COPYRIGHT INFRINGEMENT IN ARTIFICIAL INTELLIGENCE TRAINING: AN ANALYSIS OF THE BARTZ v. ANTHROPIC SETTLEMENT IN THE U.S. FROM A COPYRIGHT LAW PERSPECTIVE
The issue of the unauthorized use of copyrighted works in the training of generative artificial intelligence models is one of the most controversial topics in copyright law worldwide. In the United States (“U.S.”), the most concrete and highest-stakes outcome of this debate emerged in the Bartz v. Anthropic case heard in San Francisco. In the lawsuit filed against the artificial intelligence company Anthropic PBC, alleging that it had used millions of books without permission to train its large language models named Claude, the parties reached a $1.5 billion class action settlement—the largest in U.S. copyright law history—which received final approval on July 20, 2026.[1]
The significance of the case extends beyond the monetary amount. The partial judgment issued on June 23, 2025, during the trial divided the use of works in AI training into two categories based on whether the works were obtained legally or from pirated sources; this distinction marked the first judicial assessment in U.S. law to delineate the boundaries of the fair use doctrine in the context of generative artificial intelligence.[2] This article will examine the specific factual background of the case, the court’s fair use assessment, the scope and binding nature of the settlement, and the potential implications under Turkish law.
I. The Origins of the Dispute and the Specific Facts of the Case
The lawsuit was filed in August 2024 by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson in the U.S. District Court for the Northern District of California.[3] As the proceedings progressed, the case evolved into a class-action lawsuit involving approximately 500,000 works registered with the U.S. Copyright Office and bearing ISBN or ASIN numbers.[4] Bartz alleged that four of her novels, Graeber two of his nonfiction works, and Johnson three of his nonfiction works had been copied and used without permission.[5]
At the factual core of the dispute lies a permanent digital library created by Anthropic while preparing to train its Claude models. According to internal company communications, the purpose of this library was to create a book archive that would be as comprehensive as possible and stored indefinitely ; novels and nonfiction works by the plaintiff authors were also included in this archive .[6] Anthropic obtained this content through two separate methods: on the one hand, it downloaded over seven million books for free from piracy platforms such as LibGen and Pirate Library Mirror (“PiLiMi”); on the other hand, it purchased millions of books—which partially overlapped with the former—through legal channels, disbound the volumes, scanned the pages, stored them in a digital and searchable format, and destroyed the printed copies.[7]
According to records, each work selected for training was reproduced through four distinct technical stages; even Anthropic itself has been unable to determine the total number of these reproductions with reasonable certainty. In contrast, the findings in the case file indicate that these training copies were not shared directly with third parties; for when a model was integrated into the publicly available Claude service, it operated alongside a separate filtering layer that monitored user inputs and model outputs. Therefore, the plaintiffs did not allege that the Claude service provided users with an infringing copy of their works; they based their claims on two separate grounds: the inclusion of the works in the library after being obtained from pirated sources, and the reproduction of the works during model training.[8]
The plaintiffs further argued that the unauthorized use of their works in model training could lead to the generation of text capable of competing with their own works and reducing demand, and that it could also eliminate a potential market for licensing the works for training purposes. The court, however, rejected this claim of market harm on the grounds that there was no concrete evidence that Claude produced output that competed with the plaintiffs’ works in a way that constituted infringement; it concluded that the competitive but non-infringing content produced by artificial language models did not constitute a legally protected harm under the fourth fair use factor (effect on the market).[9]
II. The Court’s Fair Use Assessment
In his partial ruling dated June 23, 2025, Judge William Alsup evaluated three distinct forms of use independently of one another. First, he found the use of works to train large language models to be highly transformative—on the grounds that the generated output serves a function entirely different and new from that of the source work—and thus deemed it to fall within the scope of fair use.[10] Second, the court also deemed the digitization of lawfully purchased printed books and their use in place of destroyed copies to fall within the scope of fair use. In its third and most decisive finding, the court ruled that copies downloaded from pirated platforms and stored indefinitely in a permanent library could not benefit from a fair use defense, and that this matter required separate adjudication.[11]
A key point to emphasize in the reasoning behind the decision is that the court did not separate its determination of fair use from the manner in which the work was acquired: while the transformative nature of educational use was acknowledged, this acknowledgment was deemed valid only with respect to copies obtained in accordance with the law.[12] In other words, the transformative nature of the educational activity was not deemed sufficient on its own; the legality of the work’s initial acquisition was accepted as a separate threshold. In August 2025, the court certified the case as a class action covering all rights holders whose works are included in the LibGen and PiLiMi datasets and meet certain criteria; however, it limited this certification solely to claims regarding the use of pirated sources. The finding of fair use with respect to educational activities remained binding only for the three authors who initiated the lawsuit.[13]
III. Scope and Binding Nature of the Settlement
While a trial was scheduled for December 2025 to determine damages for the use of pirated materials, the parties reached a $1.5 billion settlement on September 5, 2025, instead of proceeding to a jury trial.[14] The settlement received preliminary approval on September 25, 2025, and was definitively approved by Judge Araceli Martínez-Olguín on July 20, 2026. In its final approval order, the court noted that if the case had proceeded to trial, the plaintiffs would have faced the risk of an unfavorable outcome in light of the unraised claims presented, whereas the settlement provided significant benefits to class members.[15]
Under the settlement, the fund—comprising approximately 500,000 works included in the LibGen and PiLiMi datasets and duly registered with the U.S. Copyright Office—will be distributed equally among the class members after deducting attorneys’ fees and administrative expenses; an average amount of approximately $3,000 per work will be split equally between the publisher and the author—unless otherwise agreed upon in a contract.[16] However, the settlement does not overturn the ruling that educational activities constitute fair use. Indeed, in a statement made following the settlement, Aparna Sridhar, Associate General Counsel at Anthropic, emphasized that the finding that the use of books in AI training constitutes fair use under copyright law remains in effect.[17]
The key distinction to note here is that the settlement is not a court ruling but a voluntary agreement between the parties; therefore, it does not establish binding precedent.[18] The key legal determination shaping the industry is the partial ruling dated June 2025; the settlement, however, signifies that the parties chose to terminate the agreement due to the financial risk posed by that ruling—a risk that, given the statutory caps on damages for over seven million works derived from pirated sources, could have resulted in a significantly higher financial liability. As part of the settlement, Anthropic is also required to destroy copies obtained from pirated sources. Andrea Bartz, one of the plaintiff authors, expressed her hope that this process would be one of the first steps toward creating a fairer environment for creative professionals in the age of artificial intelligence.[19]
IV. Assessment Under Turkish Law
In terms of classification, the relevant point of comparison under Turkish law is the Law on Intellectual and Artistic Works No. 5846 (“FSEK”). Article 22 of the FSEK broadly defines the right of reproduction to include the reproduction of a work by any method, even in part; thus, the use of a work as input in model training will be considered direct reproduction without having to meet a separate doctrinal threshold, as is the case under U.S. law.
The main difference lies in the structure of the exception regime. Under U.S. law, fair use is a flexible defense mechanism based on the case-by-case weighing of four factors (the purpose and character of the use, the nature of the work, the proportion of the work used, and the effect on the market), and it allows courts to make decisions by considering the transformative nature of the use. In the FSEK, however, no such general exception exists; the limitations set forth in Article 30 and subsequent provisions are subject to the principle of limited enumeration and cannot be extended by analogy.[20] It does not appear possible for training a model on a commercial scale to fall under any of these exceptions. Consequently, there is no equivalent in Turkish law to the fair use defense that Anthropic can invoke in the U.S. with respect to works lawfully acquired; the use of works by a provider operating in Turkey for training purposes without consent would constitute reproduction under Article 22 of the FSEK, regardless of the method of acquisition.
With regard to the use of pirated sources, the two legal systems reach a similar conclusion: in both U.S. and Turkish law, obtaining a work from an unlawful source constitutes a separate legal violation, regardless of the purpose of use. Under the Copyright Law (FSEK), this situation may give rise to actions for the removal and prevention of infringement pursuant to Article 66, claims for damages of up to three times the value pursuant to Article 68, and criminal liability pursuant to Article 71 et seq. Furthermore, the claim of market damage discussed in the case file also warrants separate consideration under Turkish law: Since the claim for damages under Article 68 of the FSEK is calculated based on the normal usage fee of the work, independent of the infringement’s impact on the market, it can be argued that the “fourth factor” discussion in U.S. law does not serve the same function under Turkish law. With regard to the collective representation of rights holders, the functional equivalent of the class action mechanism in the U.S. is the claim process conducted through professional associations established under Article 42 et seq. of the Copyright Law; however, since Turkish law does not provide for a U.S.-style class action mechanism, rights holders will need to assert their claims individually or through professional associations.
V. Implications for Practice
The key lesson the Bartz v. Anthropic case leaves for the industry is that the method of acquiring training data creates a layer of risk independent of its intended use. The distinction between legally obtained sources and pirated sources is significant not only under U.S. law but also under any legal system that regulates the data supply chain. From the perspective of AI developers, ensuring that the source of the training dataset is verifiable and that supplier contracts include detailed provisions regarding representations and indemnity clauses related to the data source has become a minimum requirement for legal risk management.
The fact that the settlement does not establish binding precedent does not mean the matter has been definitively resolved. Similar lawsuits against other AI companies are currently ongoing in various federal courts, and how the distinction between fair use and pirated sources—as well as the debate over the fourth factor regarding market harm—will be addressed in other judicial districts and on appeal will be the defining question in the coming period.
VI. Conclusion
The outcome of Bartz v. Anthropic demonstrates that the training of artificial intelligence may be evaluated under the fair use doctrine, but that this evaluation remains contingent on the work having been lawfully obtained. The use of works obtained from pirated sources in education has been excluded from the fair use defense even under U.S. law and ultimately resulted in the highest-value copyright settlement in history. From the perspective of Turkish law, however, since the Copyright Law (FSEK) does not provide for a general exception similar to fair use, educational activities fall directly under the scope of the right of reproduction—regardless of the legality of the source. In this context, the risks for artificial intelligence developers operating in Turkey or subject to Turkish law extend to a broader scope than in the U.S.; the question of whether will provide for a legal text and data mining exception remains a pressing issue.
[1]Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement, The Authors Guild, July 20, 2026, https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/ (Accessed August 14, 2026).
[2]District Court Issues AI Fair Use Decision, Goodwin, June 23, 2025, https://www.goodwinlaw.com/en/insights/publications/2025/06/alerts-practices-aiml-district-court-issues-ai-fair-use-decision (Accessed August 14, 2026).
[3]Princeton University Press statement on the proposed Bartz v. Anthropic class action settlement, https://press.princeton.edu/news/barz-v-anthropic (Accessed August 14, 2026).
[4]What to Know About the $1.5 Billion Bartz v. Anthropic Settlement, Copyright Alliance, https://copyrightalliance.org/participating-bartz-v-anthropic-settlement/ (Accessed August 14, 2026).
[5]Bartz v. Anthropic PBC, 787 F.Supp.3d 1007 (N.D. Cal. 2025), vLex, https://case-law.vlex.com/vid/bartz-v-anthropic-pbc-1094702944 (Accessed August 14, 2026).
[6]Landmark Ruling on AI Copyright: Fair Use vs. Infringement in Bartz v. Anthropic, ArentFox Schiff, https://www.afslaw.com/perspectives/alerts/landmark-ruling-ai-copyright-fair-use-vs-infringement-bartz-v-anthropic (Accessed August 14, 2026).
[7]The Bartz v. Anthropic Settlement: Understanding America's Largest Copyright Settlement, Kluwer Copyright Blog, https://legalblogs.wolterskluwer.com/copyright-blog/the-bartz-v-anthropic-settlement-understanding-americas-largest-copyright-settlement/ (Accessed August 14, 2026).
[8]Landmark Ruling on AI Copyright: Fair Use vs. Infringement in Bartz v. Anthropic, ArentFox Schiff, n. 6.
[9]Two Federal Courts Rule That Reproduction of Books to Train LLMs Is Fair Use, But With Caveats, Fenwick, https://www.fenwick.com/insights/publications/two-federal-courts-rule-that-reproduction-of-books-to-train-llms-is-fair-use-but-with-caveats-and-strikingly-different-views (Accessed August 14, 2026).
[10]Landmark Ruling on AI Copyright: Fair Use vs. Infringement in Bartz v. Anthropic, n. 6.
[11]Judge Alsup Grants Partial Summary Judgment to Anthropic..., Chat GPT Is Eating the World, June 24, 2025, https://chatgptiseatingtheworld.com/2025/06/24/judge-alsup-grants-partial-summary-judgment-to-anthropic-ruling-training-copies-were-fair-use-but-judge-rules-no-fair-use-in-pirated-copies-of-books-used-to-build-a-central-library-they-are-infring/ (Accessed August 14, 2026).
[12]Anthropic Reaches Settlement with Authors in First-of-Its-Kind AI Copyright Infringement Lawsuit, NPR, September 5, 2025, https://www.npr.org/2025/09/05/nx-s1-5529404/anthropic-settlement-authors-copyright-ai (Accessed August 14, 2026).
[13]Bartz v. Anthropic Settlement: What Authors Need to Know, The Authors Guild, https://authorsguild.org/advocacy/artificial-intelligence/what-authors-need-to-know-about-the-anthropic-settlement/ (Accessed August 14, 2026).
[14]Anthropic Settles with Authors in First-of-Its-Kind AI Copyright Infringement Lawsuit, NPR, n.d. 11.
[15]Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement, p. 1.
[16]Federal judge in San Francisco approves $1.5 billion class settlement in Bartz v. Anthropic copyright case, Gaudêncio Advogados, https://www.gaudencioadvogados.com/en/news/federal-judge-in-san-francisco-approves-1-5-billion-class-settlement-in-bartz-v-anthropic-copyright-case/ (Accessed August 14, 2026).
[17]Federal judge in San Francisco approves $1.5 billion class settlement in Bartz v. Anthropic copyright case, Gaudêncio Advogados, no. 16.
[18]Judge Approves Anthropic's $1.5 Billion Book Piracy Settlement, the Largest in U.S. Copyright History, TNW, https://thenextweb.com/news/anthropic-1-5-billion-copyright-settlement-approved (Accessed August 14, 2026).
[19]Authors have mixed feelings about the $1.5B Anthropic copyright infringement ruling, NPR, https://www.npr.org/2026/07/27/nx-s1-5904606/anthropic-vs-bartz-ai-copyright-lawsuit-pros-cons (Accessed August 14, 2026).
[20]Article 35 of the FSEK pertains to specific and limited quotations; Article 34 pertains to anthologies and compilations prepared for educational and instructional purposes; and Article 38 pertains to the condition of non-profit status.