Case Summary: Bartz v. Anthropic PBC, No. 24-05417 (N.D. Cal. Jun. 23, 2025)

Written by Ava Nesci

Edited by Madeline Leonard & Maclain Conlin

Bartz v. Anthropic PBC.pdf

Download PDF • 140KB

(All views expressed in this article are those of the author and do not necessarily represent the position of the Clemson Law Review.)

As a new, novel, and rapidly developing technology, artificial intelligence (AI) has recently come to the center of many legal debates, raising complex questions that the legal field is only just beginning to address. Copyright law has been at the heart of many of these early cases, and the District Court of Northern California’s recent order in Bartz v. Anthropic PBC on fair use in accordance with section 107 of the Copyright Act[1] has laid a foundational brick for building upon this area of law. In Bartz v. Anthropic PBC, the court considered how copyright infringement and fair use doctrines apply when research companies use copyrighted works to train AI systems. In the case, plaintiffs and authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed suit against Anthropic PBC for copyright infringement after discovering that the company had used both purchased and pirated copies of their books to train large language models (LLMs) to create an interactive AI named Claude and to create a central library as a general-purpose resource.[2] The dispute thus focused on the use of the author’s materials as inputs to train the LLM, not the outputs produced by Claude.[3] The court’s analysis of the use of purchased materials to train AI is significant in determining how the courts will analyze similar cases going forward.

Under the Copyright Act, the reproduction of a copyrighted work for the purposes of issuing “criticism, comment[ary], news reporting, teaching, . . . scholarship, or research” is considered a fair use, and “is not an infringement of copyright.”[4] The Copyright Act is rooted in Article 1, Section 8 of the Constitution, which states that “Congress shall have power . . . to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.”[5] Thus, in order to determine whether or not the use of a copyrighted material is fair, the court applied the four part analysis required by the Copyright Act, which requires the following: First, the court must consider the purpose and character of the use. Secondly and thirdly, the court must determine the nature of the copyrighted work and the amount of the copyrighted material used. Finally, the effect of the use on the value of the copyrighted work must be taken into consideration.[6] Moreover, when a copyrighted work is used in multiple ways, the court must analyze each infringing use separately.

The Supreme Court in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith established that when a copyrighted material is used in multiple ways, each use must be analyzed separately.[7] In accordance with the precedent from Warhol,[8] the district court found and analyzed three uses: training the LLM,[9] converting purchased copies of the author’s books from print to digital to create a central library as a permanent, general-purpose resource,[10] and downloading pirated copies of the author’s books to contribute to the general-purpose library.[11] Anthropic was not entitled to use pirated copies of the authors books, and thus the court’s analysis focused primarily on the first two uses.

The first factor of the analysis considers the purpose and character of the use, which requires the court to look at whether or not the use is transformative. To be transformative, the user must apply some degree of creativity to produce something different from the copyrighted material. However, in Feist Pubs, Inc. v. Rural Tel. Serv. Co., the Supreme Court acknowledged that the originality threshold is “extremely low,” requiring only “some minimal degree of creativity” to be considered transformative.[12] In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., a federal district court judge in Delaware refused to recognize fair use due to a lack of transformation because the AI simply reiterated relevant information from the copyrighted source when prompted as opposed to writing new content itself.[13] 

In contrast, Anthropic’s use of the material for training purposes was transformative. Anthropic used copies of the authors’ copyrighted works to iteratively map statistical relationships between text-fragments, so that the LLM, Claude, could receive text inputs and respond as if it were a human.[14] Moreover, the court in the Anthropic case equated fair use by humans to that of AI, reasoning that “like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.”[15] The authors argued against this, claiming that “computers nonetheless should not be allowed to do what humans do,” but the court rejected this notion and applied the traditional fair use analysis. The court further supported this reasoning, citing three influential circuit court precedents—Nichols v. Universal Pictures Corp., Apple Comput., Inc v. Microsoft Corp., and Swirsky v. Carey—which held that copyright does not extend to protect methods of “operation, concepts, or principles.”[16] The court also found that Anthropic’s conversion of the purchased materials from print to digital to keep in their library was transformative enough to favor fair use under the first factor.[17]

Next, the court applied the second factor of the analysis, which looks at the nature of the copyrighted work. The court cited Campbell v. Acuff-Rose Music, Inc., which recognized that “some works are closer to the core of intended copyright protection.”[18] More protection is afforded to published works, than to unpublished ones, and more protection is afforded to works of fiction than works of fact. Since the authors books were published and chosen by Anthropic because of their expressive elements, the court determined that the second factor points against fair use.[19]

Then, the court applied the third analysis factor, considering the amount and substantiality of the copyrighted work which was used to determine whether this amount was reasonably necessary in light of Anthropic’s purposes.[20] In light of Anthropic’s requirement for a monumental amount of material to train its LLM, and the requirement that Anthropic needed only to prove that their use was reasonably necessary, and not strictly necessary, the court ruled in favor of fair use. The court also ruled in favor of fair use for the digital copying of purchased books to store in their library, since maintaining the totality of each work was necessary for Anthropic’s research purposes.[21]

Finally, the fourth factor of the analysis considers the effect of the use on the value of the copyrighted materials. The court found that this factor was neutral for the conversion of purchased copies from print to digital, and ruled in favor of fair use for training purposes, specifically noting that “the copies used to train specific LLMs did not and will not displace demand for copies of Authors’ works, or not in the way that counts under the Copyright Act.”[22] The court dismissed the authors concerns that this use may lead to an explosion of competing works, reasoning that it is no different than school children using their works to learn to write well and is thus not a use which copyright holders can restrict.[23]

This decision provides insight into how courts going forward will handle copyright law as applied to AI. In determining fair use thus far, the courts have treated AI learning as equivalent to human learning, as shown by the court’s decision and rationale here to reject the author’s claim that a distinction ought to be made between humans and computers in to the application of copyright law; just as a human may learn from books and emulate a blend of their writing, computers may learn from and emulate the writing styles of various books.[24] 

In the still emerging area of artificial intelligence law, subjecting AI to the same laws and analysis humans are subject to under existing copyright law offers the most workable framework. However, the authors in this case highlight the key concern that humans and computers are not the same, and argue that they should not be treated as such. The authors in the Anthropic case specifically challenged the use of their books as inputs to train an AI, and emphasized the great extent to which the LLMs memorized components of their work,[25] suggesting that this capacity significantly differentiates AI use from human use. Although the court dismissed this argument, it may become relevant in light of the original intention behind the Copyright Act if authors and creators continue to bring suits against the use of their work to train AI. As developers continue to utilize copyrighted materials to train and improve AI, and people become more reliant on it, authors and creators may become disincentivized to create.

In a current events conversation article from The New York Times, Ricardo Nagaoka compiled a number of testimonies from how students are using AI.[26] The responses included using AI to answer questions, generate ideas, provide sample introductions for essays to aid students in learning, and developing their writing. Another article by Paul McQuiston from USC Today, similarly, looks to how college students are using AI. McQuiston explains that some students are using AI to cultivate their writing skills and mentions a tool called ABE, which is designed to help students through promoting reflection and revision throughout the writing process.[27] While helpful, this trend towards using AI to develop writing skills may degrade the value of the works which were used to train the AI, and despite the ruling in Anthropic, may elevate the importance of considering these additional factors under the fourth prong of the fair use analysis, which considers the effect of the use upon the potential market for or value of the copy.

[1] 17 U.S. Code § 107.

[2] Barts v. Anthropic PBC, No. 24-05417, slip op. at 2 (N.D. Cal. Jun. 23, 2025).

[3] Id. at 12.

[4] 17 U.S. Code § 107.

[5] U.S. Const. Art. I § VIII. 

[6] 17 U.S. Code § 107(1)–(4). The four part analysis considers: “(1) the purpose and character of the use, including whether such use if of a commercial nature or is for nonprofit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.” Id.

[7] Andy Warhol Foundation for Visual Arts, Inc v. Goldsmith, No. 21–869, slip op. at 20 (U.S. May 18, 2023) (“The fair use provision, and the first factor in particular, requires an analysis of the specific ‘use’ of a copyrighted work that is alleged to be ‘an infringement’. . . .The same copying may be fair when used for one purpose but not another.”) (cleaned up).

[8] Barts v. Anthropic PBC, No. 24-05417, slip op. at 10 (N.D. Cal. Jun. 23, 2025)

(“These factors presuppose a “use.” So, at the threshold, a court must decide whether a “copyrighted [work] has been used in multiple ways,” then evaluate each.” Warhol, 598 U.S. at 533).

[9] Id. at 11 (“Anthropic used copies of Authors’ copyrighted works to iteratively map statistical relationships between every text-fragment and every sequence of text-fragments so that a completed LLM could receive new text inputs and return new text outputs as if it were a human reading prompts and writing responses.”).

[10] Id. at 14 (“[Anthropic] retained all the copies in its central library for other uses that might arise even after deciding it would not use them to train any LLM (at all or ever again).”).

[11] Id. at 18. (“Anthropic downloaded over seven million pirated copies of books, paid nothing, and kept these pirated copies in its library even after deciding it would not use them to train its AI (at all or ever again).”).

[12] Feist Pubs, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 345 (1991).

[13] Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 765 F. Supp. 3d 382, 398 (D. Del. 2025).

[14] Barts v. Anthropic PBC, No. 24-05417, slip op. at 11 (N.D. Cal. Jun. 23, 2025).

[15] Id. at 13.

[16] Id. at 13 (cleaned up).

[17] Id. at 16.

[18] Id. at 24. (“This factor ‘calls for recognition that some works are closer to the core of intended copyright protection than others, with the consequence that fair use is more difficult to establish when the former works are copied.’ Campbell v Acuff-Rose Music, Inc., 510 U.S. at 586 (1994).”).

[19] Id. at 24.

[20] Id. at 25.

[21] Id. at 26.

[22] Id. at 28.

[23] Id. 

[24] Id. at 13.

[25] Id. at 11 (“Authors further argue — and this order takes for granted — that such training entailed ‘memoriz[ing]’ works by ‘compress[ing]’ copies of those works into the LLM. . . .The LLMs “memorize[d] A LOT, like A LOT”. . . .Regardless, the ‘purpose and character’ of using works to train LLMs was transformative — spectacularly so.”).

[26] Ricardo Nagaoka, What Students Are Saying About Using A.I. for Schoolwork, N.Y. Times, May 8, 2025, https://www.nytimes.com/2025/05/08/learning/what-students-are-saying-about-using-ai-for-schoolwork.html, accessed November 10, 2025.

[27] Paul McQuiston, AI is Changing How Students Learn – Or Avoid Learning, USC Today, Sept. 18, 2025, https://today.usc.edu/ai-is-changing-how-students-learn-or-avoid-learning/, accessed November 11, 2025.