Sony Music and Warner have sued Anthropic over what they describe as a "brazen campaign" of intellectual property theft, alleging the company engaged in "outright theft" of material including copyrighted musical works. Music Business Worldwide reported the filing first. It is not the first intellectual property claim against Anthropic, and it arrives carried by a legal team that has already won against the company once.
Some of the lawyers representing the plaintiffs in the new case also act for Concord Music Group and Universal Music Group in a case brought in January. The same lawyers ran Bartz v. Anthropic, in which a group of authors accused the company of using protected works to train products such as Claude. In Bartz, the court ordered Anthropic to pay $1.5 billion.
The reasoning behind that award is the thing to hold onto, because it shapes everything that has followed. The judge found that using copyrighted works to train a model was permitted; obtaining those works from pirate sources was not. The liability attached to acquisition, not to training. That split turned a contested question about machine learning into a much older and much simpler question about how a company got its hands on a library.
The new suit largely repeats the arguments of the earlier cases while covering a broader set of claims. The publishers accuse Anthropic of "outright piracy," alleging that the company illegally downloaded millions of copies of books via torrents, including editions containing song lyrics and sheet music.
Read the shape of that claim rather than its volume. Two record companies are suing an AI lab over books. Not over what Claude sings back to a user, not over a model producing a chorus on demand, but over the contents of a scraped library. The complaint is engineered to land on the acquisition side of the line the Bartz judge drew, which is the side where the money is. That is a sensible strategy and a narrow one, and it says something about how music rightsholders now rate their chances on the output side of the argument.
The arithmetic is what makes this serious rather than symbolic. $1.5 billion in Bartz came out of a per-work calculation applied across a pirated corpus. A songbook or a lyrics edition sitting in the same torrented pile is another work, with another owner, in a catalogue that is very large and very well documented. Sony Music and Warner have spent decades building the registries that make per-work counting straightforward. The publishing industry's institutional advantage in this fight is not legal theory; it is metadata.
Notably absent from the account of the filing is any number attached to what the publishers want, and any allegation about what the model actually produces. The case as reported is entirely about intake. If that continues to be where courts assign liability, the practical outcome is not that training on copyrighted material stops. It is that the price of a clean corpus becomes a line item, and only labs that can afford to pay it get to build. The judgment that was meant to discipline Anthropic may end up functioning as a price list for everyone behind it.