Sony Music Publishing and Warner Chappell are asking for up to $25,000 each time Anthropic allegedly removed or altered copyright information attached to a song. That figure matters to developers because the disputed act can happen long before a chatbot returns a lyric. In the publishers' telling, ordinary data preparation steps stripped away titles, writer names, owner names, and notices while text moved through collection and training systems. The 48-page complaint turns metadata retention into a separate source of legal exposure alongside the familiar argument over copying training material.
The publishers filed the case on August 28 in the US District Court for the Northern District of California. Anthropic, chief executive Dario Amodei, and co-founder Benjamin Mann are named as defendants. The filing alleges infringement involving tens of thousands of compositions and seeks as much as $150,000 per work for willful copyright infringement. The separate $25,000 request applies to each alleged violation involving copyright management information, often shortened to CMI. The Verge reported both maximum figures, while correctly presenting a multibillion-dollar total as a possible maximum rather than an award already made.
No court has ruled on the new claims. Anthropic had not responded to requests for comment before TechCrunch published its report. The complaint is the publishers' account, and some of its most important assertions depend on evidence that would be tested through discovery.
Two damage theories, one data pipeline
The complaint separates its case into four counts. It accuses all three defendants of direct infringement through torrenting, accuses Amodei and Mann of contributing to torrenting by employees, brings a wider direct-infringement claim against Anthropic, and alleges that Anthropic removed or altered CMI. The requested relief includes damages, an injunction, an accounting of training data and collection methods, and supervised destruction of infringing copies. Those are requests, not court orders.
The two statutory-damages numbers measure different alleged conduct. Up to $150,000 can be sought for willful infringement of a registered work under Section 504 of the Copyright Act. The complaint separately invokes Section 1203 for up to $25,000 per CMI violation. Multiplying either ceiling by every work or processing event would produce an eye-catching total, but the filing does not establish how a judge or jury would count violations or what amount would be awarded.
CMI is broader than a visible copyright symbol. Section 1202 of US copyright law covers information such as a work's title, the author's name, the copyright owner's name, and terms for use when that information is conveyed with a copy or display. The law bars certain intentional removals and alterations. It also addresses distribution when a party knows the information was removed without authority and knows, or has reason to know, that the act will help induce, enable, facilitate, or conceal infringement.
That knowledge requirement will matter. A missing field is not automatically a $25,000 violation. The publishers allege that Anthropic treated notices and related text as "useless junk" or "boilerplate," applied algorithms that removed them, and knew the resulting copies could conceal infringement. Anthropic will have the opportunity to contest the alleged conduct, intent, statutory fit, and count of violations.
The acquisition route is central to the case
The filing describes several ways copyrighted material allegedly entered Anthropic's systems. It says the company scraped licensed lyric sites including Musixmatch and LyricFind, drew from third-party datasets, scanned physical books, and obtained books through pirate libraries. The publishers say some of those books contained lyrics and sheet music for their compositions.
The torrent allegations build on facts aired in earlier litigation. According to the new complaint, Mann used BitTorrent in 2021 to obtain at least five million books from Library Genesis. Anthropic employees later downloaded at least two million more books from Pirate Library Mirror. Because BitTorrent peers generally upload pieces while downloading, the publishers claim reproduction and distribution as separate infringements. They also allege that Amodei approved the LibGen acquisition and that Mann directed the later download.
Exhibit A lists examples the publishers say appeared in the torrented books, while Exhibit B covers the much larger set allegedly involved in other copying, training, and output. Songs named in the filing include "Ain't No Mountain High Enough," "Livin' on a Prayer," "September," "Hallelujah," and "Paper Rings." These are composition claims from music publishers. The case is therefore about rights in lyrics, sheet music, and the underlying songs, rather than a claim that Claude trained on the corresponding commercial sound recordings.
That distinction explains why a collection of books can create a music case. A songbook can carry protected lyrics and notation along with the title, writers, publisher, and copyright notice. If a pipeline extracts the expressive text and drops the surrounding ownership fields, the publishers argue that the system has made one unauthorized copy and removed information used to trace it. Their fourth count tries to attach liability to that second step.
Bartz narrowed the dispute without ending it
Anthropic has already won an important part of the legal argument over training. In June 2025, Judge William Alsup held that using books to train Anthropic's language models was a transformative fair use in Bartz v. Anthropic. The same fair-use order found that keeping pirated books in a permanent, general-purpose central library was not justified by fair use. The source of a copy and the later use of that copy were treated as separate questions.
That split runs through the Sony and Warner complaint. Its torrent count targets acquisition and retention. Its broader count challenges scraping, training, outputs, and alleged derivatives. The CMI count focuses on what happened to attribution and ownership data as works passed through the system. A favorable ruling on one stage would not necessarily settle the others.
The earlier book case ended in a $1.5 billion settlement after the mixed ruling, which TechCrunch cites as the immediate legal backdrop. The new plaintiffs are trying to apply the documented book-acquisition history to songbooks and sheet-music collections. They also make allegations about licensed lyric sites and model outputs that extend beyond the central-library issue decided in Bartz.
Sony and Warner are joining an existing run of music-publisher cases against Anthropic. The Verge notes prior actions involving Universal Music Group, Concord, ABKCO, BMG, and Round Hill. The new filing is still distinct in parties and pleaded works, even where it draws on evidence and rulings from the book litigation.
What data teams can take from the filing now
The complaint does not create a new engineering standard, but it exposes a gap in many model-data discussions. Teams often record where a dataset archive came from and whether its text may be used. The lineage can disappear at the document level after extraction, deduplication, filtering, and tokenization. A dataset manifest that names the download source may be unable to connect one passage back to its author, owner, license, or original notice.
For systems that ingest licensed or copyrighted material, provenance needs to survive those transformations if the operator expects to audit or remove a work later. That means preserving the original item identifier and source URL, retaining rights and author fields, logging each transformation, and maintaining a reversible mapping between processed text and its source record where policy permits. These steps are an engineering inference from the claims, not a finding that every preprocessing pipeline is legally required to use one design.
The requested accounting shows why coarse records may fail under litigation. The publishers want Anthropic to identify their works in training data and disclose how the company collected, copied, processed, and encoded that material, including work done by third parties. If a provider cannot answer those questions from its records, reconstructing the lineage after model training could be expensive or impossible.
There is also a procurement lesson. A dataset arriving from a vendor or public repository can contain text whose original notices were removed upstream. Section 1202 includes a separate distribution provision tied to knowledge that CMI was removed or altered. Contracts and source declarations help, but technical inspection of samples and metadata fields can reveal whether the delivered data supports the provenance claims made about it.
The next concrete signals will come from Anthropic's answer and any early motion attacking the complaint. Watch whether the court treats preprocessing as actionable CMI removal, how it applies the statute's knowledge tests, and whether the publishers can connect specific compositions to particular copies and transformations. Those decisions will say more about the risk to AI data pipelines than the complaint's theoretical maximum damages.