Sony and Warner Sue Anthropic: Training Data and Lyrics Claims

What was filed

Sony Music Publishing and Warner Chappell Music filed suit against Anthropic on Friday, August 28, 2026, in the U.S. District Court for the Northern District of California. The complaint also names Dario Amodei and Benjamin Mann as individual defendants. It brings four counts: direct infringement by torrenting against all three defendants, contributory infringement by torrenting against Amodei and Mann, and against Anthropic alone, direct infringement and removal or alteration of copyright management information. [1]

The publishers demand a jury trial and ask for statutory damages of up to $150,000 for each work willfully infringed, plus up to $25,000 for each removal of copyright management information. They also ask the court to order every infringing copy destroyed and to require Anthropic to account for what is in Claude’s training data. The complaint identifies "tens of thousands" of their compositions, which is what puts the theoretical exposure in the billions. [1]

Atlas interpretation: The accounting demand is the part that would change the most if granted. Statutory damages are a number; a court-supervised inventory of training data is a discovery obligation that outlives this case and is usable in the next one. Song titles in the caption get the headlines, and the publishers named recognizable ones, but the relief that matters to a model developer is the audit. [1]

Where the allegations come from

The complaint alleges that in June 2021 Mann used BitTorrent to download at least five million pirated books from Library Genesis, and that Anthropic employees torrented at least two million more from Pirate Library Mirror in July 2022. Both figures are drawn from findings in Bartz v. Anthropic, the authors’ case in the same district, where the judge described the conduct as "straightforward piracy but at massive scale." [1]

Much of the rest is quoted from material unsealed in that earlier case: Mann describing Library Genesis as "sketchy AF," Anthropic’s own Archive Team calling it a "blatant violation of copyright," and a 2024 planning document about a destructive book-scanning operation reading "we don’t want it to be known that we are working on this." Separately the publishers allege that Anthropic scraped lyrics from licensed services including MusixMatch and LyricFind, and used Common Crawl, The Pile and Books3. [1]

Atlas interpretation: That sourcing is the strategic point of the filing. These plaintiffs are not starting from a theory about how models are trained; they are starting from an evidentiary record another set of plaintiffs already built and a court already characterized. The suit is closer to a collection action on established facts than to a fresh investigation, which is why it could be assembled and filed quickly. [1]

Outputs as well as inputs

The complaint does not stop at acquisition. It alleges that Claude reproduces the publishers’ lyrics verbatim in its outputs, and that the guardrails Anthropic added after the earlier Universal, Concord and ABKCO litigation are "easily circumventable by simply ‘re-prompting’" the model. [1]

Atlas interpretation: Input claims and output claims fail differently. A fair-use defense about training is an argument about what a model learns from a corpus, and it has had some success. An output claim is an argument about a specific string a specific model emitted, and a filter that a second prompt defeats is not a defense so much as an admission that the material is in there. Pleading both means a defense of the training theory does not end the case. [1][2]

The fifth suit, not the first

Music publishers have been litigating against Anthropic since October 2023, when Universal Music Publishing Group, Concord and ABKCO sued in Nashville over roughly 500 songs in a case later transferred to California. The same publishers filed again in January 2026 over more than 20,000 works, seeking more than $3 billion. BMG sued in March 2026 over 493 compositions, and Round Hill Music filed on August 17. With Sony and Warner Chappell in, the publishing arms of all three major music companies are now suing the same defendant. [1]

The complaint leans on Anthropic’s $1.5 billion settlement with book authors a year earlier over the same torrenting, arguing the company "clearly considers that to be just the cost of doing business" and that the amount is "obviously not a large enough settlement to deter infringing conduct" at Anthropic’s current valuation. It also attacks the company’s positioning directly, saying that "despite branding itself as the ‘ethical AI company,’ Anthropic has repeatedly acted in ways that belie that image." [1][2]

Atlas interpretation: Sony had told AI developers not to train on its catalogue two years earlier, and this is what that position looks like when it is enforced rather than announced. The publishers are explicit that they license to AI companies and want to keep doing so; the argument is about who sets the price, not about whether models may be trained at all. The naming of two founders personally is the escalation, and it is the part with no precedent in the earlier cases. [1]