Anthropic’s piracy chats enter music copyright case

Sony, EMI and Warner Chappell sued Anthropic on August 28, and their complaint tries to get around Anthropic’s main defense. Anthropic says no commercial Claude model was trained on pirated books. The publishers argue those books could still have reached Claude indirectly, through synthetic data and feedback produced by an internal, non-commercial model that was trained on them.

The publishers back the piracy claim with internal messages that first came out in the authors’ case, Ars Technica reports. Their account starts in July 2021, when co-founder Benjamin Mann allegedly used BitTorrent to download and upload millions of pirated books from Library Genesis (LibGen), with CEO Dario Amodei’s approval. Both men are named as individual defendants. After the FBI shut LibGen down, its contents were copied into Z-Library, and that library was later mirrored by the Pirate Library Mirror (PiLiMi). The complaint says Mann told employees to torrent PiLiMi soon after it appeared and wrote that it had dropped “just in time!” One staffer replied: “zlibrary my beloved.”

The publishers searched the public catalogs of LibGen and PiLiMi by title, author and ISBN. They say the results show Anthropic torrented “at least hundreds of books containing sheet music and song lyrics” owned by the three plaintiffs. Among the titles they name are collections of the complete works of the Beatles, Taylor Swift’s “best” songs and “VH1’s 100 Greatest Songs of Rock & Roll.” They expect discovery to show the “full extent” of the downloads.

What counts as training?

The complaint concedes that Anthropic denies using the LibGen and PiLiMi books to train its commercial models. It then argues that the denial depends on how “training” is defined. This is the central passage:

Anthropic has trained at least one of its commercially released Claude models using synthetic data created by a non-commercial AI model that was trained on text derived from LibGen and/or PiLiMi. Anthropic employed at least one non-commercial model that was trained on text derived from LibGen and/or PiLiMi to provide at least one commercial Claude model with reinforced feedback.

The argument has two routes. In the first, a model trained on the pirated books writes training text that a production model then learns from. In the second, that same model scores or corrects a production model’s outputs during fine-tuning. On either route the production model never sees the books themselves, but the publishers say it learns from a model that did.

The question matters because of the settlement in the authors’ case. According to class counsel Susman Godfrey, Anthropic certified in that settlement that it did not use LibGen or PiLiMi material in any commercial model, and it agreed to destroy the original files and every copy made from them. The firm also says the settlement released only claims based on past conduct, gave Anthropic no license for future training, and did not cover any claims based on what the models output.

The publishers point to one more use, drawn from documents recently unsealed in the authors’ case. An Anthropic witness testified that the company stopped training large language models on LibGen but kept using the dataset as part of its guardrails: long passages in Claude’s outputs were checked against it to see whether they matched source text too closely.

Lyrics in the output

The rest of the complaint is about what Claude produces. The publishers say Claude returns lyrics on request without copyright management information, which is the credit and rights data attached to a work, and sometimes includes lyrics when the user did not ask for them. Asked for a song’s chord progression, “the AI model will often generate output” containing “copyrighted lyrics alongside those chords,” the complaint says. It also says Claude mixes real lyrics with generated ones when writing “new” songs. One example it gives is a request to rewrite Eminem lyrics in “Beyonce’s style.”

The publishers argue this behavior was designed in. They allege that Anthropic staff, while testing whether the models could recommend songs similar to someone’s favorites, “repeatedly prompted the AI models for Music Publishers’ copyrighted lyrics.” They also name other sources besides the pirate libraries: lyrics scraped from lyrics sites without permission, and digital copies of “hundreds of songbooks and sheet music collections” made when Anthropic scanned physical books and destroyed them.

The output claims carry the publishers’ market-harm argument, which is where the authors lost. In June 2025, Judge William Alsup ruled that training on lawfully acquired books was transformative fair use, writing that Anthropic’s models trained “not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.” In the same ruling he held that downloading and permanently keeping pirated copies was “inherently, irredeemably infringing.” The publishers say AI-generated songs now on the charts compete directly with the writers whose work trained them. They also cite the US Copyright Office’s view that outputs which compete in the market for a type of work, song lyrics included, can dilute royalty pools even when they are not substantially similar to any one work.

The “legal/practice/business slog”

The settlement figure itself is part of the complaint. ”$1.5 billion is obviously not a large enough settlement to deter infringing conduct by a company that has parlayed such mass infringement into a staggering $2-trillion-dollar valuation,” the publishers wrote. That $1.5 billion paid authors over more than 7 million pirated books. At the time of settlement, the books in the LibGen and PiLiMi datasets carried statutory damages of up to $150,000 per work, which put Anthropic’s theoretical exposure above $70 billion.

The publishers also rely on Amodei’s testimony in the authors’ case. He said Anthropic could have bought the works legally but torrented them to avoid a “legal/practice/business slog.” The complaint adds that Anthropic never approached the publishers about licensing, as its largest rivals have. “Anthropic has never sought nor obtained any license to lawfully exploit Music Publishers’ copyrighted works for any use, let alone for AI training data or output,” it says.

The publishers want an injunction against the training and an order requiring Anthropic to “provide an accounting of the training data, training methods, and known capabilities of Anthropic’s AI models.” They say that disclosure would let rights holders in any medium see which of their works were used and how much those works may have shaped a commercial model. This is not the only publisher suit against Anthropic this month: Round Hill Music filed its own earlier in August.

An Anthropic spokesperson responded: “This is the third lawsuit from the same lawyers, recycling allegations from cases already before the courts. Training generative AI models is a transformative fair use—as the court held in Bartz—and we will defend ourselves robustly.”