Is AI training free? — The United States, fair use, and Anthropic

The previous article examined German court cases. What, then, is happening in the United States, home to the major AI companies?

To begin with the conclusion, American courts differ from German courts: they tend to view AI training as closer to fair use. It is notable, however, that many disputes over AI training are being resolved through enormous settlements rather than judgments.

This article examines three major US cases concerning AI training.

⚖️ The scales of fair use

First, let us review fair use under US copyright law. This exception permits use of another person’s work without permission. Courts assess four factors:

1. The purpose and character of the use, particularly how transformative it is.
2. The nature of the copyrighted work.
3. The amount used.
4. The effect on the market for, or value of, the original work.

In AI-training cases, the real disputes concern the first and fourth factors. Does AI reading books and learning language patterns substitute for the originals, or turn them into something entirely different — transformativeness? And does the result undermine the originals’ market — market effect?

This concerns the character and consequences of use rather than whether the work remains present. Its starting point therefore differs fundamentally from the German court’s question in the previous article: does the copyrighted work remain inside the AI model?

📕 First — Thomson Reuters v. Ross Intelligence[1]: the first rejection of fair use

The case began with a refused license. Legal AI startup Ross Intelligence wanted to use data from Thomson Reuters’ legal research service Westlaw to build a case-law search service. Thomson Reuters refused, fearing a future competitor.

Ross Intelligence then took an indirect route. It commissioned an outside provider to produce a collection of legal memoranda, but those memoranda were effectively based on Westlaw’s case summaries, or headnotes.

First, the court held that the headnotes were creative works incorporating editorial judgment, rather than mere extracts from judgments, and rejected fair use for using them in training.

The central issue was the fourth factor, market effect. The court emphasized that Ross Intelligence was building a direct substitute for Westlaw, and that it also harmed the potential market in which Thomson Reuters could license data for AI training. The latter point is particularly significant because it recognizes the existence of a training-data licensing market.

There is, however, a major qualification. Ross Intelligence’s AI was a search system, not generative AI, and the court explicitly stated that its reasoning did not automatically extend to generative AI.

[1] Thomson Reuters v. Ross Intelligence — US District Court for the District of Delaware, judgment of February 2025. Ross sought an interlocutory appeal, which the Third Circuit accepted in June 2025. Oral argument was held on June 11, 2026 (case 25-2153). No judgment has yet been issued.

For context, Ross Intelligence has already ceased business, unable to sustain its litigation costs.

📗 Second — Bartz v. Anthropic[2]: training is permitted, piracy is not

This lawsuit was filed against Anthropic in August 2024 by thriller writer Andrea Bartz and two other authors. They alleged that Anthropic downloaded copyrighted books in bulk from shadow libraries such as LibGen to train Claude, and permanently retained those books in an internal library even after deciding not to use them for training.

Judge Alsup divided Anthropic’s conduct into three categories. That separation is the key to the ruling.

1. Training language models on lawfully purchased books — fair use recognized as highly transformative.
2. Scanning purchased physical books into digital form — fair use recognized because the originals were discarded and only the format changed.
3. Downloading more than seven million books from shadow libraries and storing them in an internal library — fair use rejected.

In other words, even if the purpose of training is legitimate, the method of obtaining the data is assessed separately.

[2] Bartz v. Anthropic — US District Court for the Northern District of California, judgment of June 2025 (Judge William Alsup). Training LLMs on lawfully acquired books and digitizing purchased books were held to be fair use; downloading and retaining copies from piracy sites were not.

The third category, for which fair use was rejected, was ultimately set for a jury trial. US copyright law allows statutory damages of up to $150,000 per work. With seven million books involved, the arithmetic could threaten the company’s existence. Anthropic eventually settled for $1.5 billion, and the settlement received final approval on July 20[3]. The court described it as the largest copyright class-action settlement in history.

[3] Bartz settlement — finally approved on July 20, 2026. The total is $1.5 billion, covering approximately 482,000 works.

The authors also argued that the trained Claude could write books resembling theirs, ultimately undermining their works’ market. Judge Alsup rejected this market-harm concern. In explaining transformativeness, he likened their argument to complaining that teaching children to write well would generate competing works. Copyright protects specific expression, not a right to prevent new competing works from being created.

📘 Third — Kadrey v. Meta[4]: a victory that is not quite a victory

This case was brought by thirteen authors, including Sarah Silverman, against Meta. Meta won on the result, but the judgment deserves close reading.

Judge Chhabria strongly endorsed a theory of market dilution. The idea is that an unlimited flow of similar works from generative AI could undermine demand for originals and incentives to create. He stated that proving this could allow plaintiffs to prevail on the fourth fair-use factor — market effect — and even on fair use overall. He also directly rejected Judge Alsup’s analogy to teaching children to write as inappropriate.

Yet the thirteen plaintiffs submitted no evidence of market dilution. Meta supplied evidence that its training had not affected sales of their books, while the plaintiffs neither offered contrary evidence nor properly contested the issue. Judge Chhabria therefore had to rule for Meta. He nevertheless made clear that the judgment did not mean Meta’s use was lawful; it meant only that these plaintiffs had advanced the wrong arguments and failed to build a record supporting the right ones.

[4] Kadrey v. Meta — US District Court for the Northern District of California, judgment of June 2025 (Judge Vince Chhabria). The court recognized fair use for training Llama but expressly stated that this followed from the plaintiffs’ failure to submit evidence of market dilution. Claims relating to distribution through torrent seeding during downloading remain pending separately.

Compared with the second case, two rulings issued by the same court just two days apart adopted opposite reasoning. Even within the United States, the answer has not converged.

🧭 Closing thoughts

The American position, based on these three cases, can be summarized as follows:

1. How was the training data obtained? Lawfully purchased books are acceptable; downloads from piracy sites are problematic.
2. Was the market actually harmed? Market dilution has entered the courtroom as a theory but has not yet been proved. The situation could change the moment a plaintiff demonstrates it with data.

One case is drawing attention as a potential first demonstration of market dilution. In The New York Times’ lawsuit against OpenAI[5], the court ordered OpenAI in January to produce all twenty million anonymized ChatGPT conversation logs. If this reveals indications of market harm — for example, routine generation of content substituting for paid articles — it could provide the first proof of market injury that the authors in the Meta case could not establish.

[5] The New York Times Co. v. Microsoft Corp. & OpenAI — US District Court for the Southern District of New York, case 1:23-cv-11195 (Judge Sidney Stein). Filed in December 2023. In April 2025, the court largely denied OpenAI’s motion to dismiss, rejected the assertion that AI training is inherently transformative, and held that fair use required factual examination. The case has since been consolidated with similar suits by other publishers and authors in multidistrict litigation, MDL No. 3143, before the same court. As of August 2026 it is in discovery, with no trial date set.

The answers from the two continents examined across these articles diverge. Germany held that training involving copying beyond what is necessary requires a license. The United States viewed training as transformative and therefore fair use, while leaving qualifications concerning acquisition methods and market effects.

Which direction will South Korea take? The outcome of the same case could differ substantially depending on which question it prioritizes: in what form does a work remain inside a model, or is the use transformative and what effect does it have on the market? Discussion of introducing a TDM exception is currently under way, and that process should clarify the direction.

#BetweenTechAndRules #AICopyright #FairUse #FairUse #Anthropic #TDM #GenerativeAI