How is AI training data licensed, and what have courts decided?

What is a training data license?

It is a contract in which a content owner lets an AI developer use its material. The use can be training a model, showing excerpts in answers, or both.

The deals exist because the law is unsettled. If training on copyrighted work is always fair use, a developer has little reason to pay. If it is never fair use, every developer must pay. Courts have so far landed between those poles, so a data license is partly a payment for content and partly a payment for certainty.

This guide covers deals and court cases as of October 2026. For licenses on model weights, see open-weight model licenses.

Which content licensing deals are public?

The deals with the clearest public record are listed below. Most terms are confidential, so the dollar figures come from securities filings and press reports.

Deal	Announced	What it covers	Money (as reported)
Associated Press – OpenAI	July 13, 2023	Part of AP’s text archive dating to 1985; AP access to OpenAI technology	Not disclosed
Axel Springer – OpenAI	December 13, 2023	Content in ChatGPT answers and for training	“Tens of millions of euros” a year, per the Financial Times; not confirmed by the companies
Reddit – Google	February 22, 2024	Access to Reddit’s Data API for training and display	About $60 million a year, per Reuters; not confirmed by the companies
Reddit (all data deals)	Disclosed in S-1, 2024	Data licensing contracts signed in January 2024	$203.0 million aggregate, terms of two to three years (SEC filing)
Reddit – OpenAI	May 16, 2024	Reddit content via Data API in ChatGPT	Not disclosed
News Corp – OpenAI	May 22, 2024	Current and archived content from the Wall Street Journal, the Times of London and other titles	Possibly more than $250 million over five years, per the Wall Street Journal
New York Times – Amazon	May 2025	News, Cooking and The Athletic content for training and product summaries	$20 million to $25 million a year, per the Wall Street Journal
Disney – OpenAI	December 11, 2025	More than 200 characters for Sora video; three-year license	$1 billion equity investment by Disney; license fee not disclosed

Music moved from lawsuits to licenses. Universal Music Group settled its suit against Udio on October 29, 2025, and agreed to build a licensed platform with it. Warner Music Group settled with Suno on November 25, 2025, and signed a licensing partnership.

Two patterns stand out. Several deals cover live access, through an API or current articles, not only old archives. And at least one publisher, the New York Times, is suing one AI company while licensing to another, so suing and licensing are not mutually exclusive.

What did the Bartz v. Anthropic case decide?

Bartz v. Anthropic held that training on lawfully acquired books was fair use but that keeping pirated copies in a central library was not. The case then settled for $1.5 billion.

Authors sued Anthropic in the Northern District of California over books used to train its Claude models.

On June 23, 2025, Judge William Alsup split the issue. He held that training on books was “exceedingly transformative” and fair use. Buying print books and scanning them for training was also fair use. But downloading pirated copies from shadow libraries and keeping them in a permanent central library was not.

Under the settlement, Anthropic agreed to pay $1.5 billion, about $3,000 per work, for a class of roughly 480,000 works. Judge Araceli Martínez-Olguín granted final approval on July 20, 2026, according to the Authors Guild. The release covers past conduct only. It does not release claims about AI outputs or future conduct.

For licensing, the lesson is narrow: in this ruling, how the data was acquired mattered more than what the model did with it.

What did Kadrey v. Meta decide?

Kadrey v. Meta ended in a fair use ruling for Meta, but the judge wrote that the authors had made the wrong arguments.

On June 25, 2025, two days after the Bartz ruling, Judge Vince Chhabria granted summary judgment to Meta against 13 authors who claimed Meta trained Llama on their books. He found the use fair on that record. He suggested a different theory, “market dilution,” under which a model that floods the market with competing works could harm the originals even without copying them. The plaintiffs had not built the evidence for it.

The ruling binds only these parties, and it left market dilution, the theory the judge himself raised, untested.

What did the appeals court decide in Thomson Reuters v. Ross?

The Third Circuit held on September 29, 2026 that Ross Intelligence’s use of Westlaw headnotes to train a non-generative legal search tool was not fair use, affirming the district court.

Ross built an AI legal search tool. To train it, Ross used material derived from Westlaw headnotes, which are short summaries of court rulings owned by Thomson Reuters.

In February 2025, Judge Stephanos Bibas granted summary judgment for Thomson Reuters. He found the headnotes original enough to be protected and Ross’s use not fair use.

The appeals court affirmed on both issues, according to LawSites, which describes it as the first federal appellate decision on fair use in AI training. The opinion, by Judge Tamika R. Montgomery-Reeves, was initially sealed for redactions and has since been released; the court stressed that Ross’s tool “was not a generative AI.” Ross has said it will appeal to the Supreme Court.

The facts limit its reach. The copying took place before generative AI tools came into wide use, according to LawSites. The Ross tool was not generative, and it competed directly with Westlaw.

Where does New York Times v. OpenAI stand?

As of October 2026, the Times’s suit against OpenAI and Microsoft was still in the trial court, with no ruling on fair use and no trial date that we could find.

The New York Times sued OpenAI and Microsoft in December 2023 in the Southern District of New York. It claims they copied millions of articles to train ChatGPT and related products.

On March 26, 2025, Judge Sidney Stein let the core claims go forward, the Associated Press reported. In late June 2026, the Times filed an amended complaint that dropped one claim against OpenAI and modified one against Microsoft.

The Times has meanwhile licensed its content to Amazon.

How have courts outside the U.S. ruled?

Courts outside the U.S. have reached different results on different facts. In the UK, Getty did not claim that Stable Diffusion’s weights contained copies of its images, and the court held the model was not an “infringing copy.” In Germany, a court found that song lyrics memorized inside OpenAI’s models were a reproduction.

United Kingdom: Getty Images v. Stability AI

Getty sued Stability AI over the Stable Diffusion image model. Before closing submissions, Getty dropped its main training claims, acknowledging there was no evidence that the training took place in the UK.

On November 4, 2025, the High Court rejected Getty’s remaining secondary copyright claim. Getty did not assert that the model weights contained copies of its images, and the court said the weights did not store them, so the model was not an “infringing copy.” Getty won a narrow trademark point over watermarks in some outputs.

In December 2025, the judge granted Getty permission to appeal the secondary infringement point, according to Taylor Wessing. The appeal concerns the meaning of “infringing copy,” not whether the model stores copies. We found no Court of Appeal decision as of October 2026.

Germany: GEMA v. OpenAI

On November 11, 2025, the Munich Regional Court ruled for GEMA, the German music rights society, over song lyrics. It found that lyrics were memorized in GPT-4 and GPT-4o and that memorization inside the model was a reproduction. It held that the EU text-and-data-mining exception did not cover that memorization. The ruling is first-instance and can be appealed.

That question decides whether the model itself, not just its training, needs a license.

What do regulators say about licensing training data?

Regulators have so far preferred markets to mandates. The U.S. Copyright Office favors voluntary licensing over compulsory licensing, and the EU requires developers to honor opt-outs without requiring them to buy licenses.

The U.S. Copyright Office released a pre-publication version of Part 3 of its AI report on May 9, 2025. It concluded that voluntary licensing markets should develop without government intervention and favored them over compulsory licensing. It suggested extended collective licensing could be considered where individual deals are unworkable.

The EU took a different route. Under the AI Act, every general-purpose model provider must have a policy to respect copyright, including rights holders’ text-and-data-mining opt-outs, and must publish a summary of training content using the Commission’s template. See frontier AI laws.

Because the EU duty is to honor opt-outs rather than to pay, rights holders who want payment have to negotiate for it.

What should a developer or publisher take from all this?

The cases so far treat how data was obtained and how it was used as separate legal questions, and only trial courts, on different facts, have addressed whether a trained model contains copies. The rest follows from that:

• Pirated sources are the clearest liability in U.S. rulings so far
• A non-generative tool that competes with its source has lost on fair use at the appeals level
• No appeals court has decided whether model weights contain copies; the UK and German trial rulings turned on different facts
• Deals increasingly cover display in answers, not only training
• EU opt-outs must be respected regardless of U.S. fair use

Is AI training fair use?

Sometimes, depending on the facts. No court has held that all AI training is fair use, and none has held that all of it is infringement. The licensing deals fill that gap, priced on how each side reads the legal risk.

Case statuses are stated as of October 2026, and several appeals are pending. This guide is general information, not legal advice.