Future TechnologySony and Warner Want 150,000 Dollars a Song From Anthropic
Key takeaways
- Sony Music Publishing and Warner Chappell filed a 48 page complaint alleging Anthropic trained Claude on tens of thousands of copyrighted compositions
- The publishers want up to 150,000 dollars per infringed work plus 25,000 dollars for each instance of stripped copyright management information, and they have asked for a jury
- The case targets the line a judge drew in 2025, when training on copyrighted work was found lawful but acquiring it through piracy was not
150,000 dollars a song. That is the figure Sony Music Publishing and Warner Chappell put in front of a California federal court in a 48 page complaint against Anthropic, and the arithmetic gets uncomfortable quickly when the complaint alleges tens of thousands of compositions.
The publishers say Anthropic trained Claude on copyrighted lyrics and sheet music. They name co-founders Dario Amodei and Benjamin Mann personally alongside the company, which is unusual. Most copyright suits against AI labs stop at the corporate entity.
Where the publishers say the music came from
The complaint points at Library Genesis and the Pirate Library Mirror as the sources for lyrics and sheet music. It also alleges that licensed lyric sites, including Musixmatch and LyricFind, were scraped directly.
On top of the per-work damages, the publishers want 25,000 dollars for each instance of stripped copyright management information, the metadata identifying who owns a work. They have asked for a jury.
The 2025 ruling this is aimed at
The number hanging over all of this is 1.5 billion dollars, which Anthropic agreed to pay authors in September 2025. That settlement followed a judge drawing a specific line: training a model on copyrighted work was lawful, acquiring that work through piracy was not.
Sony and Warner have built their complaint around the second half of that sentence. The claim is narrower than the headline suggests. It concerns how the files reached the training servers, not whether a model can learn from music at all.
Why the plaintiff matters as much as the argument
Anthropic has spent the past year signing deals that assume a stable legal footing, from the Salesforce partnership to a 35 billion dollar data centre commitment with Lambda. Music publishing is a smaller industry than either, but it litigates constantly and it is very good at it.
The damages structure is what separates this from the author settlement. A fixed sum can be negotiated down. A figure that multiplies by every composition in a catalogue scales in a direction the defendant cannot control. Cases at this scale also move slowly, the way the FTC advertising case against Amazon has, and discovery costs money long before any verdict arrives.
The part worth watching is whether the claims against Amodei and Mann personally survive a motion to dismiss. That is what would change how every other lab handles its data pipeline.