Sony and Warner sue Anthropic over lyrics used to train Claude

Sony Music Publishing and Warner Chappell sued Anthropic on August 28, 2026 in the U.S. District Court for the Northern District of California. The publishers say Anthropic built Claude partly on song lyrics it had no right to use. The complaint also names two people personally: chief executive Dario Amodei and co-founder Benjamin Mann.

Anthropic rejects the case. "We disagree with the publishers' claims and we intend to defend ourselves robustly in court," the company told TechCrunch.

What the complaint alleges

The publishers describe what they call a "brazen campaign of illegally torrenting, scraping, and downloading copyrighted works." The claim splits into two separate routes.

The first is torrenting. According to Music Business Worldwide's account of the filing, Mann used BitTorrent in June 2021 to download at least 5 million pirated books from Library Genesis. It says Anthropic employees took roughly 2 million more from the Pirate Library Mirror in July 2022. Books matter in a music case because they contain printed lyrics and sheet music.

The second is scraping. The complaint says Anthropic took lyrics from licensed sites including MusixMatch and LyricFind without permission. These are services that pay publishers for the right to display those words.

There is a third claim that developers should notice. The publishers say Anthropic removed or altered copyright management information, meaning the ownership data attached to a work. That is a separate violation from copying, and it carries its own damages.

The complaint quotes Anthropic's own internal materials describing Library Genesis as "sketchy AF" and a "blatant violation of copyright."

The money at stake

The publishers ask for statutory damages of up to $150,000 for each work willfully infringed. They also seek up to $25,000 for each instance of stripped copyright information.

The number of works is where the reporting differs. TechCrunch puts it at 20,000 works. Music Business Worldwide quotes the complaint as saying "tens of thousands" of compositions. Engadget describes "thousands upon thousands." Either way, the arithmetic runs into billions.

Anthropic has been here before. In September 2025 it settled a case brought by authors for $1.5 billion, over what the new complaint calls the same torrenting conduct. Engadget notes this is the second music publishing suit against the company, after Concord and Universal Music Group sued seeking $3 billion.

The legal line this case tests

The argument is not that training on copyrighted work is illegal. Courts have been more willing to accept training itself as fair use. The publishers' case attacks the step before training: how the material was obtained.

That distinction is the whole fight. Downloading a pirated library is a copyright act on its own, regardless of what you later do with the files. Anthropic's $1.5 billion settlement suggests the acquisition route is the weaker ground, and the publishers have aimed directly at it.

What this means for developers

Provenance is now a build requirement, not a legal footnote. If you fine-tune a model, keep a record of where every dataset came from and under what terms. The question a court asks is not what your model outputs. It is where the training files came from and whether you were entitled to them.

Treat scraped data from licensed intermediaries as radioactive. MusixMatch and LyricFind pay for their rights, and scraping them does not transfer those rights to you. The same reasoning applies to any aggregator that licenses content it did not create, including stock image sites and code that carries a restrictive licence.

Do not strip metadata. The copyright management information claim is the quiet lesson here. Pipelines routinely discard attribution during cleaning because it is noise for a tokenizer. That habit creates a separate cause of action, and it is trivially avoidable by keeping the provenance record alongside the cleaned text.

Individual liability is the new part. Naming a chief executive and a co-founder personally, over a download made in 2021, changes the risk calculation for anyone who assembles training data inside a company. "I was told to get the corpus" is a weaker position than it looked a year ago.