Meta Faces Growing Copyright Lawsuits Over the Use of Pirated Books to Train AI

Meta's use of millions of books to develop its artificial intelligence systems has placed some of America's most prominent writers at the center of a far-reaching copyright dispute. Although a federal judge sided with the technology company on key claims, allegations involving pirated libraries, unauthorized file sharing, and uncompensated creative work remain unresolved. With hundreds of additional authors pursuing litigation, the case could have lasting consequences for the publishing industry's finances and the legal protections governing literature in the age of artificial intelligence.

In the spring of 2023, as Meta accelerated the development of its artificial intelligence systems, the company confronted a problem that had little to do with computing power. Its engineers needed books. The sustained narratives, grammatical consistency, and carefully organized language of published literature offered qualities that other sources of training data could not reliably provide. Meta had explored purchasing the necessary rights, with its head of generative AI discussing a licensing budget of as much as $100 million. Negotiations with major publishers had produced few results. By then, the company had already downloaded material from Library Genesis, an unauthorized digital repository containing vast quantities of copyrighted books.

After the licensing effort faltered, the question of using that collection was escalated to Mark Zuckerberg. Meta subsequently proceeded with books obtained from Library Genesis, commonly known as LibGen, to train its Llama language models. The decision would become central to a copyright lawsuit filed in July 2023 by writers Richard Kadrey, Sarah Silverman and Christopher Golden, eventually joined by other authors whose works had been acquired without their permission. Documents disclosed during the litigation revealed the extent to which Meta had considered conventional licensing arrangements before turning to unauthorized repositories, bringing the company's internal decisions and commercial priorities under unusually close judicial examination.¹

The case produced a result that surprised many observers. On June 25, 2025, U.S. District Judge Vince Chhabria ruled that Meta's use of thirteen authors' books for artificial intelligence training qualified as fair use under the evidence presented. His opinion nevertheless expressed serious concern about the economic consequences of generative AI for professional writers, particularly the possibility that systems trained on copyrighted literature could eventually produce enough competing material to diminish the commercial value of human-authored books. The plaintiffs had failed, he concluded, to establish that injury with evidence sufficient to overcome Meta's defense.²

More than a year later, Kadrey v. Meta Platforms remains before the U.S. District Court for the Northern District of California. Allegations concerning the redistribution of copyrighted material through BitTorrent networks have survived the ruling on training, and a March 2026 order permitted the plaintiffs to pursue an additional theory of contributory infringement. Hundreds of authors have since filed separate lawsuits against Meta, expanding the legal challenge to the company's acquisition and use of literary material.³,⁴

The proceedings have exposed difficult questions about the rights attached to published work and the circumstances under which those rights can be exercised. Copyright law distinguishes among reproduction, distribution and other protected uses, each of which may carry different legal consequences. The outcome of a dispute over training an artificial intelligence model cannot necessarily establish whether the methods used to obtain the underlying material were lawful. That distinction has become increasingly consequential as technology companies develop commercial systems using books whose value was established through the labor and investment of authors, editors and publishers.

Meta's interest in published literature was closely connected to the technical ambitions of Llama. Language models are trained on extensive collections of text, allowing them to identify relationships among words, sentences, and longer passages. The quality and organization of that material influence their capabilities. Books offered a particularly useful source of sustained linguistic structure for engineers attempting to develop systems that could process lengthy documents, maintain context across extended interactions, and produce coherent responses.

The court's account of Meta's development process describes repeated internal discussions about acquiring additional literary material. Employees identified books as valuable training resources and emphasized the urgency of obtaining them. Meta's expert witnesses explained that published works were useful for improving a model's ability to process extended sequences of text, partly because they maintained relatively consistent grammatical and organizational structures. These qualities distinguished books from much of the fragmented writing available through ordinary internet collection.¹

Meta's early training datasets already incorporated material from a variety of sources. According to the judicial record, approximately two-thirds of the data used to train Llama 1 and Llama 2 came from Common Crawl, a nonprofit organization that collects publicly accessible web content. Other sources included Wikipedia, GitHub, arXiv, Stack Exchange, Project Gutenberg and Books3, the last of which attracted controversy over its inclusion of copyrighted literary works.

The company released Llama 1 in February 2023 and Llama 2 that July, making the models available under licensing arrangements that permitted various research and commercial uses. In April 2024, Meta introduced Llama 3 alongside an expanded version of Meta AI, integrating its generative technology more closely with products used by billions of people. The developments occurred during an increasingly competitive period in artificial intelligence, when access to larger and more sophisticated datasets was becoming a priority for companies seeking improvements in model performance.

Published books occupied a distinctive position within that competition. Their usefulness reflected the accumulated labor of writers, editors, translators, and publishers who had developed and refined complex works over months or years. Literary fiction offered narrative continuity, developed characters and sustained thematic organization. Nonfiction supplied structured arguments, specialized knowledge, and extended explanations. Academic publishing contributed technical precision and carefully organized research. The intellectual and editorial investment that made these works valuable to readers also made them attractive resources for artificial intelligence developers.

Meta initially pursued conventional licensing arrangements. Company representatives contacted major publishers, and the head of generative AI discussed spending as much as $100 million to secure rights to literary material. The negotiations encountered obstacles associated with the structure of publishing contracts. Publishers frequently lacked authority to license books for artificial intelligence training because the relevant rights remained with individual authors. In other instances, contractual permissions were geographically restricted. Some companies did not respond to Meta's inquiries, while only one reportedly provided a pricing proposal.¹

These difficulties reflected the intricate allocation of intellectual property rights within traditional publishing. An author may grant a publisher permission to produce print editions, electronic books and audiobooks while retaining other subsidiary rights. Translation agreements, adaptations and international publishing arrangements frequently involve additional contractual divisions. Many of those agreements were negotiated before commercial artificial intelligence training had become a significant market, leaving considerable uncertainty about which parties possessed authority to license such uses.

Meta's licensing difficulties were therefore partly a consequence of the contractual arrangements governing the books it wanted. Securing permission required identifying the relevant rights holders, determining the extent of their authority and negotiating agreements covering the company's intended activities. The judicial record establishes that these efforts encountered substantial resistance and practical complications. It does not establish that lawful licensing was unavailable.

In October 2022, Meta downloaded LibGen to evaluate its usefulness for language-model development. The company initially contemplated testing the collection and subsequently seeking licenses for the material or comparable works. By spring 2023, after its negotiations had failed to produce the desired results, Meta escalated the issue to Zuckerberg and decided to proceed with books obtained from the repository. The company later acquired additional material from Anna's Archive, another online collection associated with unauthorized copies of books and academic publications. By early 2024, according to the court's findings, material from that source had also been downloaded for Meta's development efforts.¹

The sequence became central to the authors' allegations. LibGen and Anna's Archive made literary works technically accessible without necessarily possessing authorization to reproduce or distribute them. A downloadable file could provide Meta with the text required for training while leaving the copyright owner's exclusive rights intact. The company's decision placed longstanding questions about digital piracy within a commercial undertaking whose financial ambitions were considerably greater than those of the repositories themselves.

The scale of those ambitions emerged in evidence considered by the court. Internal projections discussed during the proceedings anticipated generative AI revenues of $2 billion to $3 billion in 2025, with substantially larger potential earnings over the following decade. These were forecasts rather than realized financial results, but they indicated the commercial importance Meta attached to its AI development. The authors whose books entered the training datasets had received no negotiated compensation for the uses at issue.

The legal challenge required Chhabria to examine whether reproducing the books constituted infringement and whether the purpose of training Llama brought that conduct within the protections of fair use. Meta did not dispute, for purposes of the summary judgment motions, that the plaintiffs had established an initial case of infringement of their reproduction rights. Its defense depended upon the statutory exception permitting certain unauthorized uses of copyrighted material.²

Fair use requires courts to consider the purpose and character of the challenged use, the nature of the original work, the amount copied and the potential effect upon the market for the copyrighted material. The doctrine has long accommodated activities such as criticism, scholarship, research and commentary, while also allowing courts to evaluate new technological uses. Its application to generative artificial intelligence has required careful consideration of systems that incorporate protected expression into computational processes capable of producing new text.

Meta maintained that training Llama constituted a transformative use of the authors' works. The company had incorporated the books into a general-purpose language model designed to perform a variety of tasks, rather than developing a service intended to sell or reproduce the original publications. The plaintiffs disputed the significance of that distinction, emphasizing the unauthorized acquisition of their work and its incorporation into commercial products.

Chhabria agreed that the training use was highly transformative, although he acknowledged its commercial character. The purpose of Meta's copying differed substantially from the purposes for which the authors had created their books. The model could assist users with writing, research, programming, and other activities, and the evidence did not establish that it functioned as a substitute reading service through which users could retrieve substantial portions of the original works.

The court also considered tests designed to determine whether Llama could reproduce copyrighted passages from its training material. Meta had implemented technical measures intended to reduce such reproduction, and the experiments did not demonstrate that users could reliably recover significant portions of the plaintiffs' books. Chhabria found the limited outputs insufficient to establish meaningful substitution for the originals.

The analysis required attention to the distinction between reproducing protected expression and generating material possessing similar stylistic characteristics. Copyright generally protects an author's particular expression rather than a literary style considered independently. A model's capacity to imitate elements of a writer's prose could raise substantial concerns about commercial competition without necessarily establishing infringement of the original text.

The remaining statutory considerations produced a more complicated assessment. The creative nature of the plaintiffs' books favored the authors, and the court recognized the substantial original expression contained in their work. Meta had also copied complete books, a factor ordinarily relevant to determining whether a use exceeds what its purpose reasonably requires. Chhabria concluded that the extent of copying was justified by the demands of model training, given the importance of extended linguistic structures to the development process.

The greatest difficulty arose in assessing the commercial effects of Meta's conduct. The plaintiffs argued that Llama could reproduce portions of their books and that unauthorized training had deprived them of licensing opportunities. The court found insufficient evidence that the limited reproduction demonstrated through testing would materially affect sales. It also declined to treat the loss of a hypothetical licensing fee as independently decisive when the underlying activity might otherwise qualify as fair use.

Chhabria devoted greater attention to another possible injury that the plaintiffs had inadequately developed. A generative system trained on copyrighted literature might produce substantial quantities of new books capable of competing with human-authored works. Even where the generated material contained no recognizable passages from a particular title, it could attract readers who might otherwise purchase original publications. Large volumes of inexpensive or rapidly produced material could also affect visibility within online marketplaces, making established works more difficult to discover.²

The consequences could vary considerably among literary categories. Readers seeking a particular author's memoir may place considerable value on the writer's identity and experience, while consumers of genre fiction or commercial nonfiction may be more willing to consider alternative works offering similar subject matter or narrative conventions. The economic effects of generative AI could therefore depend upon the characteristics of individual markets, purchasing habits and the capabilities of the models producing the competing material.

The plaintiffs had offered little evidence addressing those relationships. Their arguments lacked a developed analysis of the markets in which their books competed, the extent to which Llama could generate viable substitutes or the degree to which incorporating specific copyrighted works contributed to that competition. Meta presented expert testimony indicating that the release of Llama 3 had produced no discernible reduction in sales of the plaintiffs' books during the period examined.

The court concluded that the authors had failed to establish a sufficient factual basis for finding market injury. Chhabria's ruling was expressly limited to the plaintiffs and the evidentiary record before him. The decision left open the possibility that other authors, presenting different evidence of commercial harm, could prevail against similar fair-use defenses.

The qualification was especially significant because the judge recognized the potential seriousness of the economic threat. Generative AI could affect literary markets through the production of competing works, including material that did not directly reproduce the books used during training. The authors had failed to demonstrate that effect in the particular litigation, leaving the court to grant Meta summary judgment despite the concerns identified in its opinion.

On June 27, 2025, Chhabria also granted Meta summary judgment on a related claim under the Digital Millennium Copyright Act. The plaintiffs had alleged the unlawful removal of copyright management information, but their theory failed in light of the court's determination concerning the training-related copying. Separate allegations involving the distribution of copyrighted material remained unresolved.⁵

Those allegations concerned the technology Meta had used to obtain books from unauthorized repositories. The company downloaded material through BitTorrent, a peer-to-peer file-sharing protocol that allows participants to exchange portions of large files across a distributed network. Rather than receiving an entire file from a single server, a user can acquire individual segments from numerous computers and assemble them locally. The software may simultaneously upload previously downloaded segments to other participants.

The arrangement makes the technical details of acquisition legally consequential. A participant may transmit portions of a file while still downloading additional material, a process described in the proceedings as leeching. Uploading may also continue after the download is complete, commonly known as seeding. In either situation, the user's computer may contribute to the distribution of copyrighted material beyond the copies retained for its own purposes.

The judicial record establishes that Meta used BitTorrent to acquire material from LibGen and Anna's Archive. An engineer developed a script intended to prevent seeding, although that precaution did not necessarily stop uploads occurring during active downloads. The plaintiffs alleged that Meta's participation in these networks caused copyrighted works to be redistributed to other users. Meta disputed whether the evidence demonstrated that any of the plaintiffs' specific books had been transmitted.¹

The allegations introduced legal questions distinct from those resolved in the fair-use ruling. Direct infringement of distribution rights would require proof that Meta unlawfully distributed protected material. Contributory infringement could arise if the company knowingly and materially contributed to infringing conduct undertaken by other network participants. Both theories depend upon evidence concerning the operation of the software, the files exchanged and the company's role in making them available.

The distinction acquired greater significance as the lawsuit progressed. A finding that copying books for training qualified as fair use did not necessarily determine whether uploading those same works to other users was permissible. The technical process through which Meta acquired its data could therefore expose the company to liability independent of the legal status of its trained models.

On March 25, 2026, Chhabria permitted the plaintiffs to file a fourth amended complaint adding a claim for contributory infringement. His order sharply criticized their attorneys for failing to advance the theory earlier, given that the underlying allegations concerning BitTorrent uploads had been known for some time.³

The plaintiffs had sought to add direct distribution claims in November 2024. They subsequently argued that later discovery clarified the extent of Meta's involvement in uploading material, justifying an additional claim that the company had contributed to infringement by others. Chhabria rejected much of their explanation for the delay, concluding that counsel had already possessed sufficient information to recognize the potential theory.

The court also examined a related lawsuit brought by Entrepreneur Media, which had advanced a comparable contributory-infringement claim against Meta. Because discovery in the proceedings had been coordinated, and Meta was already defending the same general theory elsewhere, Chhabria determined that allowing the amendment would impose relatively limited additional prejudice.

The decision also reflected concern for authors who might eventually become members of a certified class. Although no class had been certified, denying the amendment could complicate the rights of individuals whose claims might later be affected by the proceedings. Chhabria granted permission to add the contributory-infringement claim, modify the proposed class definition, and include three additional business entities as plaintiffs. He declined to authorize class discovery at that stage.³

The order marked another stage in the lawsuit's transformation. The initial dispute had concentrated on whether Meta could lawfully train Llama using copyrighted books. The remaining claims required examination of how the books had been obtained, whether the company's computers transmitted copyrighted files, and whether its participation in peer-to-peer networks facilitated unlawful distribution.

The expanding litigation became still more consequential in October 2026, when nearly 300 authors filed individual copyright lawsuits against Meta in the Northern District of California. Attorneys from Freedman Normand Friedland and Stris & Maher indicated that additional claims could follow on behalf of approximately 1,000 writers. The actions involved fiction writers, nonfiction authors and journalists, reflecting the breadth of literary material potentially incorporated into Meta's development process.⁴

The individual filings represented a departure from reliance upon the proposed class action in Kadrey. Each author could pursue claims based upon particular copyrighted works, registrations and alleged conduct. The additional proceedings might also allow plaintiffs to develop evidence of commercial injury different from that considered in Chhabria's 2025 ruling, whose application had been expressly limited to the thirteen authors before the court.

Some of the new complaints introduced allegations concerning Meta's acquisition and scanning of printed books. According to reporting by Publishers Weekly, plaintiffs claimed that the company had purchased physical volumes and reproduced their contents for training without obtaining the necessary permissions. Those allegations remain unadjudicated and require separate consideration from the acquisition practices established in the original Kadrey proceedings.⁶

The legal distinction matters because purchasing a physical book ordinarily transfers ownership of that copy without transferring the exclusive copyright governing reproduction of its contents. Scanning a purchased volume may therefore implicate rights beyond those obtained through the transaction, subject to any applicable exceptions or defenses. The new lawsuits may require courts to consider the circumstances of particular acquisitions and the purposes for which the resulting digital copies were created.

The authors are seeking statutory damages, which may be available under federal copyright law when the relevant legal conditions are satisfied. In cases of willful infringement, the statute permits awards of up to $150,000 per infringed work, although entitlement to such damages depends upon matters including registration, the circumstances of the infringement and judicial findings. The filing of numerous claims does not establish liability or determine the amount that any plaintiff might recover.

By October 19, Magistrate Judge Thomas S. Hixson had issued a discovery order addressing disputes across Kadrey and several related proceedings. The coordinated treatment of discovery reflects substantial overlap among the cases, particularly concerning Meta's acquisition of training material and the internal records documenting those practices.⁷

The increasing number of lawsuits has placed Meta's development process under sustained legal examination, with different plaintiffs pursuing claims based on reproduction, distribution and the acquisition of protected material. The outcome of one proceeding may influence arguments in others, although differences in the works involved, the evidence presented and the rights asserted will remain significant.

A separate copyright dispute involving Anthropic has provided an important comparison. In Bartz v. Anthropic, authors challenged the company's acquisition and use of copyrighted books to develop its Claude language models. In June 2025, U.S. District Judge William Alsup concluded that training on copyrighted books constituted fair use under the circumstances examined, while separately finding that Anthropic's acquisition and retention of pirated books in a central library was not justified by that purpose.⁸

The proceedings subsequently produced a $1.5 billion settlement. On July 20, 2026, U.S. District Judge Araceli Martínez-Olguín granted final approval to the agreement, establishing one of the largest financial resolutions of a copyright dispute involving generative artificial intelligence and literary works.⁹

The settlement demonstrated the substantial potential financial consequences of unauthorized book acquisition, although it did not establish a general requirement that every AI training use must be licensed. The circumstances differed from those in Kadrey, particularly concerning the retention of pirated material and the allegations involving BitTorrent distribution. The cases nevertheless illustrate how courts may examine separate acts of reproduction, storage and distribution independently of the ultimate technological purpose for which the material was acquired.

For authors and publishers, the litigation has also raised questions about the economic value of literary work in markets increasingly influenced by generative systems. Copyright has traditionally supported compensation through advances, royalties, licensing arrangements and subsidiary rights. Artificial intelligence introduces additional commercial uses, some of which may occur without producing recognizable reproductions of the original text.

The economic consequences remain difficult to measure. A model trained on fiction may generate stories that contain little protected expression from any individual source while competing for the attention and spending of readers. That competition could affect different literary categories in substantially different ways. A historical biography, a technical reference book, and a popular romance novel serve different markets, and their susceptibility to substitution by AI-generated material may vary accordingly.

Chhabria's discussion of market dilution identified the difficulty of proving those effects. Evidence that AI-generated books are becoming more numerous would not necessarily establish that the inclusion of a particular author's work in a training dataset caused a measurable loss. A persuasive claim would require analysis of consumer behavior, the markets for comparable books, the capabilities of specific models, and the relationship between the training material and the resulting competition.

The problem is especially complicated because commercial injury may emerge gradually. Readers' purchasing habits can change over years, while publishers may adjust acquisition budgets, commission fewer works or redirect investment toward categories perceived as more commercially secure. Such decisions could influence authors' earnings without producing an immediate decline in sales attributable to a single model's release. Determining whether those developments result from artificial intelligence, broader market conditions or a combination of influences would require careful economic investigation.

The litigation also exposes uncertainty within publishing agreements negotiated before generative AI became commercially significant. Authors and publishers may disagree over who controls the right to license books for training, how to allocate compensation, and whether contractual language drafted for earlier technologies covers these uses. Meta's unsuccessful negotiations demonstrated the practical difficulty of resolving such questions across extensive collections of copyrighted works.

The legal consequences will vary internationally. Copyright systems differ in their treatment of text and data mining, licensing, exceptions, and enforcement. European law provides specific arrangements for certain automated uses of protected material, while other jurisdictions apply different statutory and contractual frameworks. Decisions issued by American federal courts do not establish binding international copyright rules, although they may influence commercial practices adopted by multinational technology companies.

The distinction between direct legal authority and commercial influence is important. Meta's decisions about acquiring data, negotiating licenses, and managing infringement risk affect operations extending beyond the United States. The resolution of American litigation may therefore influence how the company and its competitors approach literary rights in other markets, even where local law requires a different analysis.

For the publishing industry, the implications extend into contractual negotiations that will shape the ownership and exploitation of literary work for years to come. Authors need clarity about which rights they retain and how those rights may be licensed. Publishers must establish the authority under which they negotiate agreements involving new technologies. Developers must evaluate the legality of acquiring, retaining and using copyrighted material, including the methods by which it enters their systems.

The legal disputes surrounding Meta have revealed how difficult those questions become when the commercial value of literary expression is realized through technologies that transform its original form. The text of a book may contribute to the development of a model without appearing recognizably in its output. The economic benefits of that contribution may accrue to companies whose products compete with the publishing industry. Copyright law must evaluate those circumstances through established principles, even as courts confront technological practices with few direct historical equivalents.

The original Kadrey decision has established significant boundaries for that inquiry, particularly concerning the evidence necessary to demonstrate market harm. Its limits remain equally consequential. Meta prevailed on the training-related claims advanced by thirteen authors, while allegations involving distribution and contributory infringement continue. Hundreds of additional writers have pursued separate lawsuits, and the Anthropic settlement has demonstrated that disputes over unauthorized book collections can carry substantial financial consequences even where training itself receives fair-use protection.

The books at the center of these proceedings were created within an industry whose economic arrangements depend upon control of reproduction, licensing and distribution. Their incorporation into artificial intelligence systems has introduced commercial uses that many authors could scarcely have anticipated when they signed their publishing agreements. The resulting disputes concern the interpretation of established rights under circumstances transformed by technological development.

The remaining Kadrey proceedings will determine whether Meta's acquisition practices violated the rights of the plaintiffs whose claims remain before the court. The broader litigation will continue to test the relationship between fair use, market injury and the economic value of human-authored material. Those decisions will influence the conditions under which books are acquired for technological development and the extent to which their creators can participate in the financial benefits derived from them.

The dispute began with a familiar asset in an unfamiliar market. Meta wanted books because their language, structure, and accumulated intellectual content could improve its technology. The authors wanted recognition of the legal and economic rights attached to that work. The courts are now determining how those interests can be reconciled within a copyright system developed long before machines could generate entire libraries of new text. The resulting decisions will help establish how much control writers retain over the commercial uses of their work, and what protection that control affords as literary markets continue to change.

References

1. Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025). Order on cross-motions for partial summary judgment, June 25, 2025. https://law.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv03417/415175/598/

2. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025). Order denying plaintiffs' motion and granting Meta's motion for partial summary judgment, ECF No. 598. https://law.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv03417/415175/598/

3. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. March 25, 2026). Order granting leave to file fourth amended complaint and denying class discovery, ECF No. 700. https://law.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv03417/415175/700/

4. Poritz, I. (2026, October 20). Meta sued by hundreds of authors in AI training copyright cases. Bloomberg Law. https://news.bloomberglaw.com/class-action/meta-sued-by-hundreds-of-authors-in-ai-training-copyright-cases

5. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 27, 2025). Order granting Meta's motion for partial summary judgment on the Digital Millennium Copyright Act claim, ECF No. 601. https://docs.justia.com/cases/federal/district-courts/california/candce/3%3A2023cv03417/415175/601

6. Nawotka, E. (2026, October 21). Hundreds sue Meta for copyright infringement, rejecting class action payouts. Publishers Weekly. https://www2.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/101578-hundreds-sue-meta-for-copyright-infringement-rejecting-class-action-payouts.html

7. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. October 19, 2026). Discovery order addressing related proceedings. U.S. Government Publishing Office. https://www.govinfo.gov/app/details/USCOURTS-cand-3_23-cv-03417/USCOURTS-cand-3_23-cv-03417-54

8. Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal. June 23, 2025). Order on partial summary judgment concerning AI training and copyrighted books.

9. Bartz v. Anthropic PBC, No. 3:24-cv-05417-AMO (N.D. Cal. July 20, 2026). Order granting final approval of class-action settlement, ECF No. 680. https://law.justia.com/cases/federal/district-courts/california/candce/4%3A2024cv05417/434709/680/