SAN FRANCISCO, 30 AUG 2026 — Sony Music Publishing and Warner Chappell Music have sued Anthropic in the Northern District of California over the material used to train Claude. The complaint takes the unusual step of naming chief executive Dario Amodei and co-founder Benjamin Mann as individual defendants.
What is alleged
The suit, filed on 28 August, accuses Anthropic of a campaign of illegally torrenting, scraping and downloading copyrighted works at scale for model training.
The specific allegations are unusually granular. The complaint says lyrics were scraped from licensed sites including MusixMatch and LyricFind, that a destructive scanning operation was run on second-hand books, and that Common Crawl, The Pile and Books3 were used. It alleges Mann used BitTorrent to download more than five million pirated books and that other employees obtained at least two million more from Pirate Library Mirror.
Damages sought are up to US$150,000 for each infringed work, plus US$25,000 for every instance in which identifiable copyright information was removed.
The billions figure is arithmetic, not a demand
Coverage has described this as a multi-billion-dollar lawsuit. No such sum appears in the complaint. The figure is arrived at by multiplying a statutory maximum by an asserted corpus.
Statutory damages under US copyright law run to US$150,000 per work only for wilful infringement, and courts have discretion across a wide range below that. A claim covering a large catalogue therefore produces an enormous theoretical maximum and a much smaller realistic exposure, and the two are routinely reported as the same figure.
The second head of damages is the more distinctive one. Removal of copyright management information is a separate statutory violation, and alleging it as a distinct count is a way to multiply exposure independently of the infringement claim. Whether stripping metadata during a scraping pipeline meets that standard is contested, and it is one of the questions this case would decide.
Naming the founders is a choice
Individual liability for corporate infringement requires showing that a named person directed or personally participated in the conduct, and it is a high bar that most such claims do not clear.
Naming Amodei and Mann has two immediate effects, regardless of whether the claim against them survives. It creates discovery obligations that reach individuals rather than only the company, and it raises the personal stakes of the litigation for the people who decide whether to settle.
It also follows from the specificity of the allegations. A complaint that names a specific co-founder and a specific protocol is asserting personal conduct rather than institutional negligence, and personal conduct is the basis for individual liability. Whether the evidence supports it is the whole question.
This publication has an obvious interest in this dispute, since Claude is an Anthropic product. The facts above are the allegations as filed. Anthropic has said it will defend itself, and nothing has been proved.
Acquisition and use are separate questions
Headlines often collapse a distinction that matters here, between how the material was obtained and what was done with it afterwards.
Whether a model may be trained on lawfully acquired copyrighted material is the unsettled question fair use doctrine is being asked to answer. Downloading that material from a piracy source raises a much older question, one that can be answered without any novel view of AI.
That is why the complaint's detail about BitTorrent and Pirate Library Mirror matters more than its detail about model training. A defendant can argue that training is transformative and still have a problem with how the corpus arrived, because the two are judged separately.
It also explains a pattern visible across this whole wave of litigation: plaintiffs plead acquisition heavily, because it is the count least dependent on a court accepting a new theory. Any organisation assembling a training corpus should read that as guidance about where its own exposure sits.
All three major publishers are now litigating
With Sony Music Publishing and Warner Chappell filing, the publishing arms of all three major music companies are in litigation against Anthropic.
That matters for a subtle reason. Music publishing is unusually concentrated and unusually well documented: ownership of a composition is registered, licensing is administered through established societies, and the industry has litigated technology adoption for a century. It is, in short, the rights-holder class best equipped to bring this kind of case.
That is why music keeps arriving first. The underlying question of whether training on copyrighted material without a licence is permitted applies equally to books, journalism, images and code, but music publishers have the registries and the litigation habit to act on it.
What this means for the region
The immediate consequence for readers here is about licensed data rather than about American damages.
We have written that five APAC markets have five different AI rulebooks and agree on the one thing that matters, which is training data licensing. A US ruling that scraping without a licence is infringement would harden that convergence considerably, because it would establish a cost for the alternative.
For regional model builders the practical implication is the same either way. Documenting the provenance of training data is becoming a requirement rather than an assumed liberty. An organisation fine-tuning on a corpus it did not license, or on weights trained on one, is carrying a risk that is currently unpriced and increasingly visible.
That is worth acting on before the case concludes, because litigation of this size is measured in years and procurement questions are already being asked now.
View author profile → · Editorial policy
