Training Data and Copyright: The Cases So Far

A large language model learns from trillions of words. Most of those words were written by someone, and many of them are under copyright. Whether copying them to train a model is legal in the United States turns on one sentence of federal law and a stack of lawsuits still working their way up. As of October 2026, no appeals court has ruled on a generative AI training case, and the trial courts that have ruled split on important details.

Where training data comes from

• Web crawls. Common Crawl, a nonprofit, has archived billions of web pages since 2008. Most large models start with filtered versions of it. Companies also run their own crawlers.
• Books. Some licensed, some scanned from purchased copies, and some downloaded from shadow libraries like Library Genesis. That last source is now the most legally expensive one.
• Code. Public repositories, under a mix of open source licenses.
• Licensed deals. News publishers, stock photo libraries and forums like Reddit have signed paid licensing deals with AI companies.
• User data. Chats and uploads from consumer products, depending on settings. Lesson 5 covers this.
• Synthetic data. Text generated by other models, increasingly used for later training stages.

Companies rarely publish a full list. That opacity is itself a policy fight; the EU AI Act, covered in lesson 8, requires general purpose model providers to publish a summary of training content.

The fair use test

US copyright law, 17 U.S.C. 107, lets courts excuse some unlicensed copying as fair use. Judges weigh four factors:

1. The purpose and character of the use, including whether it is commercial and whether it is transformative.
2. The nature of the copyrighted work (creative fiction gets more protection than facts).
3. The amount and substantiality of the portion used.
4. The effect of the use on the potential market for or value of the work.

AI companies argue training is transformative: the model learns statistical patterns and does not store or sell the books. Rights holders argue the copying is wholesale, commercial and builds a product that competes with the originals and with the licensing market for training data. Both arguments have won pieces of rulings.

The cases, as of October 2026

Case	Court	What it is about	Status, October 2026
New York Times v. Microsoft and OpenAI	S.D.N.Y., Judge Sidney Stein, in MDL 25-md-03143	Training on millions of Times articles; outputs that reproduce them	Core infringement claims survived dismissal (March 2025). Cross-motions for summary judgment filed Sept 4, 2026, replies due Nov 6. No trial date.
Authors Guild v. OpenAI	S.D.N.Y., same MDL	Fiction and nonfiction authors' books in training	Active, consolidated with the Times case
Bartz v. Anthropic	N.D. Cal.	Books, including pirated copies, used to train Claude	$1.5 billion class settlement, final approval July 20, 2026; about $3,000 per work; payments expected no earlier than late 2026
Kadrey v. Meta	N.D. Cal., Judge Vince Chhabria	Books, including shadow library copies, used to train Llama	June 2025: training held fair use on that record; a separate claim about sharing files through torrenting was not decided in that ruling
Thomson Reuters v. Ross Intelligence	Third Circuit	Westlaw headnotes used to train a non-generative legal search tool	Sept 29, 2026: appeals court affirmed, not fair use. Ross says it will seek Supreme Court review.
Getty Images v. Stability AI (UK)	High Court, then Court of Appeal	Photos used to train Stable Diffusion	Nov 4, 2025: Getty lost secondary copyright claim, won narrow trademark points. Appeal on the copy question pending.
Getty Images v. Stability AI (US)	N.D. Cal., Judge Trina Thompson	More than 12 million photos used for training	Now in federal court in California. April 2026: most claims survived a motion to dismiss.

Bartz v. Anthropic, in more detail

In June 2025 Judge William Alsup ruled that training on books was "exceedingly transformative" and fair use. He also ruled that downloading millions of pirated books to build a permanent central library was a separate use and not protected. Facing a trial on statutory damages for those pirated works, Anthropic settled for $1.5 billion, which the plaintiffs' lawyers called the largest publicly reported copyright recovery. Judge Araceli Martinez-Olguin, who now presides over the case, granted final approval on July 20, 2026, overruled the objections and cut class counsel's fee to about $101.5 million. The Authors Guild describes the payout as about $3,000 per work, four times the $750 statutory minimum. Anthropic must destroy the pirated files. Class counsel said payments would not begin before the end of 2026, in two rounds.

Settlements create no legal precedent. This one still sent a price signal and turned "where did the data come from?" into the most dangerous question an AI company can be asked.

Kadrey v. Meta

Two days after Alsup, Judge Vince Chhabria also found training on books was fair use, but for a narrower reason: the authors had not offered evidence of market harm. In a passage that did not decide the case, he wrote that if plaintiffs could show AI-generated books diluting the market for their genres, training would in most cases not be fair use. Unlike Alsup, he treated the shadow library downloads as part of the training purpose. Two judges in the same courthouse, same month, different routes. Watch for a plaintiff who actually builds a market dilution record.

New York Times v. OpenAI

Filed December 27, 2023. In March 2025 Judge Stein let the direct and contributory infringement claims go forward and dismissed some others. Discovery fights have been their own story: OpenAI was under a preservation order to keep consumer ChatGPT logs, including deleted chats, until September 26, 2025, and in late 2025 the court ordered production of a de-identified sample of 20 million logs. In September 2026 the Justice Department filed a statement of interest supporting OpenAI's position. A ruling on fair use could come in the months after replies are filed in November 2026.

What the Copyright Office has said

The US Copyright Office issued its report on copyright and AI in three parts:

• Part 1, July 31, 2024: digital replicas (deepfakes of real people). It recommended a new federal law protecting people against unauthorized digital replicas.
• Part 2, January 29, 2025: copyrightability. Protection requires human-determined expressive elements. Prompts alone are not enough. Using AI as a tool does not bar protection for the human parts. No change in law needed.
• Part 3, pre-publication version May 2025: training. It concluded that some training uses are likely fair and others are not, that uses producing outputs which compete with the originals weigh heavily against fair use, that market dilution is a legitimate concern under factor four, and that voluntary licensing should be allowed to develop without government intervention for now. The Office said the final version would not change its analysis.

Copyright Office reports do not bind courts. Judges can and do cite them.