Copyright and AI Collide: Three Key Decisions on AI Training and Copyrighted Content from 2025

The battle over whether U.S. copyright law permits artificial intelligence (AI) training on copyrighted works is no longer a theoretical debate. In 2025, three federal district court decisions began to sketch the boundaries of what counts as fair use in this context.

1. Bartz v. Anthropic: Training as Exceedingly Transformative

In Bartz v. Anthropic PBC, book authors challenged Anthropic’s use of their books, both lawfully acquired and pirated, to train the Claude large language models. On June 23, 2025, Judge William Alsup of the Northern District of California issued a summary judgment opinion in relation to whether various actions taken by Anthropic constitute fair use.

With regard to the use of copyrighted materials to train LLMs, Judge Alsup described the use as “transformative—spectacularly so,” which bears on the first fair use factor. The opinion analogizes model training to human reading and learning. That purpose, the court held, is fundamentally different, i.e., transformative, from the authors’ purpose in writing books to be read for entertainment or education purposes.

The same order blessed Anthropic’s “print-to-digital” process as fair use. Anthropic bought physical copies of books, scanned them, destroyed the paper copies, and stored searchable digital versions for internal use.

The opinion drew a sharp line, however, at Anthropic’s acquisition of millions of pirated books downloaded from shadow libraries. He viewed the acquisition of these books and the training of the LLM based on these books to be distinct steps, each of which would be evaluated with regard to fair use. He found that, while the training was fair, use, the acquisition of libraries of stolen books was not fair use on the record before him. There was no doctrinal “get out of jail free” card for downloading pirated copies simply because the intended end use (training) might be transformative. The Bartz case since settled based primarily on this decision, with Anthropic agreeing to pay up to $1.5 billion dollars.

2. Kadrey v. Meta: Fair Use But With a Market-Harm Warning Label

Two days later, in Kadrey v. Meta Platforms Inc., Judge Vince Chhabria of the Northern District of California likewise held that training an LLM on copyrighted books—including books obtained from shadow libraries—was fair use on the record before him. After discovery, the parties filed cross-motions for partial summary judgment on fair use. Judge Chhabria denied plaintiffs’ motion and granted Meta’s. But while Judge Chhabria held in favor of Meta, the takeaway for LLM developers was not all positive and was very factually specific to the evidence of record in that case.

On the first factor, the court agreed that using books to train the Llama models was highly transformative. The books exist to be read; Meta’s purpose was to extract patterns and statistical relationships from the corpus to power a general-purpose text generator. Unlike Judge Alsup, Judge Chhabria refused to carve the shadow-library downloads off as a separate, non-transformative act of piracy. He analyzed Meta’s copying of books from those sites together with its use of the copies to train Llama as a single instance of reproduction to be judged under fair use.

The reason Judge Chhabria’s decision wasn’t a complete victory for LLM developers like Meta lies in its discussion of the fourth fair use factor, which looks to the effect of the use on the potential market for the original work. He found no triable issue that Llama was regurgitating plaintiffs’ books in a way that would serve as a market substitute for them. He also concluded that the plaintiffs had not shown the existence of a well-defined, likely-to-develop licensing market for training on their books that Meta had usurped. The latter finding is especially interesting since the decision issued, there have been a number of licensing deals struck between copyright owners (like media companies) and LLM developers.

In that sense, Kadrey is as much a roadmap for future plaintiffs as a win for Meta: it affirms fair use in this instance while warning that a richer evidentiary record might tip factor four in a different direction.

3. ROSS Intelligence: When Training Cannibalizes the Market

If Bartz and Kadrey occupy one side of the developing fair use landscape, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. sits at the other. ROSS set out to build an AI-driven legal research tool that would compete directly with Westlaw. After Thomson Reuters refused to license its content, ROSS obtained “Bulk Memos” generated from Westlaw’s copyrighted headnotes and used them in its training pipeline. On February 11, 2025, Judge Stephanos Bibas, sitting by designation in the District of Delaware, granted partial summary judgment to Thomson Reuters, holding that the headnotes were protectable and that ROSS’s fair-use defense failed as a matter of law. To be clear – ROSS Intelligence differs substantively from Bartz and Kadrey in that while the technology at use in ROSS Intelligence is AI-related, it is not generative AI.

On the purpose-and-character factor, ROSS argued that its use was intermediate and transformative: headnotes were copied only to generate numerical weights, never displayed to users, and the resulting system was not a database of headnotes but an AI tool.

On the market-effect factor, the court emphasized that ROSS was trying to build a substitute for Westlaw using Westlaw’s own curated content. That kind of commercial substitution lies at the core of factor four’s concerns. With both the first and fourth factors cutting sharply against ROSS, the court held that the training use was not fair use.

The result in ROSS underscores that AI developers may be on shakier ground when their systems are designed to replicate a rightsholder’s core product rather than to generate new works that only indirectly compete with it. Notably, the ROSS Intelligence decision is currently on appeal in the Third Circuit, where the court will be deciding at least two core questions: whether short quotes/paraphrases of judicial holdings in headnotes are copyrightable at all, and whether fair use protects ROSS’s internal use of those headnotes as training data.

What These Cases Signal

Taken together, these three decisions show that U.S. fair-use doctrine is not marching in a single direction for AI training and it will take some time for appellate decisions to start providing a more unified approach. Whether training generative AI models on copyrighted material is infringing or constitutes fair use is moving target. The decision in Thomson Reuters v. Ross Intelligence is on appeal to the Third Circuit, and the decisions in Kadrey v. Meta (still in discovery) and Bartz v. Anthropic (settled) are not binding. Numerous other cases that address these issues are also making their way through the courts and may not result in binding precedent any time soon. In addition, output-based infringement may increasingly become a factor. Whether training a generative AI model constitutes fair use will, at the present moment, generally depend on a variety of factors, including how transformative the training and outputs are, and whether the outputs compete with, or dilute the market for, the original works. Of course, the output of the model may also infringe copyrighted material if it is identical or substantially similar to a copyrighted work, but consideration must also be given to whether the model output impacts the market for the work.

The courts take a dim view of uses (training or outputs) that substitute for the copyright owners’ work. Thomson Reuters v. Ross Intelligence held that that Westlaw’s headnotes and Key Number system are original works protectable by copyright and that Ross’s use of the copyrighted headnotes to train its own AI search platform was not fair use. The decision emphasized factor four of the fair use analysis set forth in 17 U.S.C. § 107—the infringing’s uses effect on the copyrighted work’s “potential market” or “value.” Bartz v. Anthropic’s finding of fair use relied on the determination that copies of the works “did not and will not displace demand for copies of the Authors’ works …” If the proposed use serves as a replacement for a paid product, the likelihood of successfully claiming fair use decreases.

Even if training generative AI models is considered “transformative,” use of tainted source material may result in liability. Recent decisions including Kadrey v. Meta (still in discovery) and Bartz v. Anthropic highlight the distinction between transformative training practices and wholesale copying from unauthorized sources such as shadow libraries. Even where training itself may be framed as fair use, reliance on pirated or unlicensed content can undermine that defense. Clean, well?documented sourcing is therefore critical to reducing exposure.

Future decisions may also flesh out the concern in regarding market harm, including dilution of the market for copyrighted works. While the fair use analysis has traditionally focused on whether the secondary use displaces sales of the original work, the copyright owner is also harmed if an AI system produces outputs that erodes the demand for the owner’s work. Kadrey v. Meta noted the danger of this “’indirect’ substitution” that may “enable the rapid generation of countless works that compete with the originals, even if those works aren’t themselves infringing.” The Copyright Office has taken a preliminary position that section 107 encompasses any “effect” on the potential market, and that the “speed and scale at which AI systems generate content pose a serious risk of diluting markets for works of the same kind as in their training data.” This dilution, according to the Copyright Office, can come in the form of works in the same genre as the copyrighted works or works that are stylistically similar.

Coping with Legal Uncertainty

Given the unsettled state of the law as of the end of 2025, overlooking any of these risks could result in significant legal exposure. And, as always, the viability of a fair use defense turns on a fact?specific application of the four statutory factors under 17 U.S.C. §?107. Accordingly, both the methods used in training and the potential as well as actual uses of generative AI should be carefully documented and subjected to rigorous legal analysis for possible copyright implications. Although this area of practice cannot yet be navigated without risk, early and deliberate attention to these issues provides the best foundation for managing uncertainty and mitigating exposure.