Thirty-five music publishers, including affiliates of Sony Music Publishing and Warner Chappell Music, filed suit against Anthropic, CEO Dario Amodei, and co-founder Benjamin Mann in the U.S. District Court for the Northern District of California on August 28, 2026. The complaint alleges that in developing Claude, Anthropic acquired and reproduced tens of thousands of musical works without authorization and generated their lyrics as outputs. The starting point is the acquisition of roughly 5 million books from LibGen and roughly 2 million books from PiLiMi, facts that came to light in a separate authors' lawsuit. 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. To make sense of this case, it's necessary to separate four distinct acts: "acquiring pirated material," "training the model," "Claude outputting lyrics," and "stripping copyright notices." These four acts will not necessarily receive the same legal treatment.
What the 35 Publishers Are Fighting Over Isn't Sound Recordings—It's Musical Works
The rights the plaintiffs in this case are asserting do not concern the recorded sound tracks that record labels manage. The U.S. Copyright Office defines a musical work as a composition consisting of music and any accompanying lyrics, treating it as a distinct category of work separate from a sound recording, which fixes a performer's rendition. Consequently, even for a text-based system like Claude, a claim can be made that it acquired books containing lyrics and sheet music and reproduced those lyrics.
The 48-page complaint lists 24 Sony-affiliated companies and 11 Warner Chappell-affiliated companies as plaintiffs. The defendants are Anthropic, Dario Amodei, and Benjamin Mann. The claims break down into four counts. The first alleges direct infringement by all three defendants via BitTorrent. The second alleges contributory infringement by Amodei and Mann. The third alleges direct infringement by Anthropic spanning scraping through training and output. The fourth challenges the removal or alteration of copyright management information (CMI).
CMI includes elements such as the work's title, author's name, and rights holder's name. The plaintiffs allege that Anthropic stripped this information during data preprocessing, treating it as "unnecessary boilerplate," and that its outputs likewise reproduced lyrics without any source attribution. However, beyond the fact that information was dropped during preprocessing, the case will also need to examine whether Anthropic knew this would facilitate or conceal infringement. The mere fact that names weren't included in outputs does not by itself settle the fourth claim.
The Acquisition of 7 Million Books Drew In a Different Set of Rights Holders
The central fact underlying this complaint traces back to the June 2025 summary judgment in Bartz v. Anthropic. According to that ruling, Mann acquired at least roughly 5 million books from LibGen via BitTorrent in June 2021. In July 2022, an Anthropic employee acquired at least roughly 2 million additional books from PiLiMi. Because BitTorrent transmits file fragments being received to other users, the publishers argue that both reproduction and distribution rights were infringed.
The Bartz ruling drew a distinction between two types of conduct. Copies made to train Claude and its predecessor models were found highly transformative, and the court ruled that this constituted fair use for the books at issue in that case. The court also upheld as fair use the practice of taking legitimately purchased physical books, cutting and scanning them page by page, discarding the paper, and replacing them with digital versions—characterizing this as a format conversion for preservation and search purposes. On the other hand, the court found that storing more than 7 million pirated books in a permanent, centralized library available for future uses was not justified as fair use.
In the class action concerning this piracy-based acquisition, a $1.5 billion settlement received final approval on July 20, 2026. The court's order covers 482,460 works, recording payments at an estimated roughly $3,000 per work. However, this is a settlement concerning book rights—it is not a ruling that established liability for the musical works at issue in the current case. When a single book contains both novel text and lyrics or sheet music, the acquisition of that one book can give rise to separate claims from different rights holders. The 35 publishers in this case have stepped into that gap.
How Far the Bartz Fair-Use Ruling Reaches
Lining up the two cases by category of conduct reveals why Anthropic's likely defense—that "training is fair use"—would not, by itself, answer the complaint as a whole.
| Layer of Conduct | Sony/Warner Publishers' Claim | Scope Addressed in Bartz |
|---|---|---|
| Acquisition from LibGen/PiLiMi | Reproduced and distributed books containing musical works via torrenting | Permanent pirated-material library not justified as fair use |
| Model training | Reproduced tens of thousands of musical works during the training process | Training copies of the books in that case were fair use |
| Claude's output | Generated verbatim or near-verbatim copies or derivatives of lyrics | Bartz plaintiffs did not allege infringing output |
| CMI removal | Removed or altered titles and author names during preprocessing and output | Bartz's fair-use finding did not resolve this claim |
What this comparison establishes with certainty is only that acquisition and training cannot be lumped into the same category. The plaintiffs here have added output and CMI on top of that. The complaint alleges that guardrails introduced after previous music publisher lawsuits can still be bypassed with prompts that re-request specific lyrics, citing records from another lawsuit. Exactly which model, at what point in time, and with what input produced how much of any given lyric will need to be established through evidence going forward.
What the Bartz ruling found to be fair use was the training copies at issue in that specific case—it reached no conclusion regarding pirated acquisition, the lyrics output now being disputed, or CMI removal.
The plaintiffs also argue that outputs generating "new lyrics" or writing "in the style" of a particular songwriter constitute market substitution. However, producing text with a similar feel to an existing work is a different matter from reproducing the protected expression of a specific work. This will require comparing verbatim or near-verbatim outputs against records of which works were actually used in training.
Why the Maximum $150,000 Figure Shouldn't Be Read as a Total Claim
The remedies the complaint explicitly seeks are: up to $150,000 per work, up to $25,000 per CMI violation, injunctive relief, an accounting, and destruction of infringing copies—it does not specify a total dollar amount or demand full retraining. Multiplying the statutory maximum by the plaintiffs' claim of tens of thousands of works would yield an enormous figure. But the complaint itself states no aggregate damages amount. Whether each work is registered and whether the plaintiffs actually hold rights to it, which specific acts constitute infringement, and how many instances of willfulness or CMI violations will be found—none of this is yet known. At this stage, presenting a simple multiplication as the claimed total would be premature.
The non-monetary demands are more concrete. Beyond a permanent injunction to halt the infringement, the plaintiffs are asking Anthropic to report on the training data it used, its collection and processing methods, and the model's known capabilities. They further demand that infringing copies be destroyed under court supervision, with a sworn report on the steps taken.
There's a boundary here too. The complaint explicitly calls for destruction of infringing copies, but it does not explicitly demand deletion of the Claude model itself or an order to retrain everything from scratch. The scope of the permanent injunction could change as proceedings continue, but there is currently no basis for characterizing this as "a lawsuit demanding complete retraining."
Three Types of Evidence That Will Decide the Outcome
Three bodies of evidence will be decisive in the proceedings ahead. First, which acquisition channel each song listed in Exhibits A and B came through, and which of the 35 companies holds which rights. Second, under what input conditions and to what extent Claude reproduced the protected lyrics of these works. Third, whether Anthropic's process of removing CMI can be tied to an awareness that doing so would facilitate or conceal infringement.
Anthropic told Axios that it "disagrees with the publishers' claims and will vigorously contest them in court." A detailed answer to the complaint has not yet been made public. Whether the company will dispute the acquisition pathways and the presence of the works in question, broadly assert the training fair-use finding from Bartz, or demonstrate that its current guardrails already suppress such outputs—the defense it chooses will determine how much of the four claims remains standing.
The figure of 7 million books is striking, but it does not by itself guarantee victory for the 35 publishers in this case. What will ultimately decide the outcome is a mapping showing that each specific song existed within the pirated library, a record demonstrating that protected expression emerged from the trained Claude, and internal evidence showing intent behind the CMI removal. If all three come together, a settlement that was supposed to end with books could extend into a separate reckoning over musical work copyrights.