Sony Music Publishing, Warner Chappell Music and other music publishers have sued Anthropic in the U.S. District Court for the Northern District of California, alleging that the company used tens of thousands of copyrighted musical works—including lyrics—in the development and operation of Claude. The complaint names Anthropic, CEO Dario Amodei and co-founder Benjamin Mann. The case is unresolved, and the allegations have not been established by a court.
What the publishers are alleging
The complaint presents several related theories of liability. The publishers allege that Anthropic obtained or processed protected material through BitTorrent, pirate-library sources, lyric websites and other datasets. They also allege that copyrighted works were copied or used in the development of Claude, and that Claude produced lyrics verbatim or nearly verbatim in response to prompts.
Those are different legal questions, even when they appear in the same lawsuit:
- Acquisition: How did Anthropic allegedly obtain the books, lyrics or other protected material?
- Model development: Was that material copied or used directly, indirectly or through another model during Claude’s development?
- Output: Did Claude allegedly reproduce protected lyrics or other expressive material?
- Copyright-management information: Did the alleged conduct involve the removal or alteration of information identifying copyrighted works and their rights holders?
The publishers describe four broad categories of claims involving direct and contributory infringement, torrenting, copying and model use, and copyright-management information. The exact legal theories may develop as the litigation proceeds.
The complaint also seeks statutory damages and other remedies. Any amount requested in a complaint is a demand, not money Anthropic already owes. A court could reject the claims, award a different amount or the parties could resolve the dispute through an agreement.
Why the data path matters
The central issue is not simply whether copyrighted music existed somewhere in a training dataset. The legal exposure can depend on how the material was acquired, what was copied, how it was used in model development and what the model later produced.
That distinction matters because the complaint’s alleged chain includes several stages. A work could be acquired unlawfully, used in a training process, influence another model or dataset, and later be connected to an output. Proving one stage would not automatically resolve every other stage. The lawsuit instead puts the entire chain under scrutiny.
A separate ruling in related Anthropic book litigation treated training on lawfully acquired books as potentially transformative fair use while treating allegedly pirated acquisition as a separate issue. That does not decide the Sony-Warner music case, but it helps explain why “AI training on copyrighted material” and “obtaining copyrighted material through alleged piracy” are not interchangeable claims.
The practical question for the publishers is therefore more demanding than showing that Claude can discuss a song. They would need to connect the alleged acquisition and copying practices to the relevant legal claims, while also addressing the role of model training and any allegedly infringing output.
The evidence described in the allegations
The complaint and related reporting describe several possible routes through which the publishers may try to connect protected works with Anthropic’s systems:
- records associated with alleged BitTorrent activity;
- catalog information for books said to contain lyrics or sheet music;
- alleged scraping of lyric services including Musixmatch and LyricFind;
- internal messages cited by the publishers;
- information about datasets and model-development practices; and
- tests intended to show whether Claude reproduces protected text.
The publishers also advance an indirect theory involving synthetic data and feedback from models allegedly derived from LibGen or Pirate Library Mirror (PiLiMi). In that theory, material would not need to move straight from a pirate dataset into a commercial Claude model to become part of the alleged chain.
That remains an allegation, not a finding about a particular commercial Claude model. The disagreement over whether disputed books or other material reached commercial Claude models is one of the issues the case would have to resolve.
The distinction is important for readers, too. A model’s ability to produce a particular kind of text does not by itself establish where that text came from or whether a specific work was used during training. Conversely, evidence about how material was acquired would not, by itself, answer every question about training, outputs or legal responsibility.
Why this is not the same as Anthropic’s other copyright cases
Anthropic faces several copyright disputes involving different plaintiffs, works and legal theories. Combining their figures creates a misleading picture of the Sony-Warner lawsuit.
| Proceeding | Reported scope | What makes it distinct |
| Sony-Warner complaint | Tens of thousands of musical works alleged | The current case centers on music, lyrics, alleged data acquisition, model development, outputs and copyright-management information. |
| Concord, Universal Music Publishing Group and ABKCO Music action | More than 20,000 works reported | A separate music proceeding with its own parties, claims and procedural history. |
| BMG action | 493 compositions reported | A separate music-related action involving different works and claims. |
| Bartz v. Anthropic | Books and alleged use of pirated material | A separate book-copyright proceeding, not a ruling on the Sony-Warner music allegations. |
The earlier Concord case also matters as legal context, not as a decision in this lawsuit. Its procedural history includes disputes over preliminary relief, secondary liability and the treatment of copyright claims. None of those events resolves whether Anthropic is liable in the Sony-Warner matter.
What remains unresolved
The complaint does not establish that Anthropic is liable, that Claude definitely trained on every disputed work or that any alleged output infringed a particular copyright. It also does not turn a requested remedy into an award.
The case will hinge on several connected questions: what material Anthropic allegedly obtained, whether and how it was copied, whether it entered a model-development process, what Claude produced, and how those facts fit U.S. copyright law. The fair-use question is only one part of that analysis; the alleged provenance of the data and the treatment of outputs could matter separately.
For now, the useful bottom line is straightforward: the lawsuit tests the legal distance between alleged piracy, AI training and allegedly infringing outputs—but it has not yet closed that distance with a court ruling.
