Anthropic's piracy chats enter music copyright case

Music publishers Sony, EMI, and Warner Chappell have sued Anthropic, alleging that its torrenting of pirated books included songbooks and sheet music and helped Claude reproduce copyrighted lyrics while AI-generated songs compete with human writers.

The complaint, reported by Ars Technica, argues that Anthropic’s $1.5 billion settlement with book authors was too small to deter further infringement.

At a glance

• Internal messages from 2021 show Benjamin Mann urged employees to torrent PiLiMi, while a colleague answered “zlibrary my beloved,” according to the publishers’ complaint in court.
• Metadata allegedly indicates that Anthropic torrented at least hundreds of books containing sheet music and lyrics owned by Sony, EMI, and Warner Chappell.
• Anthropic says the lawyers are recycling allegations from earlier litigation and maintains that training generative AI models qualifies as transformative fair use under Bartz.

The dispute places training-data practices alongside a separate market question: whether AI-generated music can substitute for human songwriting even when outputs are not substantially similar to one particular work. That distinction may give music rightsholders a stronger basis for proving harm than book authors had in their case, while raising difficult questions about how courts measure competition from synthetic songs.

Anthropic’s torrenting allegedly began in July 2021 and moved through two pirate-library copies

According to the complaint, co-founder Benjamin Mann personally used BitTorrent in July 2021 to download and upload millions of pirated books from Library Genesis, also known as LibGen. Anthropic co-founder and CEO Dario Amodei allegedly approved the torrenting, and both executives are named individually as defendants in the publishers’ lawsuit.

The FBI shut down LibGen by the end of 2021, the complaint says, but pirates had already copied its contents into Z-Library. After Z-Library was also shut down, Anthropic allegedly obtained another copy through the Pirate Library Mirror, or PiLiMi. Mann then directed employees to torrent the mirror soon after it appeared, describing its timing as “just in time!”

The complaint says Anthropic torrented at least hundreds of books containing publishers’ music

Music publishers say catalogs from LibGen and PiLiMi exposed bibliographic metadata, including titles, authors, and ISBNs. They claim that this information shows Anthropic torrented at least hundreds of books containing sheet music and song lyrics from compositions controlled by the plaintiffs. The publishers expect discovery to reveal the full extent of the activity.

The complaint also alleges that Anthropic destroyed physical books to create unauthorized digital copies of hundreds of songbooks and sheet-music collections. Publishers say the company used other unauthorized sources as well, including pirated material scraped from lyrics websites. Anthropic denies using the torrented books to train commercial Claude models, but the publishers argue that definition of training may omit synthetic-data and feedback stages.

Anthropic’s own litigation record links LibGen material to model guardrails and feedback

The publishers’ complaint says Anthropic trained at least one noncommercial model on text derived from LibGen or PiLiMi, then used that model’s synthetic data or behavioral feedback with at least one commercial Claude model. Recently unsealed filings from the authors’ case also allegedly show continued use of the LibGen dataset to test whether outputs closely matched source text after direct training on it stopped.

Publishers say Claude can return copyrighted lyrics without copyright-management information, sometimes even when a user requests only a song’s chord progression. They allege the models can combine actual lyrics with generated lines, imitate the “heart” of popular works, and produce lyrics after employees repeatedly prompted them while testing recommendations for comparable songs.

Anthropic’s spokesperson called the lawsuit the third action from the same lawyers and said it recycles allegations already before the courts. The company maintains that training generative AI models is transformative fair use, citing the ruling in Bartz. The publishers respond that the ruling turned on book authors’ inability to establish market substitution or other harm in their market.

Прозрачность обучения в суде The publishers seek an injunction against continued use of the disputed material and an accounting of Anthropic’s training data, methods, and known model capabilities. They also argue that Anthropic never obtained licenses for their works, unlike some major AI rivals. The court has yet to determine whether the alleged use of music works creates market harm under the reasoning applied to books.