Unsealed Anthropic staff chats ("zlibrary my beloved") revealed in music publisher suit
Internal messages and allegations of personal torrenting by co-founders are cited as evidence that Anthropic systematically laundered pirated books and songbooks to train Claude.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Major music publishers, including Sony and Warner Chappell, unsealed internal Anthropic Slack messages as 'smoking gun' evidence in their copyright lawsuit against the AI lab, according to Ars Technica.
- The complaint alleges co-founder Benjamin Mann personally used BitTorrent to download pirated books from LibGen with CEO Dario Amodei's approval, and staff chats praised mirrors of pirate site Z-Library.
- Publishers argue Anthropic utilized a 'data laundering' loop, training non-commercial models on pirated data to generate synthetic data for commercial training, and reproducing lyrics verbatim.
A copyright lawsuit filed by major music publishers — including Sony, EMI, and Warner Chappell — against Anthropic has unsealed internal Slack messages and personal data logs as evidence that the AI lab systematically pirated copyrighted works, Ars Technica reported on August 31. The complaint alleges Anthropic built its valuation through a 'brazen campaign' of illegally torrenting and scraping musical compositions, songbooks, and sheet music.
The unsealed messages include internal exchanges where an Anthropic employee wrote 'zlibrary my beloved' after a mirror of the blocked pirate site Z-Library went live, and co-founder Benjamin Mann commented that a mirror dropped "just in time!" The publishers also accuse Mann of personally using BitTorrent to download millions of pirated books from Library Genesis (LibGen), an activity they allege was approved by CEO Dario Amodei. Both founders are named individually as defendants.
While Anthropic has previously denied training its commercial Claude models on pirated books, the lawsuit alleges a 'synthetic data laundering' pipeline. The publishers claim Anthropic trained non-commercial models on pirated datasets, then used those systems to generate synthetic training data or provide reinforcement feedback to optimize the commercial Claude models. The suit also documents instances of Claude reproducing lyrics verbatim when asked for chord progressions, and mimicking the styles of artists like Taylor Swift and Eminem.
A spokesperson for Anthropic dismissed the lawsuit, telling Ars Technica that it 'recycles allegations from cases already before the courts' and that training generative models constitutes transformative fair use. The lawsuit follows a historic $1.5 billion settlement Anthropic paid to book authors earlier in 2026, which the music publishers argue was not a large enough deterrent.
If a court accepts the publishers' argument that training non-commercial models on pirated data to generate 'clean' synthetic training data constitutes systematic copyright infringement, the legal firewall around synthetic data will crumble. Furthermore, naming co-founders personally as defendants on BitTorrent logging evidence sets a severe precedent of individual liability for AI training practices, moving the debate from corporate policy to personal legal jeopardy.