Nobody can give you a clean yes or no yet. Two US trial judges have held that training a generative model on books can be fair use, an appeals court has held that training a different kind of AI tool on a competitor’s content was not, and the biggest case, The New York Times against OpenAI and Microsoft, is awaiting a ruling on fair use. Whether training AI on copyrighted work is legal in the United States depends on how the model was built, where the training data came from, and what the finished product does to the market for the originals.

This post lays out the decisions in order, separates the settled points from the contested ones, and explains why a $1.5 billion settlement did not answer the underlying question. It covers US federal law, with a short note on the United Kingdom.

The scoreboard as of October 2026

Case Court Result so far
Bartz v. Anthropic N.D. Cal. Training on books held fair use (June 2025); building a library from pirated copies held not fair use; $1.5 billion class settlement finally approved July 20, 2026
Kadrey v. Meta N.D. Cal. Meta won on fair use for this record (June 25, 2025), but the judge said better evidence of market harm could change the outcome
Thomson Reuters v. ROSS Intelligence D. Del., then 3d Cir. Not fair use; affirmed by the Third Circuit on September 29, 2026
New York Times v. OpenAI and Microsoft S.D.N.Y. Summary judgment motions filed in early September 2026; no fair use ruling yet
Getty Images v. Stability AI UK High Court Getty’s secondary infringement claim rejected; permission to appeal granted

First, what “legal” is being asked

Copyright gives the owner exclusive rights, including the right to copy. Building an AI model involves a lot of copying: collecting works, storing them, cleaning them, and feeding them through training. Each step can look like infringement on its face. The question then becomes whether a defence applies, and in the US that defence is usually fair use, a four-factor test weighing the purpose of the use, the nature of the work, how much was taken, and the effect on the market for the original.

The Copyright Office reached that same structure in its report on training, issued in pre-publication form in May 2025. It said building and deploying a generative system with copyrighted material may be prima facie infringement at several stages, from data collection to training to retrieval, and that fair use must be decided case by case. It rejected the idea that training is automatically transformative and also rejected the comparison to a human learning from what they read. Its answer was a spectrum: noncommercial research with no public output is more likely fair, while commercial uses that produce competing content are less likely to be. The report’s formal status is unclear, because the Register of Copyrights and the Librarian of Congress were removed around the time it appeared.

Two trial courts lean toward fair use for training

In June 2025, Judge William Alsup ruled in Bartz v. Anthropic that training a language model on books was fair use, and that scanning print books the company had bought was too. He drew a firm line elsewhere. Downloading and keeping pirated copies from shadow libraries to build a permanent central library was not fair use, whatever the later purpose.

Two days later, Judge Vince Chhabria ruled for Meta in Kadrey v. Meta, but on narrower ground. Factors one and three favoured Meta, factor two favoured the authors, and the fourth factor decided it: the authors offered no real evidence that Meta’s model harmed the market for their books. The judge floated a theory he called market dilution, in which a flood of AI-generated works in a genre competes with the originals, and he said that a plaintiff with a better record could win on it. He also spent more time on Meta’s use of pirated sources than Alsup had. Commentators noted the two judges did not follow the same path.

The lesson is not that training is safe. It is that fair use was decided on records, and the authors in Kadrey lost partly for what they did not prove.

An appeals court says no, for a different kind of AI

On September 29, 2026, the Third Circuit affirmed the ruling against ROSS Intelligence in the case brought by Thomson Reuters. ROSS had used Westlaw’s editorial headnotes to build a legal-research tool that competed with Westlaw. According to law firm summaries of the opinion, the court found the headnotes original enough to protect, held that the use was commercial and only minimally transformative, noted that entire headnotes were copied, and found market harm in both the existing legal-research market and an emerging market for licensing headnotes as AI training data.

That last point deserves attention. The court treated licensing for AI training as a real, developing market, which makes unlicensed copying harder to justify. It is the first federal appellate decision on fair use in AI training.

It is not a ruling on ChatGPT or Claude. ROSS was not a generative system, and the court distinguished it from Bartz and the OpenAI litigation by saying ROSS’s platform could not generate original expression. The opinion did not settle the disagreement between Alsup and Chhabria over market harm from generative AI, and no federal appeals court has yet addressed that. Reports on the ruling differ on one procedural detail: IPWatchdog described it as issued under seal on the day, while Plagiarism Today reported that the opinion was unsealed on September 30.

What the $1.5 billion settlement did and did not decide

The Bartz piracy claim was headed to trial when it settled. On July 20, 2026, Judge Araceli Martinez-Olguin granted final approval of a $1.5 billion class settlement covering more than 400,000 books, at roughly $3,000 per work. The Authors Guild describes the release narrowly: class members gave up claims about Anthropic’s past acquisition and copying of their works through August 25, 2025. Claims over model outputs and future conduct were not released.

That matters for how you read the number. The settlement priced the risk of the piracy claim, which Alsup had already held was not fair use. It said nothing new about whether training on lawfully obtained books is allowed, and it binds no other court. Other AI companies can still be sued over how they acquired data, and over what their models output.

The case that may come next: The New York Times v. OpenAI

The Times sued OpenAI and Microsoft in 2023. In early September 2026, all three sides filed motions for summary judgment before Judge Sidney Stein in Manhattan. The Times argues the defendants copied its works at scale to build commercial substitutes and says binding Supreme Court and Second Circuit precedent, not recent California decisions, should govern. OpenAI argues existing law allows training on publicly available works and leans on Kadrey and Bartz. Microsoft makes similar points about its Copilot product.

The Department of Justice also filed a statement of interest supporting OpenAI’s fair-use position, describing training as transformative. That is notable because it is the first time, per the reports I found, that the federal government has publicly taken a side in these suits. A statement of interest does not bind the judge. Axios reported that Stein is expected to rule in the coming months, with possible trial next year, but no date has been set. A decision here would come from a Second Circuit court, a different forum than the California rulings, so it could move the balance in either direction.

Outside the US: a note on the UK

In the United Kingdom, the High Court rejected Getty Images’ secondary infringement claim against Stability AI. The judge reasoned that model weights, though they are articles, are not infringing copies because they never stored or contained a copy of a copyrighted work. Getty’s trademark claim over watermarks that appeared in some generated images partly succeeded, and Getty has permission to appeal the secondary infringement point. UK courts did not decide infringement during training itself. US readers should not assume that logic transfers, since US fair use has no British equivalent.

What is settled and what is not

  • Reasonably settled: Acquiring copies from pirate sites carries real liability risk; the Bartz settlement is the proof. Copying a competitor’s expressive content to build a rival product is hard to defend as fair use, at least for non-generative tools, after the Third Circuit.
  • Contested: Whether training a generative model on lawfully obtained works is fair use. Two district judges said yes on their records, with caveats. No appeals court has ruled.
  • Open: How much market dilution counts as harm; whether outputs that closely track source works infringe; and whether licensing markets for training data can defeat fair use.

Related questions

Can an author stop an AI company from training on their book? Not through any clear right today. Authors can sue, and success so far has turned on how the data was obtained and on proof of market harm.

Is it legal for me to use a model that was trained on copyrighted work? The training lawsuits target the companies that built the models. Users can face separate risk if specific outputs copy protected material, which is why some vendors offer indemnities.

Will Congress settle this? No bill I verified resolves it, and the Copyright Office has urged letting licensing markets develop without government intervention for now.

This is general information, not legal advice, and it reflects reports available on October 10, 2026.

Sources