Training on Thin Ice: Thomson Reuters v. ROSS and the Future of Fair Use for AI Systems

October 6, 2025

By Atreya Mathur

In 2020, Thomson Reuters, owner of the widely used legal research platform Westlaw, filed suit against ROSS Intelligence Inc., a start-up developing an AI-driven legal research tool. The dispute centers on Westlaw’s “Headnotes,”editorial summaries and points of law that precede judicial opinions on the platform.^[1] In their complaint, Thomson Reuters alleged that ROSS copied thousands of these headnotes to train and populate its competing system, rather than independently creating its own summaries, and did so without authorization.^[2] The complaint alleged copyright infringement and tortious interference, and included claims that ROSS misappropriated proprietary data by using Westlaw headnotes without authorization.^[3] ROSS denied liability, arguing that its use of the headnotes was protected under fair use and necessary for innovation in legal technology.^[4]

Importantly, this case involves a non-generative AI system. According to Thomson Reuters and the district court’s findings, ROSS obtained and used copyrighted Westlaw headnotes in developing its legal-research tool.^[5] Unlike generative AI platforms, which are designed to produce new content such as text, images, or music from learned patterns, ROSS’s system was built to organize and deliver legal information for research purposes rather than to generate new expressive works.^[6] This distinction matters because the court’s analysis focused on whether ROSS’s use functioned as a substitute for Westlaw’s editorial content and harmed its market, rather than on whether AI training alone constitutes transformation or creative output.^[7]

In February 2025, the United States District Court for the District of Delaware issued a significant ruling on cross-motions for renewed summary judgment. The court held that ROSS had infringed the copyright in thousands of Westlaw headnotes and that its fair use defense failed as a matter of law for a large subset of the works.^[8] The court emphasized the commercial nature of ROSS’s conduct, the lack of meaningful transformation, and the realistic risk of market substitution.^[9] On March 18, 2025, ROSS Intelligence filed its Opening Brief in support of its Motion for Interlocutory Appeal and Stay Pending Appeal. ROSS contends that the district court erred in its summary judgment ruling, arguing that the Westlaw headnotes are not copyrightable and that its use of them constitutes fair use.^[10] On April 1, 2025, Thomson Reuters filed its Opposition to ROSS’s Motion for Interlocutory Appeal and Stay Pending Appeal. Thomson Reuters argues that the district court’s decision was correct and that ROSS’s use of the headnotes infringes upon its copyrights.^[11] On May 23, 2025, the district court granted ROSS’s motion for interlocutory appeal and stay pending appeal, certifying two questions for the Third Circuit: (1) whether the Westlaw headnotes and Key Number System are original, and (2) whether ROSS’s use of the headnotes was fair use. The court noted that these issues present substantial grounds for a difference of opinion and have the potential to materially advance the litigation.^[12] On September 22, 2025, ROSS filed its Opening Brief in the Third Circuit, challenging the district court’s rulings on copyrightability and fair use. ROSS argues that the headnotes are not copyrightable and that its use of them was transformative and non-infringing.^[13]

Although not yet binding precedent, this decision is among the first federal rulings to directly address fair use in the context of training and operating AI systems. The Delaware court’s reasoning is already shaping industry commentary and influencing briefs in other AI copyright disputes, signaling how courts may scrutinize the use of protected content in building AI tools, a question with immediate implications for both legal technology and creative sectors, including the art world.

Factual background

Westlaw, owned by Thomson Reuters, is one of the most established legal research platforms in the United States. Its value extends beyond providing the full text of judicial opinions (which are in the public domain) to the proprietary editorial additions surrounding those opinions.^[14] Central among these are headnotes, short numbered summaries that distill specific points of law and link cases to Westlaw’s Key Number System.^[15] These headnotes have been recognized as copyrightable editorial works and form a critical component of Westlaw’s platform and competitive advantage.^[16]

ROSS Intelligence, a start-up founded in 2015, sought to transform legal research with an AI assistant capable of answering lawyers’ questions in plain English.^[17] While ROSS initially relied on public domain case law, Thomson Reuters contends that the company took a shortcut: instead of creating its own editorial content, ROSS allegedly relied on thousands of Westlaw headnotes, obtained indirectly through a contractor called LegalEase, to train and populate its system.^[18] Importantly, ROSS did not directly ingest Westlaw’s headnotes or the Key Number System. ROSS engaged LegalEase to create “Bulk Memos” (collections of legal questions and potential answers compiled by lawyers) that it used as training data.^[19] LegalEase gave those lawyers a guide explaining how to create those questions using Westlaw headnotes, while clarifying that the lawyers should not just copy and paste headnotes directly into the questions.^[20] Thomson Reuters maintains that, regardless of these instructions, the process amounted to large scale use of its protected works.^[21]

ROSS also attempted to license the content directly from Thomson Reuters, but negotiations failed because Thomson Reuters viewed the proposed AI tool as a competitor to Westlaw.^[22] ROSS denies direct copying, emphasizing that its system paraphrased legal principles rather than reproducing headnotes verbatim. It has framed the litigation as an effort by Thomson Reuters to monopolize access to legal information and block AI innovation in the sector.^[23]

In 2020, Thomson Reuters filed suit in the U.S. District Court for the District of Delaware, asserting claims for copyright infringement, tortious interference, and misappropriation. Although ROSS shut down its platform under litigation pressure, it continued to defend the case, making fair use the centerpiece of its argument.

After extensive discovery and an initial round of summary judgment, the court issued a renewed summary judgment ruling in February 2025. The court concluded that ROSS had reproduced 2,243 Westlaw headnotes and held that both the headnotes and the Key Number System satisfied the relatively minimal threshold of originality required for copyright protection.^[24] The court rejected ROSS’s merger and scènes à faire defenses, finding that neither the headnotes nor the Key Number System were dictated by necessity.^[25] It also rejected ROSS’s fair use defense, determining that the first and fourth statutory factors, the purpose and character of the use and the effect on the market, favored Thomson Reuters.^[26]

This ruling, now certified for interlocutory appeal to the Third Circuit, establishes a fact-specific framework for evaluating AI training and copyright. Although it arises from a non-generative AI context, where the AI replicates and organizes existing content rather than producing new creative works, it raises questions with broad implications for AI research tools and other applications that use copyrighted material as training input.

Core legal issues decided by the district court

The district court’s memorandum opinion considered (a) whether the headnotes and related Westlaw editorial material are copyrightable, (b) whether ROSS actually copied (factually and substantially) those materials in the Bulk Memos, and (c) whether ROSS’s uses were fair use under the four statutory factors (purpose/character of use, nature of the copyrighted work, amount/ substantiality, and market effect).^[27] Some of the important conclusions of the court are as follows:

Copyrightability: The court reaffirmed that while the underlying judicial opinions are public domain, the short editorial summaries, organizational structure, and classification systems created by Westlaw’s editors reflect sufficient creativity to be treated as protected works.^[28] This distinction is significant because it draws a boundary between what is public domain (which anyone can use) and proprietary editorial content (which remains protected).

Copying and Substantial Similarity: The court found that ROSS’s Bulk Memos were not simply inspired by public domain cases but reproduced material substantially similar to Westlaw’s headnotes.^[29] The evidence showed both access and use: LegalEase lawyers drafted questions and answers referencing headnotes, and the resulting datasets were used to train ROSS’s system.^[30] This finding matters because AI developers frequently argue they only use “factual” inputs or paraphrased material; the court here signaled that paraphrase-based training can still amount to copying when it tracks protected expression closely enough.^[31]

Fair Use Analysis: The court conducted a detailed four-factor fair use analysis and rejected ROSS Intelligence’s defense for a substantial subset of the challenged Westlaw headnotes.

Factor 1: Purpose and Character of the Use: The court found that ROSS’s use was largely commercial, as it was intended to provide a competing legal research tool rather than for scholarly, educational, or transformative purposes.^[32] While ROSS claimed some transformative intent in training AI, the court concluded that the use functionally replaced Westlaw’s services and thus favored Thomson Reuters.^[33]

Factor 2: Nature of the Copyrighted Work: The court emphasized that the headnotes are creative editorial works.^[34] They reflect the editorial judgment of Westlaw’s staff in selecting, summarizing, and organizing key points of law.^[35] Unlike raw facts or purely functional information, these headnotes demonstrate originality, so this factor weighed against ROSS.^[36]

Factor 3: Amount and Substantiality of the Portion Used: ROSS copied a large volume of headnotes wholesale. The court highlighted that even if some copying could be justified for research purposes, the scale and comprehensiveness of the copying went far beyond what would be considered fair use.^[37] This factor therefore also favored Thomson Reuters.

Factor 4: Effect on the Market: The court found this factor particularly significant. ROSS’s use directly competed with Westlaw by threatening subscription revenues and licensing markets for the headnotes.^[38] The court concluded that the copying usurped the market for Thomson Reuters’ editorial work, a central component of Westlaw’s value.^[39]

Considering all four factors together, the court determined that fair use did not apply as a matter of law to a substantial portion of the materials. While acknowledging that some uses of AI or other limited excerpts might present different considerations, the opinion firmly rejected ROSS’s fair-use argument in the context of its commercial, wholesale copying of Westlaw’s headnotes.^[40]

Rejection of Merger and Scènes-à-Faire Defenses: ROSS also invoked two classic copyright-limiting doctrines, merger and scènes-à-faire, to argue that Westlaw’s headnotes were too ‘functional’ to be protected.^[41] The Delaware court rejected both defenses, explaining that there are many ways to summarize points of law and that nothing about judicial opinions requires the specific selection, arrangement, and phrasing used in Westlaw’s headnotes.^[42] Under the merger doctrine, when there is only one or a very limited number of ways to express an idea, the expression and idea are said to “merge” and copyright protection drops away.^[43] Scènes-à-faire excludes from protection stock elements that are dictated by external factors, such as industry standards, common forms, or necessary legal terminology.^[44] ROSS essentially argued that because the headnotes summarized public domain legal rules, there was little room for creative choice and therefore no protectable expression. The Delaware court rejected both defenses. It found that Westlaw’s editors exercised judgment in phrasing, emphasis, and organization when drafting headnotes, and that there are many different ways to summarize a point of law.^[45] In other words, the expression had not “merged” with the underlying idea. Likewise, while legal terms of art are unprotectable, the particular selection and arrangement of those terms, which cases to link, how to number and group them, and what wording to use, reflected enough originality to avoid the scènes-à-faire bar.

This is significant for AI cases because it shows that simply labeling material as “functional” or “factual” will not automatically strip it of protection; courts will scrutinize whether the creator’s editorial choices are genuinely constrained or whether there was room for creativity and variation.

Why the decision matters

Although Thomson Reuters v. ROSS Intelligence emerged from a dispute over legal research tools, the court’s reasoning extends far beyond the legal-tech industry. It represents one of the first federal decisions to grapple with fair use in the context of an AI developer using vast quantities of copyrighted text to train and operate a model. By stressing limited transformation, and tangible market harm, the court sketched out a framework that could guide how similar disputes are resolved across creative fields such as visual art.

First, the nature of the training data matters. The Delaware court treated Westlaw’s headnotes as highly original editorial content whose value depended on exclusive access.^[46] The concern for artists is analogous: even though illustrations, photographs, and other creative works are automatically protected by copyright, AI training presents a unique challenge. When AI platforms ingest large datasets of copyrighted works without permission, they can generate new outputs that may compete with the original works, effectively creating substitutes without licensing fees. For example, an AI trained on a collection of an artist’s portfolio could produce new images in a similar style, potentially reducing demand for commissioned work. Similarly, photographers whose images are scraped for AI training may see stock-photo sales threatened by AI-generated alternatives. While the copyright remains technically in place, enforcing it becomes practically difficult: identifying unauthorized use, proving infringement, and seeking remedies are more complicated when the copying occurs at scale behind the scenes of AI systems. This ruling signals that courts may view systematic ingestion of copyrighted works into AI models as a significant commercial harm, rather than as a neutral or fair use activity.

Second, the court’s treatment of transformative use is especially important. ROSS argued that its system was transformative because it used Westlaw’s headnotes to “teach” an AI model rather than to reproduce them verbatim.^[47] The court rejected that reasoning, holding that training alone does not create a new purpose if the resulting product effectively competes with the original.^[48] This signals to AI art platforms that the mere fact of ingestion or machine-learning does not automatically satisfy the transformation requirement. The decision also aligns with the Copyright Office’s recent report on generative AI, which emphasizes that transformation must be assessed case by case and cannot be assumed simply because an output is machine-generated.^[49] For artists, this narrows one of the central defenses often asserted by AI platforms: that training on copyrighted works is inherently transformative. If courts follow this approach, AI-generated outputs that reproduce stylistic or functional elements of an artist’s work, without adding sufficient new expression, could be seen as infringing, even if no exact copying occurs.

Third, the court’s treatment of market harm is also of consideration. While copyright law already protects artists’ works, courts have sometimes been reluctant to treat potential displacement by AI as actual harm. In Kadrey v. Meta, Judge Chhabria conducted a full four-factor fair use analysis and ruled for Meta on summary judgment, finding that the plaintiffs failed to present admissible evidence showing that Meta’s training of its Llama AI models had demonstrably harmed the market for their works, despite Mark Zuckerberg, founder of Meta stating that he had given the green light for the model to be trained on copyrighted works.^[50] The opinion recognizes the theoretical possibility that widespread AI outputs could dilute demand for original creative works, but emphasizes that such market-dilution theories require concrete evidentiary support, such as sales loss, licensing displacement, or other economic indicators, which the plaintiffs did not produce on this record.^[51] The court therefore concluded that, given the record before it, the plaintiffs had not met their burden to show the sort of cognizable market harm that would defeat a fair use defense.^[52] In ROSS, however, the court credited concrete evidence showing that ROSS’s product could directly undermine Westlaw’s subscription revenue.^[53] Rather than treating market harm as a theoretical possibility, the court recognized it as a real and measurable risk. For artists, this sets an important precedent.^[54] If they can demonstrate that AI-generated images are substituting for commissions, licensing, or stock-photo sales, courts may view the economic impact as tangible, strengthening infringement claims or encouraging settlements. The ruling makes clear that when credible evidence shows that AI outputs threaten an existing market for the original work, courts would likely take that harm seriously rather than dismiss it as speculative.

Finally, the procedural position of the case is also interesting. Because the dispute is now on interlocutory appeal, the Third Circuit will consider these issues relatively soon, potentially producing one of the first appellate level analyses of AI related fair use.

Limitations of the decision

The ROSS ruling has important and specific limits. First, it arises from a non-generative AI context: ROSS was developing a legal-research tool that used copyrighted headnotes to provide a functional alternative to Westlaw, rather than generating new creative content in the way generative AI platforms like ChatGPT, Midjourney, or Stability AI do. This distinction matters because the court’s analysis focused heavily on the substitutive nature of the product– the fact that ROSS’s tool could directly replace Westlaw subscriptions, rather than on broader questions about training AI to produce entirely new works. In other words, the case does not establish any blanket rule that any use of copyrighted works in AI training is automatically infringing.

Second, the decision is highly fact-specific. Westlaw’s headnotes are short, editorially curated summaries that are treated as highly original yet functional content. The court’s fair use analysis, including its evaluation of commerciality, transformation, and market harm, was grounded in the particular features of these headnotes and the competitive market for legal research. Different types of works or AI applications could lead to very different results. For example, courts may weigh the factors differently where the use is perhaps more transformative, where outputs are creative rather than substitutive, or where the use is primarily noncommercial or educational.

The scope of the ruling is also limited to the factual record before the court. The district court did not attempt to define sweeping principles about AI training generally; it applied the four-factor fair use test to the specific evidence of ROSS’s copying and the potential market impact on Westlaw. Consequently, while the ruling is influential, it cannot be read as automatically controlling in cases involving generative AI, large scale image datasets, or other creative outputs.

Practical impact of the case

The Ross decision signals that AI developers face heightened legal risk when they build tools on top of protected editorial or derivative materials. Westlaw’s headnotes were proprietary summaries. Many image and text datasets similarly include curator or author written captions, descriptive metadata, or other tag sets that rights holders consider added value. By recognizing headnotes as protectable expression, Ross strengthens a rightsholders’ leverage to claim that unlicensed ingestion of such editorial content (similar to Headnotes) for training or competing products amounts to infringement.

The opinion also undercuts the assumption that “intermediate” copying (the process of reproducing a copyrighted work as an interim step in the creation of a new copyrighted work) during training is automatically exempt from liability.^[55] The Delaware court treated the dataset creation process and training copies as part of the infringement analysis rather than a neutral pre-processing or interim step.^[56] This is particularly significant for AI art platforms that scrape images or metadata. Courts may examine not only the training process but also whether the resulting system functions as a substitute for the original market, such as licensed prints, stock photos, or derivative services, when evaluating harm and fair use.

Equally important, Ross shows that the “transformative use” defense will be scrutinized for both function and substance. The court emphasized that the mere fact of training an algorithm does not automatically confer a new purpose if the downstream product serves the same role as the protected work.^[57] For image generation, this suggests defense teams will need to make robust showings that the AI outputs add new expression, meaning, or aesthetic value and do not simply reproduce or displace an artist’s market. “In the style of” may not suffice if the end product competes with the artist’s commissions or print sales.

Finally, the ruling increases the commercial incentives for licensing and consent-based datasets. Platforms and model builders may now find it safer to (a) license curated collections and metadata, (b) build opt-in libraries from consenting artists, or (c) focus on public domain and Creative Commons works. Thomson Reuters’ interim win reinforces a licensors’ bargaining power, making it harder for developers to rely on a “fair use” fallback. In addition, when an AI system allows customers to generate near replicas or works “in the style of” a living artist that can substitute for commissions or licensed reproductions, the Ross reasoning makes it easier for courts to treat that displacement as concrete market harm, strengthening artists’ claims if their works were used without permission.

Food for Thought for Artists, Platforms, and Developers

For artists and their representatives, the Ross decision reinforces the importance of proactively documenting how and where their works are used. Keeping records of appearances in datasets or on AI art platforms, sending takedown or cease-and-desist letters when appropriate, and registering works with the Copyright Office can strengthen later infringement claims. Collective licensing or rights-management initiatives may also give artists leverage in negotiating with platforms and model developers.

For platforms, provenance and licensing should become a priority. Maintaining detailed records of the source of each image or text sample, building filters that exclude known proprietary or editorial content, and obtaining explicit releases/licenses from contributors can reduce exposure to claims like those raised in Ross. This is especially important for metadata and captions, which may look “factual” but could qualify as protectable expression.

For developers of generative models, safer practices include: AI training using content in the public domain or under appropriate licenses; if possible, generate or simulate data instead of scraping large amounts of proprietary or copyright protected material; and be transparent with users. Clearly disclose what data is used, and allow creators or users to opt out if feasible. None of these steps guarantees immunity from litigation, but they improve a platform’s defense and may support fair use or other defenses if disputes arise.

What’s next?

Because the Delaware court certified its February 2025 rulings on originality and fair use for interlocutory appeal under 28 U.S.C. § 1292(b), the Third Circuit may be the first at the appellate level to have the opportunity of addressing fair use in the training and operation of an AI system.

One scenario is full affirmance. The Third Circuit could agree with the district court that Westlaw’s headnotes are protected by copyright and that ROSS’s use of them does not qualify as fair use. Such a ruling would turn the Delaware decision from a persuasive opinion into binding precedent within the circuit and an influential reference point nationwide. For disputes over AI generated art, an affirmance would reinforce plaintiffs’ claims that mass ingestion of copyrighted works to build competing services is likely infringing and that “transformative” training by itself does not erase market harm. It could also encourage rightsholders to demand licenses or pursue litigation rather than rely on unsettled fair use defenses.

A second scenario is partial reversal. The Third Circuit could agree that the headnotes are protected but remand on whether ROSS’s use was transformative or whether market harm was adequately shown. Such a decision would give AI developers more room to argue fair use case by case, especially where their models are not direct substitutes for the originals. For artists and creative industries, a partial reversal might preserve some space for experimentation and non-commercial or non-substitutive AI uses, while still warning against wholesale copying.

A third possibility change. It is possible that the appellate court reframes the fair use analysis altogether, offering new guidance on how factors such as transformation and commercial purpose should be weighed in the AI context. For the art world, such a reframing could be significant. If the Third Circuit were to shift the emphasis, for instance, by prioritizing the originality of human input or redefining what counts as “transformative” use, it might reshape how courts assess AI tools that generate images, music, or writing in the style of existing artists. This could leave artists either with stronger grounds to challenge unauthorized uses of their work or, conversely, facing a more permissive legal standard that legitimizes AI-generated imitations.

Thomson Reuters v. ROSS is one of the first and probably most consequential American AI court rulings to apply traditional copyright doctrines to material used in building and operating AI systems. Given that the district court found many Westlaw headnotes to be protectable and rejected fair use for ROSS on key facts, the decision increases legal pressure on AI model builders to obtain licenses, to avoid using editorial datasets without consent, and to carefully document how training data is obtained and used. The decision is particularly relevant for AI art where artists’ work and portfolios are used without permission; however, the case is fact-specific and the Third Circuit’s interlocutory review could alter or clarify the legal standards. Keep updated with the case HERE and HERE.

About the Author

Atreya Mathur is the Director of Legal Research at Center for Art Law.

Select References:

1. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
2. Id. ↑
3. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB, Complaint (D. Del. Nov. 2023), available at https://admin.bakerlaw.com/wp-content/uploads/2023/11/ECF-1-Complaint.pdf ↑
4. Id. ↑
5. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
6. Id. ↑
7. Id. ↑
8. Id. ↑
9. Id. ↑
10. ROSS Intelligence Inc.’s Opening Brief in Support of Its Motion for Certification for Interlocutory Appeal Under 28 U.S.C. § 1292(b) and for Stay Pending Appeal, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB (D. Del. Mar. 18, 2025), ECF No. 787, available at https://admin.bakerlaw.com/wp-content/uploads/2025/05/ECF-787-Rosss-Opening-Brief-ISO-Motion-for-Interlocutory-Appeal-and-Stay-Pending-Appeal.pdf ↑
11. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB, Plaintiffs’ Opposition to Defendant’s Motion for Certification for Interlocutory Appeal Under 28 U.S.C. § 1292(b) and for Stay Pending Appeal (D. Del. Apr. 1, 2025), ECF No. 791, available at https://admin.bakerlaw.com/wp-content/uploads/2025/05/ECF-791-Plaintiffs-Opposition-to-Rosss-Motion-for-Interlocutory-Appeal-and-Stay-Pending-Appeal.pdf ↑
12. Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB, Memorandum Opinion (D. Del. May 23, 2025), ECF No. 804, available at https://law.justia.com/cases/federal/district-courts/delaware/dedce/1:2020cv00613/72109/804/ ↑
13. ROSS Intelligence Files Opening Brief in 1st Appeal of AI Training in Suit by Thomson Reuters, Chat GPT Is Eating the World, available at https://chatgptiseatingtheworld.com/2025/09/22/ross-intelligence-files-opening-brief-in-1st-appeal-of-ai-training-in-suit-by-thomson-reuters/ (Sept. 22, 2025) ↑
14. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
15. Id. ↑
16. Id. ↑
17. See Mannes, John, ROSS Intelligence Lands $8.7M Series A to Speed Up Legal Research with AI, TechCrunch (Oct. 11, 2017), https://techcrunch.com/2017/10/11/ross-intelligence-lands-8-7m-series-a-to-speed-up-legal-research-with-ai/ ↑
18. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
19. Id. ↑
20. Id. ↑
21. Id. ↑
22. Id. ↑
23. Id. ↑
24. Id. ↑
25. Id. ↑
26. Id. ↑
27. Id. ↑
28. Id. ↑
29. Id. ↑
30. Id. ↑
31. Id. ↑
32. Id. ↑
33. Id. ↑
34. Id. ↑
35. Id. ↑
36. Id. ↑
37. Id. ↑
38. Id. ↑
39. Id. ↑
40. Id. ↑
41. Id. ↑
42. Id. ↑
43. See Rich Goldstein, Understanding the Copyright Merger Doctrine, GOLDSTEIN PATENT LAW, available at https://www.goldsteinpatentlaw.com/copyright-merger-doctrine/ and 17 U.S.C. § 102(b) (2022). ↑
44. See Re:Create Coalition, Scenes-a-faire and Why It Matters (Nov. 15, 2024), available at https://www.recreatecoalition.org/scenes-a-faire-and-why-it-matters/ ↑
45. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
46. Id. ↑
47. Id. ↑
48. Id. ↑
49. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (Pre-Publication Version, May 2025), available at https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf ↑
50. See Kadrey et al. v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, Order Denying Plaintiffs’ Motion for Partial Summary Judgment and Granting Meta’s Cross-Motion for Partial Summary Judgment (N.D. Cal. June 25, 2025), ECF No. 598, available at https://law.justia.com/cases/federal/district-courts/california/candce/3:2023cv03417/415175/598/ and Meta says copying books was ‘fair use’ in authors’ AI lawsuit, Reuters (Mar. 25, 2025) available at https://www.reuters.com/legal/litigation/meta-says-copying-books-was-fair-use-authors-ai-lawsuit-2025-03-25/? ↑
51. Id. ↑
52. Id. ↑
53. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-CV-613-SB (D. Del. Feb. 11, 2025), available at https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf ↑
54. The District Court also cited Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 529–31 (2023) throughout its analysis for the fair use four factor test in the case. ↑
55. Id. ↑
56. Id. ↑
57. Id. ↑