AI and Fair Use Update

Posted on by

Originally published April 6, 2025 AI in the World of Entertainment, Arts, and Sports Law. Updated August 17, 2026 by Elliot Zimmerman.

When this article first ran, not a single U.S. court had squarely decided whether training an artificial intelligence model on copyrighted works is fair use. Sixteen months later, three federal judges have ruled on the question, a federal appeals court has heard the first appeal, the largest copyright settlement in American history has been approved, and the Register of Copyrights has spent more than a year litigating her own firing. Here is where things stand.

The Fair Use Rulings That Changed the Landscape

Every one of these decisions runs through the four-factor analysis of 17 U.S.C. § 107 as reframed by the Supreme Court in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023),1 which tied transformativeness to the specific use rather than the general character of the new work. That framing has done real work in the AI cases.

Thomson Reuters v. ROSS Intelligence: The First Decision, and the First Appeal

In February 2025, Judge Stephanos Bibas, sitting by designation in the District of Delaware, reversed his own earlier view and held on summary judgment that ROSS Intelligence’s use of Westlaw headnotes to train a legal research tool was not fair use. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025).2 He awarded factors one and four (purpose and market harm) to Thomson Reuters, and factors two and three to ROSS, then held the balance defeated the defense.

The case is now the first AI training dispute to reach a federal appellate court. The Third Circuit heard oral argument on June 11, 2026, with the panel pressing both sides on transformativeness and market definition.3 ROSS leaned heavily on the Third Circuit’s decision in American Society for Testing & Materials v. UpCodes, Inc. (3d Cir. Apr. 7, 2026),4 which found transformative use in the republication of standards incorporated into law despite verbatim copying. A decision remains pending and will be the first appellate word on AI training and fair use.

Bartz v. Anthropic: Buy the Book, Don’t Torrent It

On June 23, 2025, Judge William Alsup drew a line that has shaped every case since: using lawfully acquired books to train a large language model was fair use, but downloading and retaining pirated copies from shadow libraries was not. Bartz v. Anthropic PBC, No. 24-cv-05417 (N.D. Cal. June 23, 2025).5 Facing a piracy trial and statutory damages across roughly half a million works, Anthropic settled for $1.5 billion.

Judge Araceli Martínez-Olguín granted final approval on July 20, 2026,6 ending the largest copyright class settlement in history. Key terms: approximately $3,000 per work across a works list of 482,460 titles, a 91.3% claims rate, attorneys’ fees trimmed from the requested $187.5 million to roughly $101.6 million, and a release limited to past conduct (outputs and future conduct are expressly carved out). Anthropic must also destroy the pirated files. Practitioners advising content owners should read the release language closely; it is narrower than the headline number suggests. The Authors Guild maintains a claims timeline and FAQ.

Kadrey v. Meta: A Win for Meta, but a Roadmap for Plaintiffs

Two days after Bartz, Judge Vince Chhabria granted Meta summary judgment on fair use in Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. June 25, 2025).7 Meta also prevailed on the plaintiffs’ claims under 17 U.S.C. § 1202 for removal of copyright management information.8

The reasoning matters more than the result. Judge Chhabria found the training use transformative and rejected the “lost licensing market” theory as circular, but he identified market dilution (the flooding of markets with AI-generated substitutes) as potentially the decisive issue, and said plainly that better evidence on that point could have sent factor four to a jury or even won the case for the plaintiffs. As commentators noted at the time, this was a ruling on a record, not a blessing of the practice. Expect every plaintiff’s expert report from here forward to be built around market dilution.

Case-by-Case Updates

New York Times v. Microsoft/OpenAI: Now the Center of a Multidistrict Litigation

Judge Sidney H. Stein largely denied the motions to dismiss in the spring of 2025, allowing the core infringement claims in The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y.),9 to proceed. On April 3, 2025, the Judicial Panel on Multidistrict Litigation centralized the Times case and the parallel author and publisher class actions before Judge Stein as In re OpenAI, Inc., Copyright Infringement Litigation, MDL No. 3143.10

Two developments since then are worth flagging. First, on October 27, 2025, Judge Stein refused to dismiss output-based infringement claims, finding that a reasonable jury could deem certain ChatGPT outputs substantially similar to the underlying books.11 That is a meaningful crack in the “training only” framing of these disputes. Second, after a bruising discovery fight, the court ordered production of a 20-million-conversation sample of de-identified ChatGPT logs, and Judge Stein affirmed that order in early January 2026 over OpenAI’s user-privacy objections.

The case remains in active discovery. Fair use has not been decided. The docket is public on CourtListener, and Judge Stein’s opinions in the MDL are available through RECAP.

Andersen v. Stability AI: Headed to a Jury

The artists’ class action has survived. Andersen v. Stability AI Ltd., No. 3:23-cv-00201-WHO (N.D. Cal.).12 Judge William Orrick’s August 12, 2024 order dismissed the § 1202 claims with prejudice but let direct and induced infringement and Lanham Act trade dress theories proceed against Stability AI, Midjourney, DeviantArt and Runway. A third amended complaint was filed in February 2026, and the court has set a jury trial for April 5, 2027, with plaintiffs anticipating a two-week trial.13

The central question, whether outputs generated “in the style of” an artist can constitute infringing derivative works when style itself is not protectable, has never been tried to a jury in this context. Track it on the CourtListener docket.

Disney, Universal and Warner Bros. v. Midjourney: The Studios Enter

On June 11, 2025, Disney and NBCUniversal sued Midjourney in the Central District of California, seeking injunctive relief that would effectively require output filtering for their characters. Disney Enterprises, Inc. v. Midjourney, Inc., No. 2:25-cv-05275 (C.D. Cal.).14 Warner Bros. filed a parallel action, and the cases were consolidated in November 2025 with the Disney matter as the lead.

This is the case entertainment counsel should watch most closely, because it is about outputs and recognizable characters rather than the more abstract training question. Midjourney has asserted fair use and gone on the offensive in discovery, demanding that the studios disclose their own internal AI use. A magistrate judge limited that request to consumer-facing content, and Midjourney has moved to have the ruling reviewed under Fed. R. Civ. P. 72(a). Expert disclosures are scheduled for late 2026. Docket here.

Getty Images v. Stability AI: The First Foreign Judgment

On November 4, 2025, the High Court of England and Wales handed down Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch),15 the first trial judgment anywhere on generative AI and copyright. Getty abandoned its primary copyright claims mid-trial over territoriality problems (the training occurred outside the UK), and the court rejected the surviving secondary infringement claim, holding that the model weights contain no copies of the works. Getty won only narrow trademark findings tied to watermarks appearing in outputs.

The practical lesson is jurisdictional: where training happens matters enormously. Getty’s parallel U.S. action, filed in Delaware in 2023, now proceeds in the Northern District of California.16

Music: from litigation to licensing

The recorded music industry has largely converted its lawsuits into deals. The RIAA’s June 2024 actions, UMG Recordings, Inc. v. Suno, Inc., No. 1:24-cv-11611 (D. Mass.),17 and UMG Recordings, Inc. v. Uncharted Labs, Inc. (Udio), No. 1:24-cv-04777 (S.D.N.Y.)18, have largely resolved through settlement. Universal settled with Udio in October 2025; Warner settled with Udio and then with Suno in November 2025, in each case paired with licensing arrangements for new, licensed models. Universal and Sony have continued litigating against Suno. Notably, the American Federation of Musicians filed suit in June 2026 against Universal and Warner, alleging member recordings were licensed into these AI deals without the compensation or consent the collective bargaining agreement requires. It is the first significant challenge aimed at the settlements themselves.

Publishing has gone the other way. Concord Music Group, Inc. v. Anthropic PBC, filed in the Middle District of Tennessee in October 2023 and transferred to the Northern District of California, continues over roughly 500 compositions.19 The publishers filed a second, much larger action on January 28, 2026 covering roughly 21,000 works and seeking over $3 billion, this one built on the distribution and torrenting theory that proved so costly for Anthropic in Bartz.20 BMG filed its own suit in March 2026. An amended complaint landed in July 2026.

Right of publicity: from “Sky” to Sora

The Scarlett Johansson/OpenAI dispute over the “Sky” voice now reads as a preview. It raised the theory of Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988),21 in which the Ninth Circuit held that deliberately imitating a distinctive voice to sell a product is actionable. In late 2025, OpenAI’s Sora video app generated a far larger controversy when users flooded it with realistic videos of deceased public figures (Robin Williams, George Carlin, Martin Luther King Jr.) because likeness protections initially applied only to the living. OpenAI reversed course, moving to an opt-in model for living figures and permitting estates of recently deceased figures to request blocks.

For entertainment counsel, the takeaway is that post-mortem publicity rights (California’s Civil Code § 3344.1 among them, with other states offering shorter terms or none at all) are now a front-line commercial issue rather than an academic one.22 Midler remains good law, but a state-by-state patchwork is a poor fit for a global platform, which is precisely the argument driving federal legislation.

Regulatory and Legislative Developments

The Copyright Office Reports, and the Fight Over the Register

The Office completed its three-part study: Part 1 on digital replicas (July 2024), Part 2 on copyrightability (January 2025), and, on May 9, 2025, a pre-publication version of Part 3 on generative AI training.23 Part 3 rejects the argument that training is categorically transformative or non-expressive, concludes that model weights may infringe where outputs are substantially similar to inputs, treats retrieval-augmented generation as involving reproduction, and takes market dilution seriously. The Office has said a final version will follow without substantive changes to the analysis; the pre-publication text remains the operative document.

What happened next is now part of the story. The Librarian of Congress was dismissed the day before Part 3’s release, and Register of Copyrights Shira Perlmutter was removed the day after. She sued. The D.C. Circuit granted an injunction pending appeal in September 2025, reasoning that the removal was likely unlawful because only the Librarian of Congress may remove the Register under 17 U.S.C. § 701, and on June 30, 2026 the Supreme Court declined to disturb that injunction. Perlmutter v. Blanche, No. 1:25-cv-01659 (D.D.C.), inj. pending appeal granted, No. 25-5285 (D.C. Cir. Sept. 10, 2025), stay denied sub nom. Blanche v. Perlmutter (U.S. June 30, 2026).24 Perlmutter remains in office while the litigation proceeds. Whatever one makes of the merits, the episode has left the Office’s policy output under a cloud of contested authority at exactly the moment courts are looking to it for guidance.

Human Authorship Is Settled; at the Edges, It Is Not

On March 2, 2026, the Supreme Court denied certiorari in Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied, No. 25-449 (U.S. Mar. 2, 2026),25 leaving intact the D.C. Circuit’s holding that “authorship” under 17 U.S.C. § 102(a) requires a human being. Purely AI-generated material remains unregistrable, and applicants must still disclose and disclaim AI-generated portions.

The unresolved question is how much human contribution suffices. Prompting alone has not been enough (the Office denied registration to works generated through hundreds of iterative prompts), but no court has drawn the line. Document human creative control contemporaneously; it is the cheapest insurance available to a client using these tools in production.

Legislation

  • NO FAKES Act. The Generative AI Copyright Disclosure Act, H.R. 7913, 118th Cong. (2024), died with that Congress.26 The action has moved to digital replicas: the NO FAKES Act of 2026, S. 4591, 119th Cong.,27 would create a federal, licensable, descendible property right in voice and visual likeness, with a notice-and-takedown regime and preemption of state digital replica laws enacted after January 2, 2025. The Senate Judiciary Committee advanced it unanimously on June 18, 2026. It has not been enacted.
  • DEFIANCE Act. Passed the Senate by unanimous consent on January 13, 2026; pending in the House.
  • TAKE IT DOWN Act. Enacted in May 2025, addressing non-consensual intimate imagery, including AI-generated material, with platform removal obligations.
  • States. Tennessee’s ELVIS Act, California’s digital replica statutes and New York’s 2025 amendments continue to do the real work in this area. Roughly thirty states now regulate election deepfakes. Note the preemption clause in NO FAKES: clients relying on newer state statutes should understand what a federal bill would displace.

Where This Leaves Us

Four things have become clear since the original version of this article:

  1. How the data was acquired matters more than what was done with it. Bartz, the Anthropic settlement and the publishers’ torrenting suits all turn on piracy, not on training. Training may well be fair use; downloading from a shadow library to get there is a separate, expensive act.
  2. Market dilution is the battleground. Judge Chhabria and the Copyright Office arrived at the same place from different directions. Plaintiffs who can prove that AI output floods the market for works of the same kind have a real path on factor four; those who cannot will lose on a thin record, as the Kadrey plaintiffs did.
  3. Outputs are the more dangerous exposure. The output-based rulings in the OpenAI MDL, the studios’ character-focused claims against Midjourney and the guardrails stipulation in the Anthropic lyrics case all point the same direction. Clients deploying generative tools should be auditing what comes out, not just what went in.
  4. Licensing is arriving ahead of the law. The music industry moved from mass litigation to signed deals in about eighteen months. Expect the same pattern elsewhere, with the unresolved question, as the musicians’ union suit against the labels shows, being who at the bottom of the chain actually gets paid.

The next milestones to watch: the Third Circuit’s decision in Thomson Reuters v. ROSS, the first summary judgment briefing on fair use in the OpenAI MDL, and the April 2027 Andersen trial, the first time a jury will be asked what an AI-generated image in an artist’s style actually is.

Notes

  1. Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023). ↩
  2. Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025) (Bibas, J., sitting by designation) (granting partial summary judgment on direct infringement as to 2,243 headnotes and rejecting fair use). ↩
  3. Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir.) (interlocutory appeal; argued June 11, 2026 before Restrepo, Montgomery-Reeves and Bove, JJ.). Argument audio is posted on the Third Circuit’s website. ↩
  4. Am. Soc’y for Testing & Materials v. UpCodes, Inc. (3d Cir. Apr. 7, 2026). ↩
  5. Bartz v. Anthropic PBC, No. 24-cv-05417 (N.D. Cal. June 23, 2025) (Alsup, J.) (order on cross-motions for summary judgment re fair use). ↩
  6. Bartz v. Anthropic PBC, No. 24-cv-05417 (N.D. Cal. July 20, 2026) (Martínez-Olguín, J.) (order granting final approval of class settlement and entering final judgment; overruling 53 objections; reducing class counsel’s fee request from $187.5 million to approximately $101.6 million and service awards from $50,000 to $15,000 each). ↩
  7. Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. June 25, 2025) (Chhabria, J.) (granting Meta partial summary judgment on fair use as to the thirteen named plaintiffs’ works). ↩
  8. 17 U.S.C. § 1202(b) (removal or alteration of copyright management information). ↩
  9. The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y. filed Dec. 27, 2023) (Stein, J.; Wang, M.J.). ↩
  10. In re OpenAI, Inc., Copyright Infringement Litig., MDL No. 3143, No. 1:25-md-03143 (SHS) (OTW) (S.D.N.Y.) (centralized Apr. 3, 2025). ↩
  11. In re OpenAI, Inc., Copyright Infringement Litig., No. 1:25-md-03143 (SHS) (OTW), Dkt. No. 701 (S.D.N.Y. Oct. 27, 2025) (opinion and order denying OpenAI’s motion to dismiss the class plaintiffs’ output-based direct infringement claim). ↩
  12. Andersen v. Stability AI Ltd., No. 3:23-cv-00201-WHO (N.D. Cal. filed Jan. 13, 2023) (Orrick, J.); order granting in part and denying in part motions to dismiss (Aug. 12, 2024). ↩
  13. Andersen, No. 3:23-cv-00201-WHO, Joint Case Management Statement, Dkt. No. 379 (N.D. Cal. Jan. 13, 2026) (trial set for Apr. 5, 2027 per Dkt. No. 240). ↩
  14. Disney Enters., Inc. v. Midjourney, Inc., No. 2:25-cv-05275 (C.D. Cal. filed June 11, 2025) (Richlin, M.J.); consolidated with the Warner Bros. action Nov. 4, 2025, with the Disney matter designated the lead case. ↩
  15. Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), Case No. IL-2023-000007 (Joanna Smith J., 4 Nov. 2025). ↩
  16. Getty Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135 (D. Del. filed Feb. 3, 2023); the U.S. action now proceeds as No. 3:25-cv-06891 (N.D. Cal.). ↩
  17. UMG Recordings, Inc. v. Suno, Inc., No. 1:24-cv-11611 (D. Mass. filed June 24, 2024). ↩
  18. UMG Recordings, Inc. v. Uncharted Labs, Inc. (d/b/a Udio), No. 1:24-cv-04777 (S.D.N.Y. filed June 24, 2024). ↩
  19. Concord Music Grp., Inc. v. Anthropic PBC, No. 3:23-cv-01092 (M.D. Tenn. filed Oct. 18, 2023), transferred, No. 5:24-cv-03811 (N.D. Cal.) (Lee, J.) (“Concord I”). The court denied the publishers’ motion for a preliminary injunction and separately approved a stipulation requiring Anthropic to maintain output guardrails. ↩
  20. Concord Music Grp., Inc. v. Anthropic PBC, No. 5:26-cv-00880 (N.D. Cal. filed Jan. 28, 2026) (“Concord II”). ↩
  21. Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988). ↩
  22. Cal. Civ. Code § 3344.1 (post-mortem right of publicity); see also Tenn. Code Ann. §§ 47-25-1101 et seq. (ELVIS Act). ↩
  23. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 1: Digital Replicas (July 2024); Part 2: Copyrightability (Jan. 2025); Part 3: Generative AI Training (Pre-Publication Version) (May 9, 2025). ↩
  24. Perlmutter v. Blanche, No. 1:25-cv-01659 (D.D.C.) (Kelly, J.), inj. pending appeal granted, No. 25-5285 (D.C. Cir. Sept. 10, 2025) (Pan and Childs, JJ., concurring), application for stay denied sub nom. Blanche v. Perlmutter (U.S. June 30, 2026); see 17 U.S.C. § 701 (appointment of the Register by the Librarian of Congress). ↩
  25. Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied, No. 25-449 (U.S. Mar. 2, 2026); see 17 U.S.C. § 102(a). Compare Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022) (same result under the Patent Act). ↩
  26. Generative AI Copyright Disclosure Act of 2024, H.R. 7913, 118th Cong. (2024). ↩
  27. Nurture Originals, Foster Art, and Keep Entertainment Safe Act of 2026 (NO FAKES Act), S. 4591, 119th Cong. (2026); reported out of the Senate Judiciary Committee June 18, 2026. ↩

Sources and Further Reading

This article is provided for general informational purposes and is not legal advice. Case postures change quickly; verify the current status of any matter on the docket before relying on it.