Sony and Warner sue Anthropic over thousands of song lyrics

What this is about

Sony Music Publishing and Warner Chappell Music filed a lawsuit in California on August 28, 2026 against Anthropic and co-founders Dario Amodei and Benjamin Mann. The publishers allege that copyrighted compositions were acquired at scale through torrents and websites, copied for Claude training and sometimes reproduced in model outputs.

The allegations have not been proven. They come from a complaint reported by Music Business Worldwide and TechCrunch on August 29. Anthropic can contest them, and a court must still assess the evidence and legal questions. The case nevertheless matters because it challenges training, data acquisition and model output at the same time.

What the lawsuit actually does

The plaintiffs bring four claims. They include direct copyright infringement through torrenting, contributory infringement by the two co-founders, further direct infringement by Anthropic, and alleged removal or alteration of copyright-management information.

The publishers say the case covers “tens of thousands” of compositions. Named examples include “Ain’t No Mountain High Enough,” “All I Want for Christmas Is You” and “Eye of the Tiger.” They request a jury trial, destruction of infringing copies and an account of training data.

The statutory ceiling discussed is up to $150,000 for each willfully infringed work and up to $25,000 for each alleged removal of rights information. This is not an awarded amount. Whether claims succeed, how many works qualify and what damages a court might recognise remain open.

Why it matters

Many AI disputes treat the dataset and later outputs as separate issues. This complaint connects them. It alleges unlawful acquisition through Library Genesis and Pirate Library Mirror, while also claiming that Claude reproduced lyrics and that safeguards could be bypassed.

That increases pressure on model providers to document the provenance of training collections, not only filter outputs. For rightsholders, the central question is whether licences must be negotiated before works enter training datasets.

The case may also put executive decisions under closer scrutiny. The complaint names Amodei and Mann individually. That does not establish personal liability, but it makes internal decisions and documents more important to the litigation.

In plain language

A restaurant cannot steal another chef's cookbook, copy every recipe and then argue that only the finished dishes matter. The publishers are effectively saying that obtaining the collection was unlawful and that some resulting “dishes” also resemble protected material too closely. The court must decide whether that analogy fits the law.

A practical example

Suppose a publisher identifies 10,000 protected songs that entered a training collection without permission. Damages would not automatically equal 10,000 times $150,000. A court would still need to assess ownership, use, intent, jurisdiction and possible defences for each work.

If the model returns longer passages in 200 test prompts, the parties would also dispute whether those outputs reproduce substantial protected expression and how they were produced. The gap between an allegation and court-ready proof is therefore significant.

Scope and limits

First, this is a complaint, not a judgment. The publishers' account is adversarial and must be tested in court.

Second, a theoretical statutory maximum is not a prediction of a settlement or award. Courts can reject claims, group works or set much lower damages.

Third, the case will not automatically decide whether every use of copyrighted material for training is unlawful. Acquisition methods, copies, outputs, licences and US legal exceptions require separate analysis.

SEO & GEO keywords

Anthropic, Claude, Sony Music Publishing, Warner Chappell Music, AI copyright, song lyrics, training data, copyright lawsuit, generative AI, music publishers