Universal Music v. Anthropic
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,943 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,943 words
Add missing citations, update stale details, or suggest a clearer explanation.
Universal Music v. Anthropic is the common name for a federal copyright lawsuit that Universal Music Publishing Group, Concord, and ABKCO Music filed against Anthropic on January 28, 2026, in the U.S. District Court for the Northern District of California. The publishers accuse Anthropic of building its Claude models on a foundation of piracy, alleging that the company and two of its founders used the BitTorrent protocol to download large archives of pirated books and songbooks that contained the publishers' song lyrics, and that newer Claude models still reproduce those lyrics on demand. The complaint identifies 714 works that were allegedly torrented and 20,517 additional works swept into training, for a total of more than 21,000 compositions, and seeks statutory damages that the publishers say could exceed $3 billion. Press coverage frequently rounds the exposure to roughly $3.1 billion, the figure produced by applying the $150,000 statutory maximum to the full set of works.[1][2][3]
The case is a sequel to, and distinct from, the publishers' earlier suit. In October 2023 the same coalition sued Anthropic over Claude's reproduction of lyrics in its outputs, a matter usually called Concord I. The 2026 action, sometimes labeled Concord II, drops the output focus of the first case in favor of the act of acquisition itself: it argues that downloading pirated copies was infringement the moment it happened, regardless of what Anthropic later did with the files. Because it names Dario Amodei and cofounder Benjamin Mann as individual defendants and covers more than 21,000 works, the plaintiffs have described it as one of the largest, if not the single largest, non-class-action copyright cases ever filed in the United States.[2][4]
The plaintiffs are the music publishing arms of two of the three major music groups plus a prominent independent rights administrator. Universal Music Publishing Group is represented through Universal Music Corp. and affiliated entities; Concord is the catalog owner behind compositions administered for a long list of songwriters; and ABKCO Music & Records controls early Rolling Stones publishing along with works by Sam Cooke and others. The same coalition, represented by Oppenheim + Zebrak LLP, brought the 2023 case.[1][5]
The defendants are Anthropic PBC, the San Francisco public benefit corporation that develops the Claude family of large language models, along with chief executive Dario Amodei and cofounder Benjamin Mann. Naming the two executives personally is a deliberate escalation from Concord I, which sued only the company. The suit was filed in the Northern District of California, where both the 2023 publishers' case and the related authors' case were also litigated.[2][3]
The 2023 suit, Concord Music Group, Inc. v. Anthropic PBC, was filed in the Middle District of Tennessee in October 2023 and later transferred to the Northern District of California, where it carries case number 5:24-cv-03811 before Judge Eumi K. Lee. That complaint named roughly 500 musical compositions, including Katy Perry's "Roar," the Rolling Stones' "You Can't Always Get What You Want," and Don McLean's "American Pie," and it centered on outputs: the publishers showed that Claude would generate verbatim or near-verbatim lyrics when prompted, sometimes even when asked only to write an original song on a related theme. The case also raised a claim under section 1202 of the Digital Millennium Copyright Act for removing copyright management information from scraped pages. As of the first half of 2026, that earlier case is in summary judgment briefing, with the publishers having moved on March 23, 2026 for a ruling that Anthropic infringed and that fair use does not apply.[5][6][7]
The two cases are easy to conflate but should not be. Concord I is about what Claude says; the 2026 case is about how Anthropic got the underlying material. The roughly $3 billion figure attaches to the 2026 piracy suit and its 21,000-plus works, not to the smaller 2023 matter.[3][4]
The split into two cases was not the publishers' first choice. The torrenting evidence surfaced through a separate lawsuit by book authors, Bartz v. Anthropic, where in mid-2025 the court found that Anthropic had used BitTorrent to pull millions of books from the shadow libraries Library Genesis and Pirate Library Mirror. The publishers then tried to add piracy claims to their existing 2023 case, but Judge Lee declined to let them amend the older complaint that late in the litigation. Rather than abandon the new theory, they filed it as a fresh action in January 2026.[1][2]
The core of the 2026 complaint is acquisition. The publishers allege that Anthropic, to assemble training data cheaply, downloaded millions of unauthorized digital books from Library Genesis and Pirate Library Mirror using BitTorrent, and that hundreds of those files were songbooks, sheet music collections, and other volumes reproducing the publishers' lyrics in full. They argue that the download alone was a completed act of reproduction, and that because BitTorrent shares files between peers as it fetches them, Anthropic simultaneously distributed copies back out to other users, infringing the distribution right as a separate matter. The complaint further alleges that Anthropic kept these files as a permanent "central library" that it intended to retain indefinitely, independent of whether any given book was ever used to train a model.[1][3]
The complaint reaches past the corporation to its leadership. Drawing on evidence from the authors' litigation, it alleges that Benjamin Mann personally downloaded on the order of five million pirated books via BitTorrent in mid-2021, and that Dario Amodei discussed and approved acquiring works this way despite internal recognition that the sources were, in the plaintiffs' characterization, "sketchy." One widely quoted line frames the theory bluntly: while Anthropic "misleadingly claims to be an AI 'safety and research' company," its "multibillion-dollar business empire has in fact been built on piracy."[2][3]
The 20,517 works listed in the complaint's second exhibit go to ongoing infringement through training and generation. The publishers allege that recent Claude models were trained in ways that let them memorize and emit the publishers' lyrics, and that Anthropic's output filters are a band-aid that addresses only what users see while doing nothing about the unlicensed copying baked into the models. They note that the guardrails can be defeated through ordinary jailbreaking and that Claude still produces infringing text in many cases.[1][4]
As in the 2023 case, the publishers bring a DMCA claim under section 1202, alleging that Anthropic stripped copyright notices, author credits, and other identifying metadata from the works before ingesting them, and that internal communications dismissed such notices as low-value clutter. Removing this information, they argue, made it harder to detect the infringement and is independently actionable.[1][5]
The publishers seek statutory damages of up to $150,000 per work for willful infringement under 17 U.S.C. section 504(c), plus up to $25,000 per violation for the removal of copyright management information under section 1203(c)(3)(B). Applied across the full set of more than 21,000 works, the copyright theory alone yields a ceiling north of $3 billion, which is the basis for the reported $3.1 billion figure. The dollar amount is a statutory maximum rather than a fixed demand, so the actual exposure depends on how many works a jury finds infringed and whether it finds willfulness. Beyond money, the complaint asks for a permanent injunction, an accounting of Anthropic's training data and methods, and destruction of all infringing copies.[1][3][4]
The publishers themselves phrase the headline number as damages that "could exceed" $3 billion, so it should be read as a statutory ceiling rather than a fixed demand. The contrast in scale with Anthropic's finances is part of the narrative: when the 2023 case was filed the company was valued around $5 billion, and by early 2026 reporting put its valuation in the hundreds of billions of dollars.[2][4]
The 2026 suit sits squarely inside the wave of generative AI copyright cases working through U.S. courts, and it borrows heavily from the most important of them. In Bartz v. Anthropic, a class of book authors sued over the same torrented archives. In June 2025 Judge William Alsup issued a split summary judgment ruling that has shaped every case since: training a model on lawfully obtained books was "quintessentially transformative" and protected by fair use, but assembling and keeping a library of pirated copies was not, and that piracy could be infringement on its own. Facing potentially ruinous statutory damages at trial, Anthropic agreed in late 2025 to settle Bartz for at least $1.5 billion covering roughly 500,000 works, which works out to about $3,000 per work; the court granted preliminary approval in September 2025.[8][9]
The music publishers built their 2026 complaint on the half of Alsup's ruling that went against Anthropic. Their wager is that the fair use defense that protected lawful training does nothing for files obtained by piracy, so the act of torrenting can be litigated as straightforward infringement without ever reaching the harder questions about how models learn. The case has drawn support from the rest of the industry: trade groups including the RIAA and the National Music Publishers' Association, along with book publishers and authors, have filed amicus briefs backing the publishers and calling Anthropic's unlicensed copying inexcusable.[6][10]
Anthropic has generally defended the AI-copyright cases on fair use grounds, leaning on the part of the Bartz ruling that blessed training as transformative, and it has pointed to the output guardrails it added to Claude as evidence that it takes rightsholders' concerns seriously. In the 2023 case it has disputed both the infringement and the DMCA theories and argued that the publishers cannot tie the use of lyrics as training input to any specific infringing output. When the 2026 piracy suit was filed, the company declined to comment to reporters, so its formal answer in that case will come through its court filings. The publishers anticipate the fair use argument and try to neutralize it in advance: they stress that the use was purely commercial, that whole lyrics were copied, and that AI-generated lyrics compete directly with licensed services such as LyricFind and Musixmatch, all factors that cut against fair use under the statutory test.[1][7]