Generative Artificial Intelligence

ChatGPT was the 15th most visited website worldwide in 2024, beating out household names like Netflix and Yahoo.¹ The artificial intelligence (AI) market is projected to reach a whopping $1.339 trillion by 2030, with almost 75% of businesses already using AI to help perform a business function.² The functionality and adaptability of AI to business processes is evident; however, according to one survey, over 75% of consumers are simultaneously concerned that AI could cause job losses in the imminent future.³ While AI touches virtually every industry, its impact is arguably most profound in the creative economy, as evidenced by the nearly four dozen copyright lawsuits currently filed against companies with AI platforms.⁴

Generative AI models, like ChatGPT, can be described as software products that are capable of generating text, images, videos or sound based on materials on which they have previously been “trained.”⁵ Due to the number of generative AI options available to consumers, the field is becoming increasingly competitive, so companies are racing to create the most intelligent platform. In order to separate from the pack, generative AI models need to train not just on quantitatively more information, but more importantly, on qualitatively better information. As such, companies have consistently used copyright-protected materials to train their models without getting permission from the copyright holders or compensating them for the right to use their works for this purpose.⁶

Copyright & Fair Use

Copyright is a form of intellectual property that protects original works of authorship that are fixed in a tangible medium of expression.⁷ The goal of copyright is to incentivize others to create more works of authorship, like books, films, songs and plays, by providing the creator the right to control the work for a limited duration, thereby producing more building blocks for creativity. The Copyright Act bestows six exclusive rights to copyright owners, including, of note here, the right to prevent others from reproducing the work.⁸ The ability to monopolize a work one creates may seem to go against the goal of creating more building blocks of creativity, so legal principles fill the gap to help balance these competing interests. The most important of these is copyright fair use. Under this doctrine, “the fair use of a copyrighted work … for purposes such as criticism, comment, news reporting, teaching … scholarship or research, is not an infringement of copyright.”⁹ Importantly, the foregoing examples are illustrative of fair use, not dispositive of fair use. In order to determine whether a specific use is a fair use, the Copyright Act gives four factors to be considered: (i) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (ii) the nature of the copyrighted work; (iii) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (iv) the effect of the use upon the potential market for or value of the copyrighted work.¹⁰ Considering that companies, like Meta, acknowledge using copyrighted works to train their AI platforms without permission from the copyright owner (i.e., implicating the copyright owner’s right to control reproduction), the heart of the analysis revolves around the defense of fair use. Three recent decisions illustrate how courts are analyzing the legitimacy of using copyrighted material without permission to train AI platforms.

3 Recent Court Decisions Paving the Way

Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. Ross Intelligence Inc.

In Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. Ross Intelligence Inc., the U.S. District Court for the District of Delaware handed down a groundbreaking decision earlier this year involving Ross Intelligence Inc. (Ross), a new direct competitor of Thomson Reuters’s Westlaw platform.¹¹ The facts of the case involved Ross creating its own legal-research search engine and initially attempting to get Thomson Reuters to license its Westlaw content to Ross so that Ross could train its new AI search tool. When Thomson Reuters refused, Ross’s solution was to use another platform, LegalEase, to get training for the new Ross AI search tool by using what the case refers to as “Bulk Memos.” These “Bulk Memos” were lawyers’ compilations of legal questions, with some compilations being copied and pasted directly from Westlaw headnotes. When Thomson Reuters learned that Ross built its competing product using such Bulk Memos, it sued Ross for copyright infringement.

The court first looked at the issue of whether the LegalEase Bulk Memo questions copied Thomson Reuters’s headnotes or were instead taken from uncopyrightable judicial opinions. To prove that its Westlaw headnotes were directly infringed, Thomson Reuters had to show that (1) it owned a valid copyright and (2) Ross copied protectable elements of the copyrighted work.¹²

As to (1), the court noted that originality is central to copyright but that the original threshold is “extremely low,” requiring only “some minimal degree of creativity . … some creative spark.”¹³ With the key question being whether a work is original, and not how much effort went into developing it, the court found that Westlaw headnotes (and also Westlaw’s Key Number System) are original and thus copyrightable as they “introduce creativity by distilling, synthesizing or explaining part of an opinion …” ¹⁴

As to (2), the court had to decide whether Thomson Reuters had proven both (a) actual copying and (b) substantial similarity. The parties agreed that LegalEase had access to Westlaw and used it to make Bulk Memos. After tediously reviewing thousands of headnotes, the court found that a Bulk Memo question looks more like a headnote than an underlying judicial opinion, which is strong circumstantial evidence of actual copying. Regarding substantial similarity, the court looked at whether an ordinary user of a product would find it substantially similar to the copyrighted work.¹⁵ The court’s analysis focused on this rule: “The less protectable expression a work contains, the more similar the allegedly infringing work must be to it.” As the court manually compared thousands of headnotes and Bulk Memos, the court granted summary judgment for Thomson Reuters only on the thousands of headnotes whose language very closely tracked the language of the Bulk Memo questions but not the language of the underlying case opinion.

After extinguishing Ross’s various ancillary defenses to copyright infringement, the court focused on the affirmative defense of fair use, with the first and fourth factors weighing the heaviest on the Court’s analysis.¹⁶

Factor 1:
The Purpose and Character of the Use

First, the court looked at whether the purpose and character of Ross’s use was commercial and whether it was transformative.¹⁷ Ross admitted that its use was commercial, and the court found that Ross’s use was not transformative because it did not have a “further purpose or different character” from Thomson Reuters’s.¹⁸ Rather, “Ross took the headnotes to make it easier to develop a competing legal research tool.”¹⁹

Factor 2:
The Nature of the Copyrighted Work

Second, the court analyzed the nature of the original work and noted that “[m]ore creative works get more protection.”²⁰ The court found that while Westlaw’s material had more than the minimal spark of originality required for copyright validity, the material was not that creative and thus this second factor was decided in favor of Ross.

Factor 3:
The Amount and Substantiality of the Portion Used in Relation to the Copyrighted Work as a Whole

Third, the court focused on how much of the work was used and how substantial a part it was to the relative whole. The question was whether that usage was “reasonable in relation to the purpose of the copying.”²¹ The court noted that “[w]hat matters is not ‘the amount and substantiality of the portion used in making a copy, but rather the amount and substantiality of what is thereby made accessible to a public for which it may serve as a competing substitute.²² The parties agreed that Ross’s output to an end user did not include a West headnote. Accordingly, because Ross did not make West headnotes available to the public, this third factor also went in favor of Ross.

Factor 4:
The Effect of the Use Upon the Potential Market for or Value of the Copyrighted Work

Finally, on the most important element of fair use, the court considered the “likely effect [of Ross’s copying] on the market for the original.”²³ Ross’s intention was to compete with Westlaw by developing a market substitute.²⁴ Finding in favor of Thomson Reuters, the court noted that “[t]here is nothing that Thomson Reuters created that Ross could not have created for itself or hired LegalEase to create for it without infringing Thomson Reuters’s copyrights.”²⁵ Accordingly, balancing the four factors in terms of importance and which factors favor which party, the court granted summary judgment for Thomson Reuters and ruled against fair use.

Andrea Bartz, Charles Graeber and Kirk Wallace Johnson v. Anthropic PBC

Anthropic PBC (Anthropic) is an AI software company whose offerings include Claude, an AI software service.²⁶ Since its release in March 2023, Claude has generated over 1 billion dollars in annual revenue.²⁷

In creating Claude, Anthropic aimed to develop a trained large language model or LLM from which a user could seek an answer that would “[mimic] human reading and writing”²⁸ — in other words, it could mimic the writing styles of others.²⁹ To achieve this, Anthropic prioritized training on the “well-curated facts, well-organized analyses and captivating fictional narratives,”³⁰ and in sum, the kinds of “creative expressions”³¹ found in books written by professional authors.³²

In order to train Claude, Anthropic used millions of copyrighted books. It obtained books via pirating and purchasing. As U.S. District Judge William Alsup noted, Anthropic “could have purchased books, but it preferred to steal them.”³³ All in all, Anthropic pirated over seven million copies of books, which it maintained and utilized.³⁴ It also purchased millions of print books, often in used condition.³⁵ For the physical books, Anthropic would remove the binding, cut the pages to a certain size, scan each page into a digital format and then discard the originals.³⁶

The plaintiffs, Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, are authors of books that Anthropic purchased and pirated to incorporate into their central library.³⁷ The plaintiffs/authors allege that Anthropic “infringed its federal copyrights by pirating copies for its library and by reproducing them to train its LLMs.”³⁸ They do not allege that Claude’s outputs infringe their works, as Claude does not create an exact copy, nor a “substantial knock-off.”³⁹ Claude’s outputs were “cleaned, tokenized and compressed.”⁴⁰

Anthropic filed a motion for summary judgment to determine if the uses of the works qualify as “fair uses.”⁴¹ Judge Alsup issued an order on June 23.⁴²

The court considered fair use as it relates to three specific acts of Anthropic: (1) the use of the authors’ works to train Claude,⁴³ (2) the use of purchased books to build a central library from which Claude would draw⁴⁴ and (3) the use of pirated books to build a central library from which Claude would draw.⁴⁵

As to the use of the authors’ works to train Claude, the court held that it was fair, with fair-use factors one, three and four weighing in favor of Anthropic.⁴⁶ Anthropic used copies of copyrighted works to train Claude so it could generate new responses to users’ prompts.⁴⁷ Although this process involved compressing or memorizing large amounts of material, it was likened to how a human might study writing — by reading and internalizing examples in order to produce new, original works.⁴⁸ As to factor one, the purpose and character of the use, Judge Alsup characterized this use as not only spectacularly transformative,⁴⁹ but also “quintessentially transformative.”⁵⁰ In further support of factor one, he shared that “the technology at issue was among the most transformative many of us will see in our lifetimes.”⁵¹ He found, as to factor three, that all the copying was reasonably necessary for the transformative use.⁵² As to factor four, the effect of the use upon the market for or value of the copyrighted work, Judge Alsup determined that training Claude and the information output by Claude does not “displace demand” for plaintiffs’ works in that it does not result in any copies or “infringing knockoffs” being provided to the public.⁵³ The judge again compared these processes to “training schoolchildren to write well”,⁵⁴ and thereupon, having the possibility of an “explosion of competing works,”⁵⁵ “this is not the kind of competitive or creative displacement that concerns the Copyright Act.”⁵⁶

The use of purchased books to build a central library was also found to be fair,⁵⁷ as factors one and three strongly favored fair use, with factor four being neutral and factor two slightly disfavoring it. The court emphasized that the “purpose and character” of using the purchased works, whether in the original print version or the digital version, was transformative.⁵⁸ And, because the purpose required full access to the texts, copying the entire work was justified.⁵⁹

On the other hand, the use of pirated books to build a central library was not justified as fair use.⁶⁰ The court concluded that Anthropic had no legal right to any pirated copies⁶¹ and that the “purpose and character” of using the pirated works was not transformative.⁶² This issue was set for trial on Dec. 1. In preparation for trial,⁶³ the parties engaged in “robust and extensive” and “hard-fought” discovery.⁶⁴

As it relates to the pirated works, on Sept. 5, an Unopposed Motion for Preliminary Approval of Class Settlement was filed by Plaintiffs.⁶⁵ Under the proposed settlement, Anthropic would (1) pay the certified class at least $1.5 billion, plus interest, for the “largescale copyright infringement of books from allegedly pirated datasets;” (2) destroy the pirated material; and (3) in return, receive a past release for conduct up to Aug. 25.⁶⁶ In the memo, the agreement is described as a “landmark settlement” that “will be the largest publicly reported copyright recovery in history, larger than any other copyright class action settlement or any individual copyright case litigated to final judgment.”⁶⁷ Settlement negotiations included consultation with leading membership and trade associations representing rightsholders.⁶⁸ By order dated Sept. 7, Judge Alsup scheduled a hearing on the motion but expressed that he was “disappointed that counsel have left important questions to be answered.”⁶⁹ However, during the Sept. 25 hearing on the motion, the judge called the settlement “fair” and preliminarily approved it, while still noting that there may be challenges ahead in administering the settlement.⁷⁰

Richard Kadrey, et al., v. Meta Platforms Inc.

Released just two days after the Anthropic decision, Richard Kadrey, et al., v. Meta Platforms Inc. was decided out of the same federal district court as the Anthropic case summarized above — the Northern District of California. Furthermore, in Kadrey v. Meta, Judge Vince Chhabria takes a significantly stronger stance against fair use at the outset of the opinion, saying, “This case presents the question whether [using copyrighted material to train AI models without permission] is illegal. Although the devil is in the details, in most cases the answer will likely be yes.”⁷¹

Here, 13 authors sued Meta (Facebook’s parent company) for downloading their books from online “shadow libraries” and using the books to train Meta’s generative AI model called Llama.⁷² A shadow library is an online repository that provides things like books and academic journal articles for free download, regardless of whether that media is copyrighted.⁷³ Meta initially attempted to license books from several publishers to train Llama, but quickly realized licensing was not practical since most individual authors (and not the publishers) hold the subsidiary rights to license the books for AI training.⁷⁴ Plaintiffs alleged copyright infringement based on Meta’s reproduction of their works without permission, to which Meta did not dispute, but rather, Meta claimed such use was fair use under copyright.

Factor 1:
The Purpose and Character of the Use

In assessing the transformative value of a secondary work, one must ask whether the secondary work encourages the development of new expression without diminishing the incentive to create, or rather provides the public with a substantial substitute for the original.⁷⁵ Here, Judge Chhabria said the use of the plaintiffs’ books to train Llama was highly transformative. To support this, he pointed to evidence where, even using “adversarial” prompts designed to get Llama to regurgitate its training data, Llama would not produce more than 50 words of any of the plaintiffs’ books.⁷⁶ Citing prior precedent where the Supreme Court found that Google’s use of Oracle’s computer code that could be readily used by programmers to create the Android platform was transformative, he went on to say that Llama can be used to generate diverse text and perform a wide range of functions, including to edit an email or write a skit based on a hypothetical scenario.⁷⁷ Of course, commerciality is also to be weighed, and considering Meta admitted that it expects to generate $460 billion to $1.4 trillion in revenue over the next 10 years, this cuts against fair use.⁷⁸ However, where a use is highly transformative, commerciality is less significant, so Meta won on this factor.⁷⁹

Factor 2:
The Nature of the Copyrighted Work

Considering copyright law prioritizes protecting highly expressive works, works receiving greater copyright protection include creative ones like books and movies while works receiving lesser protection include less creative ones like computer code.⁸⁰ Here, plaintiffs’ books are highly creative, tilting this factor in favor of plaintiffs; however, this factor is commonly acknowledged as the least significant factor in the fair-use analysis.

Factor 3:
The Amount and Substantiality of the Portion Used in Relation to the Copyrighted Work as a Whole

This factor focuses not on the amount of copyrighted material used by the copier, but the amount of copyrighted material made available to the public.⁸¹ Here, the plaintiffs’ books were not regurgitated in any substantial way and the court concluded that it was reasonable to use the entire book as opposed to a portion of the book when considering the highly transformative nature of Llama, as using an entire book produced a better product than using a portion of a book.⁸²

Factor 4:
The Effect of the Use Upon the Potential Market for or Value of the Copyrighted Work

Again, as noted above, the fourth factor is the single most important factor in analyzing fair use.⁸³ The plaintiffs made two arguments. First, that Llama can regurgitate text from their books, thereby harming the market for their works since consumers could just read the output rather than purchase the book.⁸⁴ However, as mentioned above, even when promoted with ‘adversarial’ prompts, Llama did not produce more than 50 words from any of the plaintiffs’ books.⁸⁵ Second, the plaintiffs argued that allowing Meta to train Llama with their books without paying a fee harmed the potential market for licensing their works for this purpose.⁸⁶ Unfortunately for the plaintiffs, a loss (or potential loss) of licensing monies is always going to be a factor for the rightsholders, so where the use is highly transformative and the licensing market is anything but established, plaintiffs fail on this point.⁸⁷

Interestingly, plaintiffs did not argue a market harm from the output potential of Llama. As Judge Chhabria says, “People could … use LLMs to create books and then sell them, competing with books written by human authors for sales and attention. People might even be motivated to make those books available for free, given how easily it will presumably be to prompt an LLM to create them.”⁸⁸ Llama and other AI models depend on the creativity of others to operate, and armed with this information, can create millions of books, images, songs and videos in a fraction of the time it takes for a person to do so.⁸⁹ The reason this is so devastating to copyright is it arguably will disincentivize people from creating expressive works, which is the exact harm copyright aims to prevent.⁹⁰ Judge Chhabria says, “Indeed, it seems likely that market dilution will often cause plaintiffs to decisively win the fourth factor — and thus win the fair use question overall — in cases like this.”⁹¹ Unfortunately for plaintiffs, they did not argue nor provide any evidence of market harm, leaving Meta to win the fourth factor and entitling Meta to prevail in summary judgment on its fair use defense.

Conclusion

Together, the Thomson Reuters, Anthropic and Meta cases mark a critical turning point in the legal landscape governing the intersection of AI and copyright. While each case presents a unique set of facts, they collectively reflect courts’ growing efforts to define the boundaries of fair use and generative AI’s reliance on copyrighted material. Courts are beginning to diverge in their approaches, with Thomson Reuters narrowly focused on data scraping for legal research while Anthropic and Meta highlight deeper concerns over authors’ rights and the commercial replication of expressive content. Certain companies have struck direct licensing agreements to train with copyrightable content, including OpenAI with the Associated Press, and Shutterstock and Getty Image’s collaborations with Nvidia and Bria; however, corporations that control large amounts of high-quality content are in a much better position to license content.⁹² The bigger challenge will be whether a licensing scheme can be created to represent the interests of individual creators, similar to the role served by performing rights organizations for songwriters in the music space.

Tennessee has emerged as a national leader in protecting creators’ rights in the age of AI. Nashville-based companies, like Monarrch and Humanable, are building tools for real-time rights tracking, watermarking and human-authorship certification, helping creators in the music space maintain control over their works. Another company, Sureel, which operates out of California, shares in the broader mission of enforcing consent-based use of creative content in AI training. Complementing these efforts, organizations like the Nashville Songwriters Association International (NSAI) are advocating for new federal legislation that reflects the needs of creators, specifically songwriters and artists. Together, all of these legal, technological and policy developments underscore Tennessee’s growing role in protecting creative rights in the current AI landscape, and highlight the urgent need for comprehensive frameworks and tools that balance innovation with meaningful protection for human creators. |||

Nathan Drake is an attorney and assistant professor of media and entertainment industries in the Mike Curb College of Entertainment and Music Business at Belmont University in Nashville. Prior to teaching at Belmont, he practiced entertainment law at Loeb & Loeb and Dickinson Wright.

Amy Bryson Smith is an attorney and chair/associate professor of music business in the Mike Curb College of Entertainment and Music Business at Belmont University. Prior to teaching, she practiced law for many years. Smith’s research and practice focus on the intersection of law and creative industries.

Mary Lauren Teague is an attorney and assistant professor of music business in the Mike Curb College of Entertainment and Music Business at Belmont. Prior to teaching at Belmont, she practiced entertainment law at Loeb & Loeb and served as an artist and songwriter manager.

NOTES
1. www.statista.com/statistics/1201889/most-visited-websites-worldwide-unique-visits/
2. www.forbes.com/advisor/business/ai-statistics/
3. Id.
4.  chatgptiseatingtheworld.com/2024/08/27/master-list-of-lawsuits-v-ai-chatgpt-openai-microsoft-meta-midjourney-other-ai-cos/
5. Kadrey v. Meta Platforms, Inc., 2025 U.S. Dist. LEXIS 121064, 2025 LX 111886, F.Supp.3d, 2025 WL 1752484 (United States District Court for the Northern District of California June 25, 2025, Filed)
6. Id. at 12.
7. 17 U.S.C. §102.
8. 17 U.S.C. §106.
9. 17 U.S.C. §107.
10. Id.
11. Thomson Reuters Enterprise Centre GmbH et al. v. Ross Intelligence Inc., No. 1:20-cv-613-SB (D. Del. 2025).
12. Feist Publications Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 361 (1991).
13. 694 F. Supp. 3d at 345.
14. Thomson Reuters, No. 1:20-cv-613-SB (D. Del. 2025).
15. Dam Things, 290 F.3d at 562.
16. Authors Guild v. Google Inc., 804 F.3d 202, 220 (2d Cir. 2015) (Leval, J.).
17. Andy Warhol Found. for the Visual Arts Inc. v. Goldsmith, 598 U.S. 508, 529-31 (2023).
18. Id. at 529.
19. Thomson Reuters, No. 1:20-cv-613-SB (D. Del. 2025).|20. 4 Nimmer on Copyright § 13F.06[A].
21. Campbell v. Acuff-Rose Music Inc., 510 U.S. 569, 586 (1994).
22. Authors Guild, 804 F.3d at 222.
23. Campbell, 510 U.S. at 590.
24. D.I. 752-1 at 4.
25. Thomson Reuters, No. 1:20-cv-613-SB (D. Del. 2025).
26. Andrea Bartz, Charles Graeber and Kirk Wallace Johnson v. Anthropic PBC, 2025 WL 1741691 (N.D. Cal. June 23, 2025).
27. Id. at 1.
28. Id.
29. Id.
30. Id. at 3.
31. Id.
32. Id.
33. Id. at 2.
34. Id.
35. Id.
36. Id.
37. Id. at 1.
38. Id. at 5.
39. Id. at 4.
40. Id.
41. Id. at 5.
42. Id. at 1.
43. Id. at 7.
44. Id.
45. Id.
46. Id. at 18.
47. Id. at 1.
48. Id. at 7.
49. Id.
50. Id. at 8.
51. Id. at 18.
52. Id. at 15.
53. Id. at 16-17.
54. Id. at 17.
55. Id.
56. Id.
57. Id at 17-18.
58. Id. at 11.
59. Id. at 16.
60. Id. at 18.
61. Id. at 16.
62. Id. at 14.
63. Id. at 19.
64. Memorandum of Points and Authorities in Support of Unopposed Motion for Preliminary Approval of Class Settlement, Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA, at 9-10 (N.D. Cal. Sept. 5, 2025).
65. Unopposed Motion for Preliminary Approval of Class Settlement, Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. Sept. 5, 2025).
66. Memorandum in Support of Unopposed Motion for Preliminary Approval, at 1.
67. Id. at 1.
68. Id. at 2.
69. Order RE Hearing on Motion for Preliminary Approval of Settlement, Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. Sept. 7, 2025).
70. Bonnie Eslinger, Anthropic Judge Greenlights ‘Historic’ $1.5B Copyright Deal, LAW 360 (Sept. 25, 2025), www.law360.com.
71. Kadrey v. Meta Platforms Inc., F.Supp.3d at 12.
72. Id. at 17.
73. Id. at 29.
74. Id.
75. Id. at 37.
76. Id. at 41.
77. Id. at 38.
78. Id. at 41.
79. Id.
80. Id. at 48.
81. Id. at 52.
82. Id.
83. Id. at 53.
84. Id. at 55.
85. Id.
86. Id.
87. Id.
88. Id. at 57.
89. Id. at 64.
90. Id. at 61.
91. Id. at 64.
92. www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf.