For years the AI industry has repeated one defensive line: a model learns from text the way a person does — read a million books, extract patterns, copy nothing. It's a convenient metaphor, and courts haven't fully bought it. Over the past two or three years a wave of lawsuits worked its way through American courts and started answering the question for real, and the answer turns out to be sharper and more specific than the "AI accused of theft" headlines suggest. The line judges are actually drawing sits somewhere other than where most people expect to find it.
Who's suing whom
The defendant list reads like a roll call of the entire industry. The New York Times sued OpenAI and Microsoft back in late 2023, arguing that millions of the paper's articles went into training corpora without a license, and attaching examples where the model reproduced chunks of text nearly verbatim. A group of authors — plaintiffs included names on the order of Sarah Silverman, Ta-Nehisi Coates, Michael Chabon — filed parallel class actions against OpenAI and separately against Meta, arguing their books ended up in datasets like Books3, assembled from shadow libraries. Getty Images is suing Stability AI over the use of hundreds of thousands of photos from its archive to train Stable Diffusion, with a smoking-gun exhibit: the Getty watermark surfacing on generated images. Universal Music and other major labels filed suit against Suno and Udio, the AI music-generation services, accusing them of training on copyrighted recordings at industrial scale. And off to the side of the flashier generative-AI stories runs Thomson Reuters v. Ross Intelligence, over a legal-AI startup that trained a model on court-decision headnotes.
This isn't one question wearing different costumes. It's several distinct legal theories, and the first rulings have started to pull them apart.
The first real loss came from an unexpected corner
The AI industry's first serious fair-use defeat didn't happen in a splashy ChatGPT or Midjourney case — it happened in the quieter Thomson Reuters v. Ross Intelligence, over a legal-tech startup that trained a case-law search model on Thomson Reuters' proprietary headnotes summarizing court opinions. In early 2025 a federal judge ruled largely for Thomson Reuters, rejecting the fair-use defense as applied to that use. The decisive thread in the reasoning: Ross was building a direct competitor to Thomson Reuters' own product — the use wasn't "transformative" in the sense fair use requires, but a substitute on the same market.
The ruling covered a narrow, non-generative case — but it set a precedent: "we're just training a model" isn't an automatic shield when the output competes head-on with the source.
Anthropic: one case, two very different verdicts
The most instructive case is Bartz v. Anthropic, a class action by book authors over the use of their works to train Claude. Judge William Alsup split the question into two parts, and that split is essentially the line this article is about.
Part one: the training itself, run on legitimately acquired books — Anthropic had purchased physical copies and scanned them — the judge ruled fair use. The reasoning: training a model on a lawfully bought book is transformative, much like a person reading a book to become a better writer — the book itself is never reproduced to the model's downstream users.
Part two proved fatal for the company. Alongside the purchased books, Anthropic's training corpus included millions of books pulled from pirate shadow libraries like LibGen and Books3. Here the judge drew a hard line: downloading pirated copies is a separate infringement, not shielded by fair-use doctrine, regardless of what happened to those files afterward. It didn't matter that the model was "just learning" — the library had already been stolen before any training took place.
The outcome was unprecedented: in the fall of 2025, Anthropic agreed to settle this suit for roughly $1.5 billion — by the reported figures, on the order of $3,000 per book, across several hundred thousand works. That's the largest monetary recovery in the history of U.S. copyright litigation — a sum no company had previously paid out for infringement in one go.
Meta: same dispute, different judge, same cautious outcome
A parallel case, Kadrey v. Meta, had a similar group of authors accusing the company of using the same shadow libraries to train Llama. Judge Vince Chhabria, also in mid-2025, likewise ruled fair use for this particular case — the plaintiffs hadn't convincingly shown direct market harm from specific model outputs. But the judge explicitly warned this wasn't a blanket pass: he singled out the "market dilution" theory — the idea that a model trained on millions of books in a given genre could flood the market with similar content and depress demand for the originals — calling it potentially serious but insufficiently argued in this particular case. In other words: it lost not because the argument is weak, but because it was made poorly.
Our record: in the Maat system, what is taken must be balanced, and what is done must be seen. Training on a purchased book is an exchange, novel in form but not in substance — money reached the author through the sale of a copy, and the scale isn't tipped. Training on a copy pulled from a shadow library is taking without balance: the work's Ka, its essence and form, extracted and appropriated, while the creator sees no compensation and often never even learns the exchange happened at all. The court in the Anthropic case, in effect, recognized exactly this difference — not philosophically, but literally, in dollars per book.
Where the line actually sits
Put these cases together and what emerges isn't one bright line but two separate axes the press usually collapses into one.
The first axis is how the data was obtained. Buying a book and training a model on it is, for now, protected. Downloading millions of books from a pirate site is not — and not because it's "AI training," but because it's ordinary copyright infringement at the copying stage, one that existed long before any neural network did. Judges plainly aren't willing to declare piracy legal just because the end goal is training a model.
The second axis is what the model produces on the way out. If a model, as in the Getty or New York Times cases, can reproduce recognizable chunks of the original — a watermark, a near-verbatim paragraph — that's a separate claim, closer to traditional infringement through the copying of an output than through the process of training itself. Those cases aren't resolved yet, but they're the ones keeping open the question the industry finds most uncomfortable: even training that's transformative by design can still produce output that isn't transformative in fact.
The line doesn't run between "AI learns" and "AI steals." It runs between how a company acquired the material and what the model does with it afterward. Those are two different fights, and the industry is already starting to lose pieces of both — not all at once, but one ruling at a time.
These cases will decide something bigger than an abstract legal question: who ends up owning the value built from centuries of human creative labor — the people who created it, or the first company to figure out how to copy it at scale. As courts keep splitting each case into "how it was obtained" and "what it produced," creators are left with a real, working lever — not despair at the industry's size, but an actual precedent, already won, denominated in dollars per book.