Thirty-five music publishers sued Anthropic on August 28, 2026. It isn't the largest AI copyright suit by dollar value, but it may be the most consequential so far, because it targets the very issue the $1.5 billion book settlement left untouched.
Just five weeks after a federal judge signed off on the largest copyright settlement in history, the music industry arrived at the same courthouse with a new filing.
The Bartz v. Anthropic settlement, approved on July 20, 2026, paid roughly $3,000 per work across nearly 482,460 pirated books and closed the most visible chapter of the AI copyright story. What it didn't answer is whether AI models infringe copyright when their outputs reproduce protected material. The settlement expressly excluded that question, and it now lies at the heart of a new 48-page complaint from Sony Music Publishing, Warner Chappell Music, and 33 affiliated publishers.
This case differs in structure, in the rights at stake, and in what a court will be forced to decide. Every AI company that runs a general-purpose language model is watching closely.
A note on the numbers: The widely cited $1.5 billion figure comes from the Bartz book settlement. It is not a demand made in the music publishers' lawsuit. The Sony/Warner complaint asks for statutory damages of up to $150,000 for each infringed composition under the Copyright Act, which is the ceiling for willful infringement. Spread across tens of thousands of named works, that exposure goes well beyond the book settlement. These are allegations and statutory maximums, not amounts any court has awarded.
What happened
On August 28, 2026, 35 music publishing companies, led by Sony Music Publishing and Warner Chappell Music, filed a copyright complaint against Anthropic in the U.S. District Court for the Northern District of California. Alongside the company, the suit names CEO Dario Amodei and co-founder Benjamin Mann as individual defendants.
The filing runs to 48 pages, and its language is far from restrained. It accuses Anthropic of "a brazen campaign of illegally torrenting, scraping, and downloading copyrighted works on a massive scale in order to develop, operate, and reap enormous profits from Anthropic's Claude series of artificial intelligence models." The named works number in the tens of thousands of musical compositions, among them "Ain't No Mountain High Enough" and "All I Want for Christmas Is You."
The plaintiffs are asking for a jury trial and statutory damages, which can reach $150,000 per work for willful infringement under the Copyright Act.
Who is involved
This is no simple two-party dispute. The 35 publishers together hold the catalogues of Sony Music Publishing, Warner Chappell Music, and dozens of affiliated labels. With this filing, the publishing arms of all three major music companies are now in litigation against Anthropic, over a body of works considerably larger than in any previous music case. Universal Music Publishing Group, Concord Music Group, and ABKCO had earlier sued Anthropic in October 2023 over 499 works, and that case was still unresolved when this bigger complaint arrived.
On the defense side, the named parties include Amodei and Mann personally, not only Anthropic as a company. Naming individual founders is unusual, and it raises the stakes beyond what a corporate indemnification structure can easily absorb.
What Anthropic is accused of
The complaint makes three separate accusations. It helps to take them one at a time, because each carries different legal weight and a different set of possible remedies.
The first is about acquisition. The publishers allege that Anthropic built its training data through Library Genesis and the Pirate Library Mirror, the same shadow libraries at the center of the Bartz book case. According to the complaint, Benjamin Mann used BitTorrent in June 2021 to obtain at least five million books from LibGen, and in July 2022 Anthropic employees downloaded about two million more from PiLiMi. Those pirated books contained song lyrics, sheet music, and musical compositions. The publishers contend that obtaining a single pirated book containing protected lyrics infringes both the book rights and the separate music publishing rights.
The second is about outputs. The complaint alleges that Claude can reproduce copyrighted lyrics word for word, or nearly so, when prompted, and that it can write derivative or competing lyrics based on protected works. This is the allegation the Bartz settlement most conspicuously left open.
The third is about copyright management information. A separate claim accuses Anthropic of removing attribution and copyright details from works that went into its datasets. That conduct is a violation under the Digital Millennium Copyright Act, independent of any infringement claim.
The publishers also try to head off the defense they expect Anthropic to raise, that the lyrics are "freely available online." "Anthropic is wrong to claim that Music Publishers' lyrics are 'freely' available on the internet," the filing states, pointing out that lyric websites themselves pay licensing fees to display those lyrics. In music copyright law, something being freely visible online has never amounted to a license to reproduce it.
Why song lyrics create a uniquely difficult problem for AI
Books and lyrics are both text, but copyright law doesn't treat them the same way, and the differences don't favor AI companies.
Lyrics are short, and that is exactly the problem. Copyright protection doesn't depend on length, so a two-line chorus from a hit song has the same legal weight as a 300-page novel. When an AI model reproduces it word for word, there's no "small portion of a large work" defense to retreat to, because the entire protected expression can fit inside one response.
Music copyright is also more fragmented than book copyright. A single song can contain several distinct copyrighted components, such as the lyrics, the sound recording, and the musical composition, and these can be divided among different holders, including the artist, the publisher, and the record label. One act of training on or reproducing a popular song can therefore trigger claims from several rights holders at once. That complexity means AI companies face compounding exposure even when they try to compensate just one party.
Music also raises a market substitution argument of its own. The complaint alleges that Claude can generate "derivative or competing lyrics" based on protected works. That matters because publishers earn licensing revenue whenever lyrics are reproduced: in greeting cards, films, advertisements, and app integrations. A model that produces convincing imitations of protected song styles, or near-verbatim lyrics on request, competes directly with that licensing market in a way a model answering factual questions does not.
The copyright questions at the center of the dispute
Bartz settled one version of the AI copyright question and left the harder ones open. To read the music case correctly, you need to know what the earlier case resolved and what it didn't.
In June 2025, Judge William Alsup, who was then presiding, issued a split ruling. He found that training generative AI models on lawfully acquired books could count as transformative fair use: "Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them, but to turn a hard corner and create something different." At the same time, he held that acquiring pirated copies was not fair use. "Anthropic had no entitlement to use pirated copies for a central library," he wrote. The settlement that followed compensated rights holders for the piracy, not for the training.
That ruling left three questions unanswered, and the Sony/Warner case raises each of them directly:
Whether piracy-based acquisition of music rights, as distinct from book rights, is a separate act of infringement. The settlement covered book copyright holders. Music publishers were outside that class, and their claims were never released.
Whether AI output that reproduces copyrighted text verbatim is infringement. Bartz explicitly left output claims out of its scope, so this remains an open legal question in every AI copyright case.
Whether Alsup's fair-use finding applies to music at all. It was a district court ruling that was never appealed to the Ninth Circuit. It sets no binding precedent beyond that court, and the music case is before a different judge.
"Conduct that was once settled for $1.5 billion is now being pursued again by different plaintiffs who hold rights to the lyrics and sheet music embedded within those books. However, what has been filed so far is only the plaintiffs' complaint. There is no detailed rebuttal from Anthropic yet, nor any factual findings from the court."
XenoSpectrum, analysis of the August 28 complaint, September 2026
What this means for AI model companies
The most immediate consequences are financial and reputational. Anthropic settled the largest copyright case in history in July, then was sued again before August ended. The publishers make the point sharply in their complaint: "Anthropic clearly considers that to be just the cost of doing business given that its entire business model continues to be built on copyright theft."
The deeper consequence is structural. The Bartz settlement resolved claims from one category of rights holder (authors and book publishers) over one type of content (books) from one source of acquisition (LibGen and PiLiMi). Those same pirated libraries also held music, academic papers, poetry, scripts, and other kinds of content. Each category has its own rights holders, and none of them were released by the Bartz settlement. The music publishers' filing is likely the first in a series of claims from different industries asserting rights over different content types embedded in the same pirated datasets.
The pattern holds for any AI company that trained on web-scraped or pirated data: settling with one class of rights holders doesn't shut out claims from the rest. And because the output liability question is still unresolved, every model capable of reproducing copyrighted text in its responses carries ongoing exposure, no matter how its training data was first obtained.
What this means for developers and AI startups
For developers building on AI APIs, output liability is the most pressing issue. Suppose courts eventually decide that a model's verbatim or near-verbatim reproduction of copyrighted lyrics is infringement. Who then carries that liability, the model developer or the application developer? That is not settled law. The terms of service between API providers and their customers will matter, as will indemnification clauses and the type of application being built.
Developers whose applications let users prompt models to reproduce song lyrics, generate music-adjacent content, or write in the style of specific artists face real uncertainty. The safest approach for now is the one Anthropic already agreed to in January 2025: keeping guardrails that block verbatim lyric reproduction in outputs. That agreement came during the earlier UMG/Concord case, and it shows the output reproduction allegation is far from theoretical. Courts have taken it seriously enough to require technical controls.
For AI startups choosing training data, the clearest practical lesson from the Bartz ruling still stands: legally acquired data is defensible, while pirated data is not, whatever the pirated copy was later used for. That distinction should guide every data acquisition decision, not as a moral stance but as a legal position a judge has already put in writing.
What creators and publishers should watch
If the output reproduction claim succeeds, it would produce the first affirmative ruling that AI models infringe copyright by generating protected text in their responses. Such a precedent would affect every AI system that produces text touching any copyrighted work.
For music publishers in particular, the case also pushes the copyright management information claim. If it's proven, stripping attribution from protected works before training on them exposes AI companies to DMCA liability separate from copyright infringement liability. Rights holders in other content categories will be watching to see whether that claim survives long enough to reach a jury.
Possible outcomes
Judging by how Bartz ended, the likeliest outcome is a negotiated settlement before any appellate precedent is set. A jury trial on statutory damages across tens of thousands of musical compositions would expose Anthropic to extraordinary financial risk, and that kind of exposure usually creates settlement pressure on both sides. Music publishers want compensation and licensing frameworks, not a trial that might yield a pro-AI fair-use ruling. Anthropic wants its fair-use reasoning on training to hold, and has good reason to avoid a jury calculating willful damages song by song.
A second possibility is a licensing deal that ends the dispute commercially rather than legally. Several AI companies have already signed licensing agreements with publishers and news organizations. Both sides could, in theory, design a licensing framework for AI use of music catalogues, and the pressure of litigation speeds those conversations along.
A third possibility, less likely but more significant, is that the case reaches the Ninth Circuit and produces binding appellate precedent on one or more of the open questions. That would affect every AI company operating in the circuit, which includes most of the major ones. And if other circuits ruled differently, it would give the Supreme Court a national split to resolve.
The broader implication for the AI industry
The music publishers' complaint lands at a particular point in AI legal history. The first major copyright settlement is complete. The fair-use ruling that came before it is a district-court decision Anthropic chose not to appeal, so it has no national force. The questions the settlement left open, output reproduction and whether rights can be separated across content types embedded in the same pirated libraries, are now live in court.
The settlement also set a price signal, though unintentionally: roughly $3,000 per pirated book. Music publishers are reading that figure against the $150,000 statutory ceiling per composition for willful infringement, and arguing that the book settlement was too cheap to change behavior. Whether a court accepts that framing will shape how the next round of settlements is negotiated, and not only for music but for every content category waiting in line.
"The most consequential thing the Bartz ruling settled may be what it did not settle. Because Anthropic chose to settle rather than let the case proceed to an appeals court, that ruling will never become binding precedent. Other federal courts are entirely free to reach the opposite conclusion on whether AI training constitutes transformative fair use."
TechTimes, analysis of the Bartz settlement, July 22, 2026
The takeaway
At bottom, this case isn't about song lyrics. It's about whether the legal framework the AI industry built around fair use and transformative training survives contact with a new content type, a new group of rights holders, and a new allegation focused on what AI models do in their outputs rather than how they were trained.
The $1.5 billion book settlement was genuinely historic, but it was a settlement, not a verdict. It left behind a single district-court fair-use opinion that remains unresolved at any binding level. Now a different judge is being asked to examine the same pirated libraries, the same training methods, and the same model, this time for a content type whose brevity, structural complexity, and robust licensing market make fair use arguments harder to build.
Every AI firm watching this case is watching it for the same reason: not because they pirated song lyrics, but because the output reproduction question, if resolved against Anthropic, has no clean limiting principle. A model that can't be prompted to reproduce copyrighted text without liability exposure is a different product than the one most of the industry has been selling.
Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.
I do freelance scraping and data automation projects. Feel free to contact me on LinkedIn or email at hi@kawsarlog.com
Top comments (0)