Anthropic’s third music suit names two founders, not just the company
Sony Music and Warner Chappell allege Ben Mann personally torrented five million books to train Claude.
By The Signal · · 5 min read
A federal complaint filed late Friday in the US District Court for the Northern District of California accuses Anthropic co-founder Benjamin Mann of using BitTorrent to download more than five million pirated books himself, and says other employees pulled at least two million more from a site called Pirate Library Mirror. Sony Music Publishing, Warner Chappell and a group of other music publishers are suing Anthropic, Mann and CEO Dario Amodei personally, seeking up to $150,000 per infringed work across "tens of thousands" of songs.
The number is $150,000, multiplied by a count nobody has published yet
Statutory copyright damages in the US cap at $150,000 per work for willful infringement, plus a separate $25,000 penalty for each instance where copyright management information — the metadata identifying an author or rights holder — was stripped out. The publishers haven't named an exact tally of infringed songs in public filings; they describe "tens of thousands" of works, which puts the theoretical ceiling at several billion dollars if a court finds willful infringement across the full set and awards the maximum on every count. That's the number driving headlines. It is also, almost certainly, not the number that gets paid.
The comparison case is instructive. In the Bartz v. Anthropic litigation, a group of book authors made a similar piracy argument, and a judge ruled last year that training an AI model on copyrighted text is legal, but obtaining that text through piracy is not. Anthropic settled for $1.5 billion — a fraction of the theoretical statutory maximum implied by the number of works involved, arrived at through negotiation rather than a jury applying the top per-work penalty to every count. Settlements track leverage and litigation risk, not the ceiling written into the statute. Sony and Warner Chappell's filing reads like an opening bid calibrated to that same eventual negotiation, not a forecast of the final check.
Why this suit names two people, not just one company
What separates this filing from the book-publisher cases is the specificity aimed at individuals. The complaint alleges Mann personally ran torrent downloads and names Amodei as a co-defendant, a structural choice that raises the stakes beyond a corporate settlement: naming founders personally is typically meant to signal that the plaintiffs consider the piracy not an operational accident but a decision made at the top. The suit also claims Anthropic scraped song lyrics directly from sites like MusixMatch and LyricFind — services that pay the labels to license that same content, meaning Anthropic allegedly took for free what its own training-data suppliers were paying for.
The specific songs named — Marvin Gaye and Tammi Terrell's "Ain't No Mountain High Enough," Bon Jovi's "Livin' on a Prayer," Leonard Cohen's "Hallelujah," Taylor Swift's "Paper Rings" — function less as evidence and more as a jury-facing device: recognizable songs make an abstract training-data pipeline argument legible to people who have never read a model card. It's the same move as pointing to a specific pallet in a warehouse fraud case rather than describing "inventory discrepancies."
The freight-yard problem: how training data actually gets sourced
Training a large language model on text requires acquiring that text at a scale no single licensing deal covers — hundreds of millions of documents, sourced faster than any negotiated agreement can move. The industry's standard practice has been to build these datasets from web scrapes and bulk archives first, and negotiate licenses later, if at all. It's the equivalent of a shipping line loading a container from a dock warehouse and sorting out the bill of lading afterward — except copyright law doesn't treat "we'll pay eventually" as a defense to unauthorized copying. BitTorrent and shadow libraries like the one named in this suit are simply the fastest freight route to volume, which is exactly why they keep surfacing in these complaints: Bartz named similar shadow-library sourcing, and the January suit brought by Concord Music Group and Universal Music Group — filed by some of the same lawyers behind this one — makes overlapping claims.
The counter-argument Anthropic has already won once
The strongest defense available to Anthropic is the one a judge already accepted in Bartz: training an AI model on copyrighted material can itself be lawful, protected as a transformative use under fair-use doctrine, provided the underlying copies were obtained legitimately. That ruling gives Anthropic a real argument to make about the music-lyrics claim — that using song lyrics as training signal, on its own, isn't necessarily the violation. But the same ruling is what makes the acquisition method the entire case. If a court applies that same logic here, the legal fight won't be about whether Claude can learn from song lyrics; it will be about whether Mann's BitTorrent downloads and the LyricFind-adjacent scraping constitute the same kind of illegitimate acquisition that cost Anthropic $1.5 billion the last time. Anthropic has not yet responded to the complaint.
What the filing does next: it joins a queue, not a courtroom
This is not an isolated case; it is now one filing in a stack. Anthropic already faces suits from Universal Music Group, Concord, ABKCO, BMG and Round Hill Music, on top of the settled Bartz matter. Courts handling overlapping claims about the same defendant and similar conduct often consolidate discovery or coordinate scheduling, which means the piracy allegations detailed here — the torrent logs, the Pirate Library Mirror downloads, the LyricFind comparison — may end up as evidence other plaintiffs cite rather than facts this case alone resolves. The constraint nobody covering the headline number mentions: statutory damages caps exist per work, but courts have wide discretion below that ceiling, and the actual dollar figure this produces will depend on how many of "tens of thousands" of works a judge or jury finds were willfully and provably pirated — a number that requires forensic access to Anthropic's training logs nobody outside the company has yet seen.