On the surface, the Friday lawsuit that Sony Music Publishing, Warner Chappell, and a group of other major publishers filed against Anthropic reads like the latest entry in an already crowded genre: music companies going after an AI lab over copyrighted lyrics. Read closer, though, and it is actually something sharper. The plaintiffs are not just arguing that Anthropic trained Claude on their songs. They are arguing, explicitly and in detail, that the company got that training data through piracy.
That distinction is not a legal technicality. It is the exact theory that just won the music and publishing industry its biggest AI settlement to date.
What the Publishers Are Alleging
Filed in the U.S. District Court for the Northern District of California, the complaint names Anthropic and co-founders Dario Amodei and Benjamin Mann as defendants. The publishers accuse the lab of a “brazen campaign of illegally torrenting, scraping, and downloading copyrighted works on a massive scale” to develop, operate, and profit from the Claude model family.
The claimed damages are not small. The companies are seeking up to $150,000 per infringed work for “tens of thousands” of compositions, plus up to $25,000 for each instance where identifiable copyright data was stripped. If a court sides with them and awards the maximum, the total could run into the billions.
The complaint gets specific about how the training material was sourced. It alleges that Benjamin Mann personally used BitTorrent to download over five million pirated books, and that Anthropic employees downloaded at least two million more from Pirate Library Mirror, an archive of pirated books. Among those millions of books, the publishers say, were volumes containing the lyrics and sheet music to dozens of their copyrighted songs.
Why the Piracy Angle Matters Now
This is where the story gets interesting for anyone who has been watching AI copyright litigation. In the earlier round of cases, the industry’s strongest hand kept being the piracy claim rather than the fair-use claim.
In the Bartz case, a judge ruled that it was legal for Anthropic to use copyrighted works to train its models in the first place — but that it was not legal to acquire that content through piracy. That nuance is what produced the landmark $1.5 billion settlement Anthropic agreed to last year. Put simply: feeding someone’s book into a training run can be fair use, but torrenting millions of pirated copies of that book to do it is not.
This new lawsuit leans hard on that proven theory. The plaintiffs explicitly ask a jury to hold Anthropic accountable for how the data was obtained, not just for what the model does with it. The phrase “one of the largest and most blatant ongoing thefts of intellectual property in history” is the tone the whole complaint is written in.
Beyond the Books: Scraped Lyrics
The complaint is not only about pirated books. It also claims Anthropic scraped lyrics from licensed lyrics platforms like MusixMatch and LyricFind — sites that pay the publishers for the right to display those songs. From the publishers’ perspective, that is a double injury: the lyrics were taken both from pirate archives and from licensed services that were already paying for access.
A few of the named songs give the story a human face: Marvin Gaye and Tammi Terrell’s “Ain’t No Mountain High Enough,” Bon Jovi’s “Livin’ On a Prayer,” Earth, Wind & Fire’s “September,” Leonard Cohen’s “Hallelujah,” and Taylor Swift’s “Paper Rings.” These are songs most people have heard at a wedding, a karaoke bar, or a radio in a jeepney. They are not obscure catalog entries.
Anthropic, for its part, pushed back in a statement: “We disagree with the publishers’ claims and we intend to defend ourselves robustly in court.” The company has not denied the underlying conduct in detail, instead framing the case as one it expects to win on the merits.
What This Means for Developers and Creators
If you build on LLMs, this case is worth watching for reasons that go beyond the music industry. The rising legal pressure around training-data provenance is starting to reshape how models get made and how they are licensed. The stakes land right on top of the same arms race I looked at when Anthropic locked in its $45 billion compute deal.
Every downstream developer who shares or fine-tunes a model inherits whatever risk sits in its training data. A model trained on pirated works, even unknowingly, becomes a liability the moment someone elsewhere is sued. That is part of the reason the market is seeing companies race to strike licensing deals and to keep open-weight models commercially clean.
For a smaller shop or a solo dev, the practical takeaway is simpler. Before you pick a foundation model or a training corpus, ask where the data came from. The license you read at the top of a checkpoint matters less than the chain of custody underneath it. This is the same scrutiny I would apply to any dependency before I pull it into a project. It is also part of the reason broader trust in AI keeps slipping even as usage climbs — the hidden provenance questions never fully go away.
The Bigger Picture
None of this happens in a vacuum. Anthropic has been in and out of court over AI copyright for years, and the same thorny questions about how creative work gets absorbed into models have been dogging the space — I wrote about the pressure on Hollywood to come clean about AI training a while back. The new suit arrives barely a week after Round Hill Music sued Anthropic and Suno over allegedly scraped music, and months after the earlier publisher action sought $3 billion over roughly 20,000 works. The same lawyers keep showing up, and they keep carrying more evidence each time.
That compounding pattern matters. One settlement can be dismissed as a one-off. A string of cases built on the same piracy theory, from the same plaintiffs’ bar, starts to look like a coordinated strategy — and it forces the whole industry to take data provenance more seriously.
I tend to watch these fights with a certain sympathy for both sides. I genuinely like a lot of what large language models unlock, and I have written before about how much of the industry’s excitement rests on borrowed creative work. But I also run a small operation myself, and I know the difference between being inspired by something and copying the artist’s work in bulk. The law is trying to find that line, and so far the piracy theory is the clearest point on it.
If you are building anything with these tools, do not treat the copyright headlines as background noise. This is the ground floor of the rules that will govern AI development for the next decade — and the music publishers just showed they are willing to fight on the steepest possible ground. It fits a wider pattern I keep coming back to: the entire AI buildout is being financed on borrowed money and borrowed content, and someone is eventually going to call in both notes.