Part 2 - The Answer Without the Author
Part 3 - The Machinery of "No"
Part 4 - Where you Got It. Where it Hurt. Little Else is Settled
The first three posts in this series described a problem: original work is ingested without consent or credit, AI answers increasingly satisfy the reader in place of the source, and the standards arriving in response defend the boundary without restoring the visit. The fair question to ask next is whether the courts have settled any of it.
Three years of litigation and one record settlement later, the honest answer is that the headline question is still open. No court has decided whether training on copyrighted work is lawful. But the rulings so far point somewhere rather more specific, and rather more useful. The courts are not really being asked whether machines may learn from text. They are being asked two much narrower questions: where did the material come from, and can anyone prove that the output harmed the market for the original? Almost everything decided so far turns on one or the other.
Where the material came from
On 20 July 2026 a federal court granted final approval to the $1.5 billion settlement in Bartz v. Anthropic: roughly $3,000 for each of some 500,000 books, the largest copyright settlement in United States history, together with an obligation to destroy the pirated files the library had been built from.
The figure is the part that gets reported. The shape of it is the more interesting half. The underlying ruling had split the question cleanly in two: training on lawfully acquired books was fair use, while keeping a permanent library assembled from shadow-library copies was not. Anthropic was penalised for how it obtained the material, not for training on it. And the release is narrower than the headline number suggests, covering the acquisition and copying of those works up to August 2025 and nothing else. Claims about outputs, and about anything that happens next, survive intact.
Take first, settle later
That sequence is worth naming, because it was not an accident and it has been said out loud. Speaking to an AI class at Stanford in 2024, the former Google chief executive Eric Schmidt described the play for an AI startup with some candour: tell the model to build your competitor and "steal all the music", get the thing in front of people, and "hire a whole bunch of lawyers to go clean the mess up" if it works. He asked afterwards for the video to be taken down, and said later that he had not meant it literally.
Walked back or not, it remains a fair description of the incentive that has operated for the past three years. Clearing rights up front costs real money immediately, and against that sits a ruling some years away that is uncertain, discountable, and may never arrive at all. Bartz is the first occasion on which that bill has actually been presented. Whether $1.5 billion is large enough to change the calculation is a genuinely open question.
Whether anyone can prove harm
The second question is where most of the cases have actually turned, and it is the one that attracts the least attention.
In Kadrey v. Meta the authors lost. It would be easy to read that as a court blessing AI training, and it was widely reported in something like those terms. The judge said something rather different and considerably more pointed: they lost because they had not proved market harm, not because training is lawful. He went further, sketching a market-dilution theory that a better-argued case could win. In July 2026 he declined to let the authors take that question up on appeal early.
Thomson Reuters v. Ross failed for the mirror-image reason. There the copying produced a product that competed directly with the original, and a fair-use defence did not survive contact with that fact.
Harm, in other words, is the hinge. It is also genuinely difficult to evidence, which is why so much depends on who is able to construct the argument well.
Nothing has been settled on appeal
It is worth being precise about how provisional all of this remains. No appeal court has ruled on any of it.
The first to hear argument did so on 11 June 2026, when the Third Circuit took up Ross. The panel spent most of its time on precisely the two questions above: whether the use was transformative, and whether it harmed the market, including the market for licensing material as training data. No decision has issued as we write.
The flagship case is no further along. New York Times v. OpenAI has produced an order to hand over 20 million de-identified ChatGPT logs, a summary judgment motion still pending, and no trial date. Getty's UK case failed on its central copyright theory in November 2025 and is now under appeal, with a parallel US action refiled in California. Until an appeal court speaks, every ruling in this area is a district-court data point rather than a rule.
A harm argument built for academic publishing
One filing from last month matters more to our own readers than any of the above. On 14 August a group of textbook authors sued OpenAI, following a parallel action against Meta in July, and their theory of harm is structurally different from the novelists'.
Textbook adoption is an institutional decision rather than a consumer one. It is made by a department, a committee, a course lead. Which means that a substitute does not have to be as good as the book. It only has to be good enough for the people doing the choosing. That is a considerably easier harm to demonstrate than a lost novel sale, and it is the first argument we have seen built around how academic publishing is actually bought rather than how trade publishing is.
What this means for smaller publishers
To summarise where this leaves things. Whether AI training is lawful in the abstract remains undecided, and will stay undecided until an appeal court speaks. What the cases actually turn on is provenance and provable harm.
For researchers, NGOs and cultural archives, that has a practical edge to it. Clean records of what you hold and where it came from are becoming legally load-bearing in a way they simply were not two years ago. The question "can you show how you acquired this?" is moving out of the territory of good practice and into the territory of evidence.
It is worth noticing, too, who is not in any of these cases. Every plaintiff named above is a large publisher, a well-resourced trade body, or an organised class with counsel behind it. Three years in, the long tail we described in the first post of this series is still not in the room.
Next in this series
Next, we will turn to the incentive trap: why an offer that appears to solve all of this for a small publisher is usually the one worth reading twice.