The Fair Use Split: How Five Court Rulings, a $1.5 Billion Settlement, and a Sanctions Fight Are Redrawing the Line Between AI Innovation and Creative Rights

The Fair Use Split: How Five Court Rulings, a $1.5 Billion Settlement, and a Sanctions Fight Are Redrawing the Line Between AI Innovation and Creative Rights

From Thomson Reuters to Disney v. Midjourney, the 2026 AI copyright landscape is resolving around a market-competition standard that leaves neither creators nor AI companies fully protected.


The Core Question: Can AI Train on Copyrighted Works?

The central legal question of the AI era is whether training generative AI models on copyrighted works constitutes fair use or infringement. In 2025 and 2026, federal courts began answering that question, and the answers diverged sharply depending on two variables: whether the AI’s output competes with the original work’s market, and whether the training data was lawfully acquired or pirated. The result is not a blanket rule but a fact-specific framework that has produced a three-way split among courts sitting in the same federal districts.[1]

The U.S. Copyright Office weighed in with its third and final report in the “Copyright and Artificial Intelligence” series, released in pre-publication form in May 2025. The 108-page report declined to issue a blanket endorsement or rejection of fair use for AI training, instead concluding that “making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries.”[2] The report emphasized that the first factor (purpose and character of the use, including transformativeness) and the fourth factor (effect on the potential market for the copyrighted work) carry the most weight in AI training cases, but that every determination requires a case-by-case balancing of all four statutory factors.[3]


The Three-Way Judicial Split

Thomson Reuters v. Ross Intelligence: Fair Use Denied

The first major ruling came in Thomson Reuters v. Ross Intelligence, where a federal court found that Ross Intelligence’s use of Westlaw’s editorial headnotes to train an AI-powered legal research tool was not fair use. The court’s reasoning turned on market substitution: Ross’s AI was a direct competitor to Westlaw’s research services, and the headnotes were original, protected editorial expression rather than raw data. The court also noted the potential effect on a developing market for AI training data licenses. This case is currently on appeal before the Third Circuit.[1]

The core logic here is that when an AI product competes directly with the source it was trained on, the fourth fair use factor market-effect analysis cuts hard against the defendant. Ross was not building a general-purpose language model; it was building a legal research tool that replaced the very product whose editorial content it ingested.

Bartz v. Anthropic: Training Is Fair Use, Piracy Is Not

In June 2025, Judge William Alsup of the Northern District of California ruled in Bartz et al. v. Anthropic that training AI on copyrighted books constitutes fair use, calling the use “transformative, spectacularly so” and “among the most transformative many of us will see in our lifetimes.”[1] The court found that copying entire works was “reasonably necessary” to train a large language model, and that because the training data is never shown to users and does not replace the market for the original books, the market-effect factor favored Anthropic. The court’s view was that the fourth fair use factor protects against replacement, not competition.

However, the court drew a sharp line on provenance: it found that Anthropic’s acquisition and storage of over seven million pirated books was not protected by fair use, expressing skepticism that “any subsequent fair use” could ever justify pirated downloads. The case later settled for $1.5 billion, with an estimated payout of approximately $3,000 per work. The settlement terms include the destruction of Anthropic’s pirated dataset and are limited to past use of training data, not outputs.[1]

Kadrey v. Meta: Fair Use Regardless of Source

Days after the Bartz ruling, another judge in the same federal district reached a different conclusion in Kadrey et al. v. Meta Platforms Inc. The court granted a partial motion to dismiss on fair use grounds for training Meta’s Llama model on the plaintiffs’ books, finding the training transformative. But unlike the Bartz court, the Kadrey court squarely held that LLM training constitutes fair use regardless of whether the underlying materials were obtained from legitimate sources or through piracy.[1]

The Kadrey court also introduced a novel “market dilution” theory, arguing that “no other use has anything near the potential to flood the market with competing works the way that LLM training does.”[4] However, the court ultimately granted summary judgment for Meta because the plaintiffs “presented no meaningful evidence on market dilution at all.” Had they been able to demonstrate market dilution, the fourth factor might have gone to a jury. This ruling signals that while training may survive fair use scrutiny, the door remains open for plaintiffs who can prove their works are being indirectly substituted by AI-generated competing content.


Disney v. Midjourney: The Image Generation Front

While the text-training cases grapple with abstract legal doctrine, the visual media cases present juries with something far more visceral: images. In June 2025, Disney, NBC Universal, and DreamWorks filed a 110-page copyright infringement lawsuit against AI image-generation company Midjourney in the Central District of California, alleging mass infringement of the companies’ intellectual properties.[5]

The complaint does not mince words, calling Midjourney “the quintessential copyright free-rider and a bottomless pit of plagiarism” and stating that “piracy is piracy, and whether an infringing image or video is made with AI or another technology does not make it any less infringing.”[5] The studios presented evidence that Midjourney can reliably generate high-quality images of copyrighted characters including Darth Vader, Yoda, Spider-Man, and Shrek, sometimes in response to exceptionally generic prompts like “animated toys” or “popular movie screencap.”[6]

What makes the Disney case potentially more dangerous for the AI industry than the text cases is the evidentiary advantage of visual media. A jury can see a Midjourney-generated image of a Marvel character and immediately grasp the alleged infringement in a way that text excerpts from training data cannot match. The studios seek injunctive relief that would effectively force Midjourney to implement copyright protections or face a shutdown of its service.[5]

Midjourney’s Counter-Strategy: The Unclean Hands Defense

Midjourney has not been passive. In a June 29, 2026 court filing, the company demanded that Disney, Universal, and Warner Bros. reveal their own internal use of AI, arguing that if the studios “are themselves training generative-AI models on third-party copyrighted works, scraping the internet and investing in tools that do the same,” this would serve as “powerful evidence that the relevant industry considers such conduct to be fair use.”[7]

Midjourney’s defense rests on the legal doctrine of “unclean hands,” asserting that the studios’ “own unlicensed download and copying of millions of images” violated copyright and showed that the studios are “engaging in the very conduct they complained about.”[7] The studios’ lead lawyer has called Midjourney’s demands a “fishing expedition” designed to distract from the AI company’s own actions. The court has not yet ruled on Midjourney’s motion.[7]


The New York Times v. OpenAI: Sanctions and Suppression

The most consequential text-training case, New York Times Co. v. Microsoft and OpenAI, has taken a sharp turn into evidence suppression allegations. In July 2026, the Times and The Daily News asked the court to sanction OpenAI for allegedly hiding tools and datasets that could identify copyrighted journalism in ChatGPT outputs.[8]

Throughout the case, OpenAI argued it lacked the ability to search its own training corpus and that producing ChatGPT conversation logs would be technically burdensome and raise user-privacy concerns. But in an April 2026 court-ordered deposition, OpenAI data privacy engineer Vinnie Monaco allegedly revealed that the company had already conducted internal searches of its training corpus for copyrighted journalism, had amassed a database of approximately 78 million de-identified ChatGPT conversations to assess infringement, and had implemented a “Bloom” filter as part of a toolset called “Project Giraffe” to detect and record regurgitation in outputs.[8]

The plaintiffs had originally requested 120 million chat logs; OpenAI negotiated the sample down to 20 million, then submitted a sample so heavily redacted that the court called it “unusable.” The plaintiffs also claim OpenAI deleted billions of ChatGPT outputs after the lawsuit was filed, in direct violation of a court preservation order.[8]

The Times’ lead counsel, Ian B. Crosby, framed the discovery dispute as evidence of the merits: “If OpenAI genuinely believed that copying our clients’ journalism was fair and legal, it wouldn’t have hid the truth about having done it.”[8] OpenAI spokesperson Drew Pusateri denied the allegations, accusing the Times of trying to access private user conversations as its case weakens.[8]


The Music AI Settlement Template

The AI music copyright fight has produced the clearest commercial resolution so far. In October 2025, Universal Music Group settled with Udio, establishing the first major-label licensing template for AI music generation. The terms include a per-generation royalty of $0.002 to $0.005 for any output drawing on UMG-controlled recordings, with higher rates for commercially distributed outputs. Udio must implement content identification to flag outputs with substantial similarity to UMG catalog tracks, and UMG retains audit rights over Udio’s training data and generation logs.[9]

This settlement effectively treats AI music companies the same way streaming platforms are treated: per-use fees with monitoring. The framework explicitly covers future model versions, meaning Udio cannot retrain on UMG content without the license. For the entertainment industry, this is the pricing benchmark that every future negotiation with an AI music company will reference.

But the settlement also created a two-tier system. Sony Music and Warner Music declined to join the UMG settlement and continue litigating against Udio, apparently holding out for better terms or a court ruling that gives them leverage. Meanwhile, Suno is fighting on fair use grounds, filing a motion for summary judgment in March 2026 arguing that training on copyrighted music is transformative use. Independent musicians have filed separate class actions against both companies, arguing that major-label settlements do not protect smaller rights holders.[9]


The Emerging Framework: What the Cases Tell Us

Reading the 2025-2026 rulings together, a practical framework is emerging that resolves around four factors:

  • Market competition is the decisive test: When AI outputs compete directly with the market for the copyrighted training data (Ross building a legal research tool to replace Westlaw, or AI-generated images replacing stock photography), courts are finding infringement. When outputs do not replace the original works (Anthropic training on books to build a general-purpose chatbot), fair use arguments survive. The fourth fair use factor has become the battleground.
  • Provenance matters, but courts disagree on how much: The Bartz court held that lawful acquisition strengthens fair use while piracy undermines it, but still found training itself to be fair use. The Kadrey court went further, holding that training is fair use regardless of source. Both courts left the piracy question for separate adjudication, meaning AI companies that trained on pirated datasets face significant retroactive liability even where the training itself is deemed transformative.
  • Output behavior is the vulnerability: Courts have consistently noted whether AI models can reproduce copyrighted content on demand. Midjourney’s ability to generate named characters from generic prompts, and ChatGPT’s alleged regurgitation of Times articles, represent the weakest positions for AI companies. Safeguards against regurgitation were critical to the fair use findings in both Bartz and Kadrey.
  • Licensing is becoming the commercial default: The UMG-Udio settlement demonstrates that licensing frameworks can replace litigation as the path forward. The Copyright Office’s report endorsed this approach, suggesting that market-based solutions are preferable to government intervention and that “premature adoption of statutory, compulsory licenses risks stifling the development of flexible and creative market-based solutions.”[3]

What Comes Next

Several unresolved threads will shape the next phase of AI copyright law:

  • Appellate review of the fair use split: The Thomson Reuters case is on appeal before the Third Circuit, and the divergence between Bartz and Kadrey on the piracy question may prompt appellate courts to weigh in on whether training on pirated works can ever constitute fair use.
  • The Disney v. Midjourney trial: If the case proceeds to trial, a jury will see side-by-side comparisons of Midjourney outputs and copyrighted characters, a visual evidentiary record that text cases cannot match. The outcome could establish a precedent for visual AI training that is more restrictive than the text-training rulings.
  • NYT v. OpenAI sanctions ruling: The court’s response to the sanctions motion will determine whether OpenAI can use its chat log sample as evidence and whether the court will accept as fact that ChatGPT logs would have shown substantial regurgitation. A sanctions ruling against OpenAI could significantly weaken its fair use defense.
  • The two-tier music licensing regime: Whether Sony and Warner extract better terms than UMG, and whether independent artists secure a separate remedy, will determine whether the licensing model scales beyond the major labels or leaves a long tail of unprotected rights holders.

The common thread across all of these cases is that courts are applying a market-competition standard rather than a blanket rule. Training AI on copyrighted works is not automatically legal or illegal; it depends on whether the AI’s output competes with the original work’s market, how the training data was acquired, and whether the model can reproduce protected content. For AI companies, the practical guidance is clear: license what you can, document your data provenance, implement safeguards against regurgitation, and prepare to defend the fair use argument for content that cannot be practically licensed. For rights holders, the strategy is equally clear: register your copyrights, document market harm from AI-generated competing content, and consider whether licensing negotiations offer a faster path to revenue than litigation.


References

[1] [SECONDARY] Norton Rose Fulbright. “AI in Litigation Series: An Update on AI Copyright Cases in 2026.” 2026. https://www.nortonrosefulbright.com/en/knowledge/publications/ce8eaa5f/ai-in-litigation-series-an-update-on-ai-copyright-cases-in-2026

[2] [PRIMARY] Crowell & Moring LLP. “U.S. Copyright Office Releases Third Report on AI and Copyright Addressing Training AI Models with Copyrighted Materials.” June 2, 2025. https://www.crowell.com/en/insights/client-alerts/us-copyright-office-releases-third-report-on-ai-and-copyright-addressing-training-ai-models-with-copyrighted-materials

[3] [PRIMARY] U.S. Copyright Office. “Copyright and Artificial Intelligence: Part 3, Generative AI Training (Pre-Publication Version).” May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf

[4] [SECONDARY] Ohio State University Libraries. “Fair Use and Artificial Intelligence 2026 Update.” March 20, 2026. https://library.osu.edu/site/copyright/2026/03/20/fair-use-and-artificial-intelligence-2026-update/

[5] [SECONDARY] Sparks, Matthew. “Disney, NBC Universal, and DreamWorks File Major IP Lawsuit Against AI Image Generator Midjourney.” Georgetown Law Institute for Technology Law & Policy. June 24, 2025. https://www.law.georgetown.edu/tech-institute/research-insights/insights/disney-nbc-universal-and-dreamworks-file-major-ip-lawsuit-against-ai-image-generator-midjourney/

[6] [SECONDARY] Espiner, Tom and Jamali, Lily. “Disney and Universal sue AI firm Midjourney over images.” BBC News. June 2025. https://www.bbc.com/news/articles/cg5vjqdm1ypo

[7] [SECONDARY] Grant, Daniel. “Midjourney strikes back: sued AI giant demands Hollywood’s secrets.” The Art Newspaper. July 9, 2026. https://www.theartnewspaper.com/2026/07/09/midjourney-demands-hollywood-AI-secrets

[8] [SECONDARY] Bellan, Rebecca. “New York Times says OpenAI hid evidence in ChatGPT copyright trial.” TechCrunch. July 9, 2026. https://techcrunch.com/2026/07/09/new-york-times-says-openai-hid-evidence-in-chatgpt-copyright-trial/

[9] [SECONDARY] AI Vortex. “Suno and Udio AI Music Lawsuit: 2026 Status & Timeline.” July 30, 2026. https://www.aivortex.io/legal/ai-case-law/suno-udio-music-ai/


AI automated and generated content. Human reviewed.


Listen to the Bears Discuss This

Episode 18 of AI Policy for the Bears

The Weekly AI Policy Brief

AI policy moves fast. Get the week’s developments, new trackers, and podcast episodes in one short email — free, ad-free, no spam.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *