Wholestory

The largest AI copyright case reached the merits on 4 September, when both sides in the consolidated In re: OpenAI litigation (25-md-3143, S.D.N.Y.) moved for summary judgment before Judge Sidney Stein. OpenAI filed twice — a 46-page brief in the authors' class case (Authors Guild v. OpenAI, 23-cv-8292, Dkt. 1221) and a 33-page brief answering the news publishers (23-cv-11195, Dkt. 1496) — arguing that pretraining is fair use as a matter of law, that nobody owns the facts a newspaper reports, and that its web-browsing feature was impliedly licensed because the papers left its crawler unblocked for months. The Authors Guild and thirteen authors, among them George R.R. Martin, John Grisham and Jonathan Franzen, cross-moved on 194 works and argue Microsoft is vicariously liable. Oppositions are due in early October, replies in November.

In its books brief OpenAI states plainly that compilations from Library Genesis — which the US Trade Representative lists as a notorious market for piracy — trained GPT-3 in spring 2020 and GPT-3.5 in late 2021, and that no other model in the case used it. It then asks the court to reject the line Judge Alsup drew in Bartz v. Anthropic, arguing that where the end purpose is transformative "the entire course of conduct undertaken in service of that transformative purpose—including intermediate copying—is fair use". The authors reply that Warhol requires each act of copying to be judged on its own.

The Seattle Times and Newsday filed a 38-page copyright and trademark complaint against OpenAI and Microsoft in the Southern District of New York, alleging the companies scraped their sites, paywalled articles included, to train and run ChatGPT, Copilot and Bing's AI features. They want damages and the "impoundment and/or destruction" of any dataset or model containing their work, and cite a 47 percent year-on-year fall in search referrals to midsize publishers in December 2025. Their trademark claim — over fabricated stories attributed to them — is the kind the Times abandoned with prejudice four weeks earlier. No docket number has been published.

A month before the summary judgment motions, Judge Sidney Stein recorded the damage a new Supreme Court decision had done to the news publishers' case. In Cox Communications v. Sony Music (146 S. Ct. 959 (2026)) the Court held that "material contribution" is not a valid theory of contributory copyright liability — the theory on which The New York Times, the Daily News and the Center for Investigative Reporting had all pleaded that OpenAI and Microsoft were answerable for their users' infringement. Facing OpenAI's motion for judgment on the pleadings, all three dropped those claims with prejudice; the Times and Daily News dropped their trademark dilution claims too. Stein then refused to let them re-plead inducement against Microsoft, holding those theories existed before Cox and switching now "would require reopening discovery into Microsoft's intent".

The United States filed a 20-page Statement of Interest under 28 U.S.C. §517 in the consolidated In re: OpenAI Copyright Infringement Litigation (No. 25-md-3143, S.D.N.Y., Judge Sidney Stein), as the cases head into summary judgment — the first time the federal executive has taken a merits position on AI training. It urges the court to reject any argument that training infringes, calling the use "transformative — spectacularly so", and attacks the "market dilution" theory Judge Chhabria floated in Kadrey as "deeply flawed": under Warhol, outputs are a separate use from training, and works merely in the same genre are not market harm. Pointedly, it says the Register of Copyrights — "currently challenging her removal" — endorsed that theory in the Office's May 2025 report, that her view "does not warrant deference", and that her reasoning was "threadbare". It grounds the position in AI "dominance" and warns mandatory licensing would build an "oligopoly".

Sony Music Publishing, Warner Chappell and about three dozen affiliated publishers filed a 48-page complaint in the Northern District of California, and the caption is the news: CEO Dario Amodei and co-founder Benjamin Mann are defendants in their personal capacities. Two counts run against the individuals — direct infringement from torrenting, and contributory infringement for the same acquisition conduct — and two against the company, including a DMCA §1202 claim for stripping copyright management information, which carries $2,500 to $25,000 per violation on its own. The publishers allege lyrics were taken from Books3, The Pile, Common Crawl and the aggregators MusixMatch and LyricFind, covering tens of thousands of compositions, and that Claude reproduces them verbatim. Anthropic calls it "the third lawsuit from the same lawyers, recycling allegations from cases already before the courts" and says training is fair use. The complaint itself was not obtained.

wikiHow, Inc. filed a copyright infringement action against nine OpenAI entities in the Southern District of New York on 21 August — wikiHow, Inc. v. OpenAI, Inc. et al, No. 1:26-cv-07171, cause 17:101 — represented by Susman Godfrey. With it came a statement of relatedness asking that the case be treated as related to the consolidated OpenAI copyright proceeding, MDL No. 1:25-md-03143, which is the structurally interesting part: the docket is consolidating rather than fragmenting. The complaint itself could not be read. Reporting describes unauthorised scraping of more than 11,000 how-to guides and verbatim reproduction by ChatGPT; that allegation is carried as reported and has not been checked against the pleading.

Judge Jacqueline Scott Corley dismissed the part of author E. Molly Tanzer's claim against Adobe that concerns NVIDIA's Nemotron model, in Tanzer v. Adobe Inc., No. 3:26-cv-04712-JSC (N.D. Cal., 20 August). The ruling decides nothing about fair use. It is about pleading: alleging that NVIDIA "acquired millions of unauthorized copies" from the shadow library Anna's Archive, and that piracy "has been commonplace with the AI development community", is conclusory — and, the court wrote, repeating a conclusory allegation several times does not make it less conclusory. Without facts supporting an inference that the model saw her books by that route there is neither Article III standing nor a direct infringement claim. Dismissal is with leave to amend by 2 October. The obstacle in front of plaintiffs is increasingly provenance, not the exception.

Round Hill Music LP and its royalty funds filed two copyright complaints in the US District Court for the Northern District of California, against Suno (reported as No. 5:26-cv-08507) and against Anthropic (No. 5:26-cv-08505). Round Hill manages a US$1.1 billion portfolio of more than 200,000 songs — among the asserted works are the Goo Goo Dolls' "Iris", Bonnie Tyler's "Total Eclipse of the Heart" and James Brown's "I Got You (I Feel Good)" — and is represented by Richard S. Busch, who won the "Blurred Lines" verdict for the Marvin Gaye family. Each complaint pleads 500 "bell-weather" compositions preliminarily and states that the plaintiffs intend to amend to list "potentially ten thousand or more" works, so that statutory damages "would ultimately exceed hundreds of millions of dollars and potentially approach or conceivably exceed $1 billion". Three things make the filings more than another entry on the docket. They plead the Digital Millennium Copyright Act alongside the Copyright Act, alleging both the removal of copyright management information and the circumvention of technological access controls to pull protected audio from licensed platforms. They take Judge Alsup's line in Bartz — that a permanent library of unlicensed copies is not fair use — and turn it against training itself, arguing that a "permanent stockpile of copyrighted material for open-ended future exploitation fails every factor of the fair-use analysis" because "the purpose is acquisitive rather than transformative". And the Suno action names a third defendant that is not a model builder at all: Bright Data, the Israeli web-scraping vendor, on a contributory theory that its tools are "specifically designed to scrape music files en masse and strip them of CMI", pleading that it passed $300 million in revenue in 2025 with $400 million projected for 2026. On memorisation, the Anthropic complaint alleges that "despite implementing partial output filters in response to prior litigation", Claude still reproduces "the vast majority of Round Hill's catalog". These are allegations; no court has tested any of them.

S.A. Jamendo, the Luxembourg subsidiary of the Winamp Group, filed a voluntary notice of dismissal without prejudice in the US District Court for the District of Massachusetts (No. 1:26-cv-12966-AK), six weeks after suing Suno over the MTG-Jamendo Dataset, which draws more than 55,000 tracks from its catalogue and was released for non-commercial research on terms requiring a licence for commercial AI training. The notice, signed by attorney Alexander T. Hornat, gives no reason and mentions no settlement, and records that no defendant had served an answer or a motion for summary judgment — leaving Jamendo free to bring the same claims again. Around the same time Poseidon Wave Media, the entity behind the instrumental duo The American Dollar, dismissed its own suit against Suno with prejudice, so it cannot be refiled; it had been filed in the Southern District of New York on 12 May 2026 and asserted 236 recordings and compositions on a claim that the duo's licensing revenue had been nearly eliminated. Jamendo's parallel suit against Nvidia over the same dataset, in the Northern District of California, has not been withdrawn, and a related commercial claim of about €16 million against Nvidia Technologies Belgium is pending in Belgium's Ghent Enterprise Court. Neither dismissal states a reason, and neither is reported as settled; Digital Music News suggests Jamendo, headquartered in Luxembourg, may prefer a European forum after GEMA's Munich win, which is that outlet's inference and not a finding of the court. The larger actions continue: Universal Music Group and Sony Music Entertainment are still suing Suno in the same Massachusetts district.

BMG and Suno announced a global strategic alliance covering BMG's recorded music and publishing repertoire, twelve days after the Regional Court of Munich I held that Suno was not entitled to train on, store or reproduce works represented by the German rights society GEMA. The agreement is built on an opt-in principle — BMG artists and songwriters who choose to participate have their rights protected and are compensated — and it also settles Suno's prior use of BMG recordings and publishing works, which is the part that matters to the price of a work: an unlicensed corpus is retrospectively paid for rather than adjudicated. No financial terms were disclosed by either party. It follows Suno's November 2025 settlement and licensing partnership with Warner Music Group, and comes six days after Suno's co-founder and chief executive Mikey Shulman published company principles committing to audio watermarking and fingerprinting and to a new downloads policy limiting mass distribution of generated tracks. BMG is simultaneously a plaintiff against Anthropic in the lyrics litigation, so the same company is buying a licence from one model builder while suing another.

The Supreme Court declined to let President Trump remove Shira Perlmutter as Register of Copyrights while her challenge to her firing proceeds, denying the administration's application to stay a lower-court order that had restored her. The one-line order in Blanche v. Perlmutter (No. 25A478) read: "The application for stay is denied. The denial of the application is not a ruling on the merits of the legal issues presented in the litigation." It reversed the picture left by the summer of 2025, when a district judge had refused Perlmutter emergency relief: the D.C. Circuit had since overturned that refusal and ordered her reinstated, calling the case a "genuinely extraordinary situation," and the administration's bid to undo that reinstatement is what the Justices turned down. The timing was pointed. A day earlier the Court had decided two removal-power cases — holding in Trump v. Slaughter that the Federal Trade Commission's for-cause protections for its commissioners are unconstitutional and that officers wielding executive power answer to the President, while carving out the Federal Reserve alone in Trump v. Cook — and it had been holding the Perlmutter application precisely until those came down. So Perlmutter keeps the office that produced the Copyright Office's AI-training report for now, but the reasoning most likely to govern whether her firing was lawful has just been written against her, and the Court left the merits to the lower courts. Congress is pulling the other way: the House passed a bill in June 2026 that would strip the Librarian of Congress of authority over the Copyright Office and hand appointment of the Librarian to a bipartisan commission.

The Regional Court of Munich I (Landgericht München I, 42nd Civil Chamber, presiding judge Elke Schwager) ruled for the German rights society GEMA against the AI music generator Suno in first-instance judgment 42 O 763/25, holding that Suno infringed copyright by training on six protected works without a licence and by reproducing them in its outputs. The court found the model had ‘memorised’ and stored the works in reproducible form, which defeats the text-and-data-mining exception; held the provider rather than the user liable; held compliance with the EU AI Act no defence; and found the U.S. training copies not protected by fair use, expressly distinguishing the American Bartz and Kadrey rulings on the ground that there the training materials were not reproduced in outputs. It took jurisdiction over the U.S. training acts under the German Collecting Societies Act (VGG), noting Suno had used stream-ripping to pull tracks from YouTube in circumvention of its ‘rolling cipher’ protection. It ordered Suno to cease the reproduction and the use of the works for training, to disclose the revenue connected to the infringements before damages are set, and confirmed an entitlement to damages whose amount is not yet fixed. It is the second time the same court has held that AI training infringes German copyright — it held in November 2025 in GEMA v OpenAI that training on song lyrics infringes — and it moves German law further from the American line in Bartz v. Anthropic, where training on lawfully acquired works was fair use and only pirated acquisition was punished: under the Munich rulings the training itself is the infringement, and licensing, not acquisition hygiene, is the price of entry. The judgment is first-instance and not final; Suno intends to appeal. The disclosure order lands on a company carrying a $5.4bn valuation and roughly $300m in annual recurring revenue, while parallel U.S. suits by Universal and Sony seek up to $150,000 per work. Accounts differ on immediate enforceability — juve-patent reports the judgment is not yet enforceable, while Forbes reports a German first-instance judgment can be enforced pending appeal, typically on posting security.

Judge Araceli Martínez-Olguín granted final approval to the $1.5 billion Bartz v. Anthropic settlement in the Northern District of California, closing the authors' pirated-books class action after a preliminary approval in September 2025 and a fairness hearing in May 2026. She took over the case after Judge William Alsup — who wrote the underlying rulings that training on lawfully acquired books is fair use but downloading pirated copies is not — retired at the end of 2025. Her order found the agreement "provides meaningful relief to the Settlement Class," noted that the estimated per-work payment of roughly $3,000 is "four times the minimum statutory damages amount for willful infringement," recorded that 92% of those eligible for an award had opted into the class, and denied all but two opt-out requests. The court separately awarded plaintiffs' attorneys $101.56 million in fees — about 6.8 percent of the fund, well under the quarter-to-third customary in class actions, a restraint the fee order attributed to the settlement's unusual size.

Justice Amit Bansal declined to grant news agency ANI Media an interim injunction against OpenAI, holding that ANI had not made out a prima facie case of infringement on either of its claims. On training, the court held that storing a literary work in electronic form is reproduction under §14(a)(i) of the Copyright Act, 1957, but that §14 is subject to §52 — and that, read broadly, the "private or personal use, including research" exception in §52(1)(a) prima facie covers a closed, internal training process. It held that "private" is not confined to individuals and can extend to a company or closed group where the material is not made available to the public, that research may be undertaken commercially, and that commercial character does not by itself defeat the exception. On outputs, it found the examples ANI relied on were published after the relevant model training cut-offs and so could not evidence memorisation, and that copyright protects the form and arrangement of expression rather than the underlying facts or news events. The court declined to import the US four-factor fair use test, weighing instead market substitution, functional difference and public interest, and noted that ANI had shown no loss of subscribers or licensing revenue and had itself offered OpenAI a licence. It upheld territorial jurisdiction notwithstanding that OpenAI's servers sit outside India. This is an interim order on a prima facie standard, not a final ruling on the merits; the suit continues.

The Regional Court of Munich I ruled in GEMA v OpenAI (Case No. 42 O 14139/24) that using copyrighted song lyrics to train generative models without a licence violates German copyright law — the first European decision to address the question directly. The court held that the text-and-data-mining exception in §44b UrhG does not apply because the models permanently memorise lyrics rather than analysing them transiently, that storage of a work inside the model is reproduction under §16 UrhG, and that ChatGPT's output of lyrics in response to prompts is communication to the public under §19a UrhG; it rejected OpenAI's argument that users rather than the platform were responsible. The remedies went beyond an injunction: OpenAI must stop storing unlicensed German lyrics on infrastructure in Germany and publish the judgment in a local newspaper, with damages to be assessed later. Personality-rights claims were dismissed. OpenAI is appealing, so the ruling is not final. GEMA, which represents more than 95,000 German composers, songwriters and publishers, brought a second suit against the music-generation service Suno before the same court in January 2025.

Concord, Universal Music Publishing Group and ABKCO filed a 71-page Second Amended Complaint against Anthropic in the Northern District of California, drawing on internal records obtained in discovery — including an alleged co-founder query for Bob Dylan lyrics and a decision to use an extraction tool that stripped copyright notices — and seeking statutory damages up to $150,000 per work. It follows a separate January 2026 suit by the same publishers covering more than 20,000 songs and seeking over $3 billion for alleged torrenting of lyrics.

In the UK, the High Court (Mrs Justice Joanna Smith) largely ruled for Stability AI, holding that a model such as Stable Diffusion, which does not store or reproduce the copyrighted works, is not an "infringing copy." The framing of the result as a sweeping AI "win" understated what happened: Getty had voluntarily withdrawn its main training-copyright claim mid-trial because the training occurred outside the UK, and the court actually ruled for Getty on some trademark (watermark) claims.

Anthropic agreed to pay about $1.5 billion to settle the Bartz class action over pirated books used to train Claude — the largest copyright recovery on record and the first major AI-era settlement. It covers an estimated 500,000 works at roughly $3,000 each, with Anthropic also agreeing to destroy the downloaded files, resolving the piracy liability that survived Alsup's June ruling and averting the December trial. Court filings had shown Anthropic downloaded more than 7 million books it knew were pirated (from Books3, LibGen and the Pirate Library Mirror) before shifting to buying and scanning physical books.

Getty Images filed a notice of voluntary dismissal without prejudice of its Delaware copyright suit against Stability AI, stating it intended to refile in the Northern District of California to move faster. Defendants had filed no answer or summary-judgment motion. The Delaware docket's termination was Getty's own venue-driven withdrawal, not a ruling on the merits.

Two days after Bartz, Judge Vince Chhabria granted Meta summary judgment that training Llama on books was fair use — but "on this record" only, and reluctantly, faulting the thirteen author-plaintiffs for failing to develop a market-harm case and noting Llama could not reproduce more than trivial snippets. He expressly disagreed with Alsup on pirated "shadow library" copies and rejected the "teaching children to write" analogy. The paired June rulings left US fair-use law genuinely unsettled.

In Bartz v. Anthropic (N.D. Cal., No. 24-cv-5417), Judge William Alsup ruled on summary judgment that training Claude on books was "exceedingly transformative" fair use, and that digitizing lawfully purchased print books for training was also fair use. But he held that Anthropic's acquisition and retention of a permanent library of pirated books was not fair use, preserving liability for the piracy and setting a December 2025 trial on damages. In July the court certified a class for the piracy claim only, so the fair-use-on-training holding bound just the three named plaintiffs.

Shira Perlmutter sued President Trump over her firing, arguing it was unconstitutional because the Copyright Office and Library of Congress sit within the legislative branch; the administration countered that the removals were lawful under the Federal Vacancies Reform Act. A federal judge denied her request for a temporary restraining order, with further hearings set for July 2025 — leaving the leadership of the US copyright system contested.

Register of Copyrights Shira Perlmutter was fired on 10 May 2025 — one day after the Part 3 AI report's release, and two days after Librarian of Congress Carla Hayden was dismissed. The administration named DOJ officials Todd Blanche as Librarian and Paul Perkins as Register; the Office then paused issuing roughly 20,000 registration certificates for about 12 business days. NPR reported the sequence between the report and the firing without asserting causation.

The U.S. Copyright Office released a "pre-publication version" of Part 3 of its AI report, on using copyrighted works to train generative AI. Its central conclusion: many training uses are likely transformative, but fair use turns case-by-case on the works used, the source, the purpose and output controls. The report introduced a "market dilution" concern — that AI output competing in the same market can harm the market for originals even without direct copying. Observers called the unusual pre-publication release, issued with a final version promised, unprecedented.

In Thomson Reuters v. ROSS Intelligence, Judge Stephanos Bibas (sitting in the District of Delaware) granted Thomson Reuters summary judgment on direct copyright infringement and rejected ROSS's fair-use defense — the first US merits ruling on fair use in an AI case. The court found ROSS copied 2,243 Westlaw headnotes, held the headnotes and Key Number System copyrightable, and found fair-use factors one (purpose) and four (market effect) favored Thomson Reuters, recognizing a potential market for AI training data. The ruling concerned non-generative AI, limiting how far it reaches.

The U.S. Copyright Office published Part 2 of its AI report, addressing when works created with generative AI can be copyrighted — reaffirming that purely AI-generated output without sufficient human authorship is not protectable while human-arranged or -modified works may be.

A federal court approved a stipulation requiring Anthropic to maintain guardrails meant to stop its Claude model from reproducing copyrighted song lyrics, in the suit brought by music publishers Concord, Universal Music Publishing Group and ABKCO over some 500 songs (filed October 2023). It resolved only the injunction question; in March 2025 the court denied the publishers a preliminary injunction against training, finding any harm compensable by damages, and the case continued into discovery.

US District Judge William Orrick denied Stability AI's and Midjourney's motions to dismiss the artists' copyright claims in Andersen v. Stability AI, finding both direct and induced-infringement claims plausible and letting the case proceed to discovery. Trial was later set for September 2026, keeping the image-generation copyright question alive alongside the text cases.

The U.S. Copyright Office published Part 1 of its report "Copyright and Artificial Intelligence," addressing digital replicas and recommending a new federal digital-replica law. It was the first installment of a three-part study that would become central to the US copyright debate — and, with Part 3, to a governance crisis inside the Office itself.

The New York Times sued OpenAI and Microsoft in the Southern District of New York, alleging millions of Times articles were copied without authorization to train ChatGPT and related models that now compete with the Times. The complaint seeks no fixed figure but says defendants should be liable for "billions of dollars" in statutory and actual damages and asks the court to order destruction of models and training data built from Times material. It was the first major American news organization to bring such a suit.

Microsoft announced its Copilot Copyright Commitment: if a third party sues a commercial customer for copyright infringement over the use of Microsoft's Copilots or their output, Microsoft will defend the customer and pay adverse judgments or settlements — provided the customer used the built-in guardrails and did not try to generate infringing material. In November 2023 it expanded the pledge (as the Customer Copyright Commitment) to Azure OpenAI Service outputs. A private-ordering response to legal uncertainty, not law.

Getty Images (US) filed a copyright and trademark complaint against Stability AI in the U.S. District Court for the District of Delaware (No. 1:23-cv-00135), alleging Stability copied more than 12 million Getty photographs — with captions and metadata — to train Stable Diffusion, and that the model reproduces distorted Getty watermarks. Claims were brought under the Copyright Act, the Lanham Act and Delaware unfair-competition law, with copyright-management-information removal alleged and a jury demanded.

Sarah Andersen and other visual artists filed a class-action copyright suit against Stability AI, Midjourney and DeviantArt in the U.S. District Court for the Northern District of California, alleging their works were used without permission to train image-generation models. It was among the first US cases to test whether training generative AI on copyrighted works is infringement.