ai copyright lawsuit: what the adjudicated record actually shows in 2026

Three courts, three different answers. A $1.5 billion Anthropic settlement, a rejected fair-use defense in Delaware, and a Getty loss in London show that the outcome turns on what was copied, not a single rule.

Every builder training a model and every buyer paying for one now asks the same question: is any of this legal? The headlines say courts are deciding it, and they are. They are just not deciding it the way a single headline can hold.

Because the answers disagree. In 2026 a Northern District of California court granted final approval to a $1.5 billion settlement in Bartz v. Anthropic, the largest copyright settlement in U.S. history (source). A year earlier a Delaware court had rejected the fair-use defense outright. A London court threw most of Getty’s case out.

Read together, those three rulings do not say AI training is legal, and they do not say it is theft. They say something more useful to anyone actually exposed: the outcome turns on what was copied and how it was obtained, not on the act of training itself.

What an ai copyright lawsuit is

An ai copyright lawsuit is a claim that an AI developer reproduced protected works without a licence, either to train a model or in what the model outputs. The disputes split along two axes the courts treat separately: whether the training use is fair use, and whether acquiring or storing the underlying copies was lawful.

Is training an AI on copyrighted work infringement?

It depends on what was copied and how, and the record already shows both answers. Training itself has been held fair use in one case and rejected as fair use in another, decided a year apart in two different federal courts.

In Thomson Reuters v. Ross Intelligence, a Delaware court found that a legal-research startup infringed by training on Westlaw editorial headnotes. Judge Stephanos Bibas held Ross had copied 2,243 headnotes and that “None of Ross’s possible defenses holds water. I reject them all” (source). The court weighed the fair-use factors and found the use commercial, non-transformative, and harmful to a market the judge named directly: “data to train legal AIs” (source).

That is a load-bearing detail, not a technicality. The court stressed that “Ross’s AI is not generative AI” (source). It answered the question for a non-generative search tool, and left the generative case open. The ruling is a partial summary judgment, and Ross has appealed to the Third Circuit.

What did the Anthropic copyright settlement decide?

On the merits, almost nothing, and that is the point. Anthropic denied the allegations and paid to avoid a damages trial, so no court ever ruled that its conduct was unlawful.

What the court did rule on came earlier, and it cut the case in two. In a June 2025 order Judge William Alsup held that training Claude on books was “transformative, spectacularly so” and therefore fair use (source). But he refused to extend that to how Anthropic got the books: the court sent to trial “the pirated copies used to create Anthropic’s central library and the resulting damages” (source).

That split is what put a price on the case. Training was safe; the piracy of the source copies was not, and it exposed Anthropic to statutory damages across hundreds of thousands of works. Rather than try it, Anthropic settled for a $1.5 billion fund at “approximately $3,000 per work” (source). The count of covered works is where the public sources diverge: “over 400,000” (JURIST), “approximately 500,000” (Authors Guild), and 482,460 through the class filters (Wolters Kluwer). TIN’s case file shows that conflict rather than averaging it away.

Have AI companies won any copyright lawsuits?

Yes, but narrowly, and the clearest win came in London on a ground that has nothing to do with fair use. On 4 November 2025 the England and Wales High Court handed down Getty Images v Stability AI and rejected Getty’s surviving copyright claim.

Getty had already abandoned its training and output claims during trial, largely because it could not show the relevant acts happened in the UK. On the one copyright claim left, Mrs Justice Smith held that Stable Diffusion is not an “infringing copy” because the model weights “do not store the visual information” from the works (source). Latham & Watkins read the holding the same way: “the model weights are not themselves an infringing copy and they do not store an infringing copy” (source).

The only infringement the court found was a trade-mark point it went out of its way to shrink, calling the case “historic” and its findings “of necessity, extremely limited” (source). Getty has been granted permission to appeal. This is a first-instance judgment, not the last word.

The three rulings side by side

Case	Court	What was copied	Holding	What it cost
Bartz v. Anthropic	N.D. Cal. (US)	~500,000 books, some from pirate libraries	Training fair use; acquiring pirated copies not	$1.5 billion settlement, no merits ruling
Thomson Reuters v. Ross	D. Del. (US)	2,243 Westlaw headnotes	Fair use rejected; non-generative tool infringed	Liability found; on appeal
Getty v. Stability	EWHC (UK)	Images alleged in training and outputs	Secondary copyright claim rejected; weights do not store works	Only “historic” trade-mark infringement; on appeal

What TIN’s case files document

Each row above is a TIN case file built on the primary court record, not a press summary. The Bartz settlement figures come from the court-authorized class notice and Judge Alsup’s order, with the conflicting work-counts shown, not merged. The Ross ruling is quoted from the signed Delaware opinion. The Getty holding is drawn from the EWHC judgment itself and its permission-to-appeal record.

That is the whole point of a case file. Anyone can summarize a headline. The value is a figure you can follow back to the document it came from.

The bottom line

The tempting read is that the courts are turning on AI, or that AI keeps winning on fair use. The record supports neither. Training on lawful copies has survived; training on pirated copies has cost $1.5 billion; training a non-generative tool on a rival’s editorial work lost on fair use; and a claim that a model “stores” the works it learned from failed on the facts of how the model works.

For an operator, the transferable lesson is not about fair use at all. It is about provenance. Every one of these cases turned on where the training data came from and what the model retained, not on whether a model was trained. The exposure lives in the supply chain of the data, and that is the part a buyer can actually audit before signing.

Sources

1. U.S. District Court, N.D. Cal., Bartz et al v. Anthropic PBC, No. 3:24-cv-05417-WHA, Order on Fair Use (Document 231), 2025-06-23. https://copyrightalliance.org/wp-content/uploads/2025/06/Bartz-v.-Anthropic-Order.pdf
2. JURIST, “Judge approves record $1.5 billion settlement involving Anthropic,” July 2026. https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/
3. U.S. District Court, D. Del., Thomson Reuters v. Ross Intelligence, No. 1:20-cv-613-SB, Memorandum Opinion (D.I. 770), 2025-02-11. https://www.govinfo.gov/content/pkg/USCOURTS-ded-1_20-cv-00613/pdf/USCOURTS-ded-1_20-cv-00613-5.pdf
4. England and Wales High Court (Chancery), Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), 2025-11-04. https://caselaw.nationalarchives.gov.uk/ewhc/ch/2025/2863
5. Latham & Watkins, “Getty Images v Stability AI: English High Court Rejects Secondary Copyright Claim,” November 2025. https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim
6. TIN case files: Bartz v. Anthropic, Thomson Reuters v. Ross Intelligence, Getty v. Stability AI (full source lists on each page).

Questions

Is training an AI on copyrighted work copyright infringement?

It depends on what was copied and how. A Delaware court rejected fair use for training a legal-research tool on Westlaw headnotes and found 2,243 infringed. A California judge held that training Claude on books was fair use, but that downloading those books from pirate libraries was not, and Anthropic settled that piracy for $1.5 billion.

What did the Anthropic copyright settlement decide?

The Anthropic copyright settlement decided nothing on the merits: Anthropic denied wrongdoing and paid to end the case. A Northern District of California court granted final approval in July 2026 to a $1.5 billion fund, about $3,000 per work across roughly half a million pirated books, the largest copyright settlement in U.S. history.

Have AI companies won any copyright lawsuits?

Yes, in part. A UK High Court rejected Getty's surviving secondary-copyright claim against Stability AI because the model's weights do not store the images, and a California judge held that the act of training on books was transformative and fair. The wins are narrow and case-specific, and both are under appeal or paired with a loss.

Sources

1. U.S. District Court, N.D. Cal. (Alsup, J.), Bartz et al v. Anthropic PBC, No. 3:24-cv-05417-WHA, Order on Fair Use (Document 231) , 2025-06-23
2. JURIST, Judge approves record $1.5 billion settlement involving Anthropic , 2026-07
3. U.S. District Court, D. Del. (Bibas, J.), Thomson Reuters v. Ross Intelligence, No. 1:20-cv-613-SB, Memorandum Opinion (D.I. 770) , 2025-02-11
4. England and Wales High Court (Chancery), Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch) , 2025-11-04
5. Latham & Watkins, Getty Images v Stability AI: English High Court Rejects Secondary Copyright Claim , 2025-11

---

This is analysis, not a verified outcome. It carries no verification badge and never will. The proof lives in the case files, where every figure is checked against the public record and the method is printed on the page.