AI training data lawsuits tracker

What did the court decide in Bartz v. Anthropic?

In Bartz v. Anthropic, Judge William Alsup ruled on June 23, 2025 that training Claude on books Anthropic had bought and scanned was fair use, but that downloading and keeping pirated copies was not. Anthropic then agreed to pay $1.5 billion, about $3,000 per book, and the court gave final approval on July 20, 2026.

The class covered books downloaded from Library Genesis in June 2021 or from the Pirate Library Mirror in July 2022. The Association of American Publishers reported a claims rate of 92.77% and called it the largest class action copyright settlement in history.

The settlement requires Anthropic to destroy the pirated files and copies made from them. It covers past conduct only, so it is not a license for future training.

See the case in the table

What did the court decide in Kadrey v. Meta?

In Kadrey v. Meta, Judge Vince Chhabria granted Meta partial summary judgment on June 25, 2025, finding that training Llama on 13 authors' books was fair use on the evidence those authors presented. He stressed that the authors lost because they had not shown harm to the market for their books.

Chhabria wrote that the ruling did not mean Meta's use of copyrighted materials to train its models was lawful in general. Claims about Meta distributing books through torrenting stayed in the case.

In May 2026 a group of publishers including Elsevier, Hachette and Macmillan filed a new class action against Meta in New York, so the training question is being tested again with different plaintiffs.

See the case in the table

What did the Third Circuit decide in Thomson Reuters v. Ross Intelligence?

In Thomson Reuters v. Ross Intelligence, the Third Circuit ruled on September 29, 2026 that Ross's copying of Westlaw headnotes to train a competing legal search tool was not fair use. It is the first U.S. appeals court ruling on fair use in AI training.

The court found the headnotes had enough originality to be protected, called Ross's use minimally transformative at best, and weighed heavily that Ross was building a competitor to Westlaw.

The court also recognized an emerging market for licensing content as AI training data and said unlicensed training can take that market from the owner. Ross's tool was not generative AI, which limits how far the ruling reaches.

See the case in the table

What did the Ninth Circuit decide in Doe 1 v. GitHub?

In Doe 1 v. GitHub, the Ninth Circuit ruled on September 16, 2026 that code generated by Copilot without the original copyright notices does not violate the DMCA rule against removing copyright management information. Removing a notice from an existing work is different from creating a new work that never had one.

The ruling closes off a theory that plaintiffs had pleaded in many AI cases, and courts elsewhere, including in Concord v. Anthropic, asked for briefing on it.

The programmers' claims that GitHub and Microsoft breached open-source license terms remain in the district court.

See the case in the table

Where does New York Times v. Microsoft and OpenAI stand?

New York Times v. Microsoft and OpenAI is pending before Judge Sidney Stein in New York, consolidated since April 2025 with the Authors Guild class action and other news cases. Both sides filed summary judgment motions on September 4, 2026, and no trial date is set.

The Times argues ChatGPT and Copilot copy its work at scale and produce substitutes for it. OpenAI and Microsoft argue training is fair use and point to Bartz and Kadrey. The Justice Department filed a brief supporting OpenAI.

A ruling on fair use here would be the first on news training in the Second Circuit, where both the Times and the authors argue that binding precedent favors them.

See the case in the table

What did the High Court decide in Getty Images v Stability AI?

In Getty Images v Stability AI, Mrs Justice Joanna Smith ruled on November 4, 2025 that Stable Diffusion is not an infringing copy under UK law because its model weights do not store Getty's images. Getty had already dropped its main training claims because training took place outside the UK.

The judge found limited, historic trademark infringement where early model versions produced images with Getty watermarks. The rest of the trademark case failed.

In December 2025 Getty was given permission to appeal the secondary infringement point. Getty's separate US case continues in California.

See the case in the table

What did German courts decide in GEMA v. OpenAI and GEMA v. Suno?

In GEMA v. OpenAI, the Munich Regional Court ruled on November 11, 2025 that GPT-4 and GPT-4o memorized nine song lyrics and that this memorization is a reproduction outside the EU text and data mining exceptions. The same chamber ruled against Suno on July 31, 2026 on similar reasoning.

In the Suno case the court held that copies memorized in models hosted in Germany fall under German law, even though training happened in the US, and that US fair use does not apply. Neither judgment was final at last report.

The Hamburg appeals court reached a different result in Kneschke v. LAION in December 2025, where building a research dataset fell within the text and data mining exceptions. The difference turned on research use and on whether the model could reproduce the works.

See the case in the table

What did the Delhi High Court decide in ANI v. OpenAI?

In ANI v. OpenAI, Justice Amit Bansal of the Delhi High Court refused an interim injunction on July 24, 2026. He found that storing ANI's articles to train ChatGPT was prima facie not infringement under the Indian Copyright Act and that ANI had not shown memorization or regurgitation.

The court did rule for ANI on jurisdiction, holding that the suit can be heard in Delhi although OpenAI's servers are in the United States. The findings are provisional and the case continues to trial.

See the case in the table

What do these lawsuits mean if you license data?

For U.S. companies, the fair use rulings so far turn on how data was obtained and whether outputs compete with the original, not on whether training happened. Judges in Bartz v. Anthropic and Kadrey v. Meta separated lawfully acquired copies from pirated ones, and the Third Circuit in Thomson Reuters v. Ross treated licensing for AI training as a real market.

The money has followed the acquisition question. Anthropic paid $1.5 billion in Bartz for books taken from pirate libraries, not for training on books it had bought. For a U.S. AI company, a written license with a documented chain of title answers the questions that drove that settlement: where each record came from, who owns it, and what use was agreed. It also gives the buyer an indemnity to negotiate, which a scraped dataset cannot.

For a U.S. data owner, the cases show what buyers check before they pay. They ask whether you own the records, whether client contracts or website terms restrict the use, whether personal data was collected under a privacy notice that allows it under the CCPA and other state laws, and whether any third-party content is mixed in. Our guide to whether you can legally sell company data for AI covers each point, and data provenance explained covers the records buyers want to see.

Licensing does not remove every risk. Illinois biometric claims like those Clearview settled, state and federal wiretap claims like those against Otter.ai, contract claims like Reddit's against Anthropic, and output claims like the lyrics cases can arise even where a license exists. The deals that followed U.S. litigation, such as Universal's and Warner's settlements with Udio and Suno, are listed in our AI data deals tracker, and published prices are in the AI data price index.

If you hold UK or EU data

There is no fair use outside the United States. German courts in GEMA v. OpenAI and GEMA v. Suno held that a model that memorizes works falls outside the EU text and data mining exception, while the Hamburg appeals court in Kneschke v. LAION let research dataset building proceed and required opt-outs to be machine-readable. Our guide to the EU AI Act and training data transparency covers the disclosure rules that apply to models sold in the EU.

We are a broker, not a law firm. We tell you what buyers will check and help you prepare, but you should take independent legal advice before signing any license.

How is this tracker compiled?

We include lawsuits and court or regulator rulings where the use of data to train or run AI models is a central issue. We favor cases that produced a ruling, a settlement or a large class, and cases that test a new data type or jurisdiction.

• Every fact comes from a court record, a party's announcement or a reputable legal or news publication, linked on each row
• Status reflects the latest report we could find. Where we could not confirm a later status, the row says "Status as last reported" with the month
• "Ruled" means a court has decided the core training question or a dispositive motion, even if other claims continue. "Pending" means no such ruling yet, including cases with only procedural or interim rulings
• Settlement values appear only where a court or party disclosed them. Undisclosed terms are described as undisclosed
• Case names follow the form used by the court or the leading published report

We review every case monthly and when a ruling is reported. This page describes what courts decided and is not legal advice.

Read more about how we research, and see our AI data licensing agreement checklist for the terms that allocate these risks in a deal.

Frequently asked questions