AI Training on Copyrighted Works: Is It Fair Use? Three 2025 Decisions Reshaping the Debate

For the better part of three years, the central unanswered question in the wave of AI copyright litigation has been a single sentence: Is training a machine learning model on copyrighted works — without a license — fair use? The question matters enormously. Generative AI models are only as good as the data they consume, and the most valuable training data is, almost by definition, someone else's copyrighted expression. Dozens of cases teed up the issue, but none reached the finish line — until 2025.

In the span of ten days in June 2025, three federal district court decisions landed that, read together, sketch the emerging architecture of the answer. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., handed down in February 2025 by Judge Bibas in the District of Delaware, became the first decision to squarely reject the fair use defense for AI training. Then, in late June, Judge Alsup in the Northern District of California issued a split ruling in Bartz v. Anthropic — a partial win for the AI industry on lawful-acquisition training, and a decisive loss on the use of pirated books. Two days later, Judge Chhabria, also in the Northern District of California, reluctantly granted summary judgment to Meta in Kadrey v. Meta Platforms, Inc., while warning that a better record of market harm could flip the result in the next case.

None of these decisions is final, and each turns on its own facts. But the analytical threads they share — particularly around transformativeness and the fourth fair use factor, market harm — already frame how every remaining AI training case will be litigated.

The Fair Use Framework, Briefly

Section 107 of the Copyright Act codifies fair use as an affirmative defense to copyright infringement. The defense is a balancing test, with four non-exclusive factors:

1. The purpose and character of the use, including whether it is commercial and whether it is transformative — that is, whether it adds "a further purpose or different character" rather than merely superseding the original.
2. The nature of the copyrighted work — whether it is creative or factual, published or unpublished.
3. The amount and substantiality of the portion used in relation to the work as a whole.
4. The effect of the use upon the potential market for or value of the copyrighted work.

The Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith reshaped factor one, narrowing what counts as "transformative" and emphasizing that commercial uses borrowing an original's primary purpose receive little First Amendment solicitude. And as the 2025 AI decisions confirm, factor four — market harm, including harm to potential licensing markets — has reclaimed its traditional place as "the single most important element of fair use," in the words of Harper & Row v. Nation Enterprises.

Thomson Reuters v. Ross Intelligence: The First Clear "No"

Thomson Reuters v. Ross Intelligence, No. 1:20-cv-613-SB (D. Del. Feb. 11, 2025), is the first reported decision to hold that training an AI system on copyrighted works is not fair use. Ross Intelligence built an AI-powered legal research platform designed to compete with Westlaw, the dominant legal research service owned by Thomson Reuters. Thomson Reuters refused to license its content to a direct competitor. So Ross turned to a third party, LegalEase, which supplied "Bulk Memos" — compilations of legal questions and answers built, in turn, from Westlaw's copyrighted attorney-authored headnotes and its Key Number System taxonomy. Ross used roughly 25,000 Bulk Memos to train its search tool.

Judge Bibas, sitting by designation, revised his own prior 2023 summary judgment ruling, conceding that "a wise man knows when he is wrong." He granted summary judgment to Thomson Reuters on direct infringement and fair use, and denied Ross's cross-motion.

Copyrightability. The court confirmed that Westlaw's headnotes and Key Number System clear Feist's "extremely low" originality threshold. They are not mere copies of uncopyrightable judicial opinions; they are original selections and arrangements. (Factor two, the nature of the works, favored Ross only modestly, because the headnotes, while original, are "not as creative as a novel.")

Factor one — transformative purpose. Applying Warhol, the court held that Ross's use was not transformative. Ross did not add a further purpose or different character to the headnotes; it copied them to build a competing legal research tool serving the same purpose as Thomson Reuters's product. The fact that Ross's underlying technology differs from Thomson Reuters's did not change the analysis. The court also rejected the "intermediate copying" analogy to cases like Google v. Oracle, Sony v. Connectix, and Sega v. Accolade — those cases involved copying functional computer code necessary for interoperability and reverse engineering, not copying of expressive written works. The unprotected facts were available elsewhere; Ross copied the headnotes because doing so was easier, not because it was necessary.

Factor four — the decisive factor. The most consequential passage recognizes that there is an "obvious" potential market for licensing copyrighted works for AI training. Judge Bibas quoted Harper & Row for the principle that factor four "is undoubtedly the single most important element of fair use," and held that factor four — together with factor one — outweighed the factors that favored Ross. The existence of a growing licensing market for AI training data is, the court reasoned, precisely the kind of potential market harm factor four is designed to capture.

Defenses rejected. The court methodically rejected Ross's innocent infringement, copyright misuse, merger, and scènes à faire defenses — "none of Ross's possible defenses holds water."

Thomson Reuters v. Ross is the clearest signal yet that training a competing AI product on copyrighted works, without a license, will face serious fair use headwinds — especially where the AI's output competes in the same market as the original.

Bartz v. Anthropic: A Split Decision

Bartz v. Anthropic, No. C 24-05417 WHA (N.D. Cal. June 23, 2025), is the first decision to find that training a large language model (LLM) on copyrighted books can be fair use — but only where the works were lawfully acquired. The plaintiffs were authors (Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson) whose books Anthropic copied to train Claude, its flagship AI assistant.

The facts are striking. Anthropic's cofounder downloaded millions of books from pirate "shadow libraries" — Books3, Library Genesis (LibGen), and the Pirate Library Mirror (PiLiMi) — knowing they were unauthorized copies. Anthropic's CEO reportedly preferred "stealing" books to a "legal/practice/business slog." Later, Anthropic spent millions purchasing print books, stripped them from their bindings, scanned them, and discarded the paper originals to build a "central library" of digitized texts — what one Anthropic executive called a way to obtain "all the books in the world."

Judge Alsup's order on fair use split the conduct into three categories:

1. Training on lawfully acquired books — fair use. The court held that using the books to train Claude to "return new text outputs" was "spectacularly" (and "quintessentially") transformative. The purpose was not to reproduce the books or compete in the book market, but to enable a model to generate responsive text. This is a meaningful win for AI developers, and a contrast with Ross, where the AI tool competed directly with the copyrighted work's market.

2. Digitizing purchased print books for a central library — fair use. Anthropic's creation of digital copies of books it had lawfully purchased and then destroyed in physical form was, the court reasoned, a fair-use "replacement" of the physical copies — a kind of format-shifting. (This holding is narrow and arguably the most vulnerable on appeal.)

3. Downloading and retaining pirated books — not fair use. This was the decisive loss. Anthropic's acquisition of over seven million books from known pirate sites, and its retention of those copies in a permanent "central library," was "inherently, irredeemably infringing." There was no transformative purpose to the initial acquisition and permanent retention of pirated copies. The court's framing here echoes the principle animating Ross: where an obvious licensing market exists, bypassing it through piracy is not excused by the downstream transformative use.

Bartz thus draws a line that the industry will now organize around: lawful acquisition can support a fair use defense for training; piracy cannot. In late 2025 and 2026, Bartz culminated in a class settlement — preliminarily approved at a reported $1.5 billion — the largest copyright settlement in U.S. history, reflecting the magnitude of the piracy finding.

Kadrey v. Meta: A Reluctant Win for the AI Industry

Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025), arrived just two days after Bartz. The plaintiffs — a group of prominent authors including Richard Kadrey, Sarah Silverman, and Ta-Nehisi Coates — alleged that Meta infringed their copyrights by training its LLaMA models on their books, including copies obtained from shadow libraries.

Judge Chhabria denied the plaintiffs' motion for partial summary judgment and granted Meta's cross-motion, finding that Meta's training use was fair use on the present record. But the decision is unusually candid about its limits.

Reluctance, in the court's own words. The opinion opens by acknowledging that "in most cases the answer will likely be yes" to whether copying protected works to train AI is illegal, because copyright law "cares, above all else" about preserving the incentive to create. Yet the court granted Meta summary judgment because — on this record — the plaintiffs had failed to produce evidence of market harm sufficient to defeat fair use.

Factor four, again, is decisive. Because Meta's use was found to be highly transformative (training a model to generate text, not to reproduce books), the plaintiffs needed a "decisive win" on market harm to prevail. Judge Chhabria found their presentation "so weak that it does not move the needle." He was explicit that a better-developed record of market harm — including evidence that AI outputs dilute the market for the original works, or that a licensing market is being undercut — could prevail in a future case.

The "derivative works" theory rejected. The court dismissed the plaintiffs' theory that Meta's LLaMA models are themselves infringing derivative works merely because they were trained on copyrighted books. Training does not, without more, recast or transform the original works into a derivative work in the statutory sense.

Kadrey is best understood as a procedural warning dressed as a defense win. The court did not endorse blanket training-without-licenses; it found that these plaintiffs, on this record, had not carried their burden on market harm. The opinion practically invites the next plaintiff to build the record the Kadrey plaintiffs did not.

The Thread That Ties Them Together: Market Harm

Read together, the three decisions converge on a single proposition: factor four — the effect on the potential market for or value of the copyrighted work — is where AI fair use cases will be won and lost.

• In Ross, an obvious licensing market existed, and Ross bypassed it to build a competing product. Fair use failed.
• In Bartz, the transformative training use was excused — but the piracy, which circumvented a licensing market, was not.
• In Kadrey, the plaintiffs simply failed to prove market harm, but the court all but instructed future plaintiffs on how to do so.

This is consistent with the post-Warhol trajectory of fair use doctrine, which increasingly asks whether the secondary use is a substitute for, or competitor with, the original in a market the copyright holder reasonably seeks to exploit. The growing, well-documented market for licensing copyrighted works to train AI — with deals struck between major publishers and AI companies — makes the "potential market" inquiry concrete rather than speculative.

Transformativeness still matters, but its gravitational pull now bends toward market harm. A use that is transformative and lawfully acquired (the Bartz training scenario) is in the strongest position. A use that is transformative but built on pirated works (the Bartz library scenario, and the facts lurking in Kadrey) is in serious jeopardy. And a use that is not transformative at all — training a direct competitor on the original's expressive content (Ross) — is in the weakest position of all.

Practical Implications

For AI companies. The single most important risk-mitigation step is to lawfully acquire training data and document the acquisition chain. Bartz makes plain that the provenance of training data is not a technicality — it is the difference between a viable fair use defense and near-certain infringement liability. Where a licensing market exists (and for books, news, code, and images, it increasingly does), bypassing it is now a documented legal risk, not merely a business preference. Companies should also avoid retaining permanent "central libraries" of copyrighted works beyond what is needed for the transformative training purpose, and should be cautious about claims that AI outputs are non-infringing when those outputs can resemble the training data.

For copyright owners. The decisions confirm that factor four is the lever. To defeat a fair use defense, rights holders should develop a robust record of (1) the existence of a licensing market for AI training, including their own and comparable deals; (2) concrete evidence of market harm or dilution from the AI outputs; and (3) where applicable, that the defendant acquired the works from unauthorized sources. The Kadrey court's frustration was specifically with an underdeveloped market-harm record — a mistake future plaintiffs will not repeat.

For both sides. Settlement is increasingly the rational path. The Bartz settlement — a classwide, court-supervised resolution at historic scale — suggests that where piracy is in the record, the infringement finding may be largely a foregone conclusion and the real negotiation is over damages and a going-forward licensing framework.

Outlook

The Ninth Circuit will have the first crack at Bartz and Kadrey on appeal, and the Third Circuit at Ross. Appellate review will test whether the district courts' transformativeness analyses, and their factor-four emphases, hold. Several issues remain genuinely open: whether the Bartz "format-shifting" of purchased print books survives scrutiny; whether the Kadrey market-harm record can be rebuilt in parallel cases; and whether training on works scraped from the open web — a fact pattern none of these three cases squarely presented — is analyzed differently from training on books.

What is already clear is that the era of treating AI training fair use as an open question is over. The answer now depends on facts: who acquired what, how, from where, and what market the AI product ultimately competes in. For a firm that has litigated copyright and entertainment matters for over a decade, the takeaway is straightforward — the law is moving fast, and the facts of acquisition and market impact will decide most of these cases.

---

This article is for informational purposes only and does not constitute legal advice. If your company is developing AI products that rely on third-party content, or you are a rights holder concerned about unauthorized use of your works for AI training, contact our intellectual property team for a consultation.