Sony and Warner Sue Anthropic Over Lyrics: Explained
The Short Answer
Sony Music Publishing and Warner Chappell Music sued Anthropic in California federal court in late August 2026 — filed in the Northern District of California and widely reported August 29–31, 2026.
The core allegations:
- Anthropic torrented, scraped and downloaded copyrighted musical compositions at scale to train Claude
- Claude reproduces copyrighted lyrics verbatim when prompted
- Anthropic used those lyrics to teach Claude to generate competing AI song lyrics
The publishers call it “one of the largest and most blatant ongoing thefts of intellectual property in history.” Anthropic co-founders Dario Amodei and Benjamin Mann are named as individual defendants.
Anthropic says it disagrees with the claims and intends to defend itself.
Last verified: September 5, 2026.
The Damages Arithmetic
This is where the “multi-billion dollar” framing comes from, and it is worth doing the math rather than accepting the headline.
The publishers seek:
- Up to $150,000 per willfully infringed work (statutory maximum)
- Up to $25,000 for each alleged removal of copyright management information
The complaint describes “thousands if not tens of thousands” of works, naming compositions associated with The Beatles, Taylor Swift, Michael Jackson and Mariah Carey among others.
At the statutory ceiling, 10,000 works reaches $1.5 billion on the infringement claims alone, before the copyright-management-information count. Statutory maximums are rarely awarded in full — but they are the anchor every settlement negotiation starts from, and Anthropic already has a $1.5 billion data point on its own record.
How This Fits the Broader Pattern
This is not an isolated action. The sequence matters:
| Action | Subject | Status |
|---|---|---|
| Authors and publishers v. Anthropic | Books | $1.5B settlement, final approval 2026 |
| Sony Music Publishing + Warner Chappell | Compositions and lyrics | Filed late Aug 2026 |
| Universal Music Group | Music | Separate suit |
| BMG | Lyrics | Separate suit |
| Round Hill Music | Music | Separate suit |
Two distinct legal theories are in play, and conflating them is the most common analytical error here.
Theory one — acquisition. How the training data was obtained. Allegations of torrenting and scraping target the acquisition of copies, which is a separate question from whether training itself is transformative. This is the theory that produced the largest recoveries so far, because “we downloaded it from a pirate source” is a much simpler factual claim than “training is or is not fair use.”
Theory two — output. Whether the model reproduces protected expression. The verbatim-lyrics allegation goes here, and it is the one with the clearest technical remedy: output filtering.
The complaint’s third element — that Claude was trained to generate competing lyrics — is a market-substitution argument, which speaks directly to the fourth fair-use factor. That is a deliberate framing choice, not incidental.
What Actually Changes for Builders
In the short term: very little. There is no indication of service disruption, no model withdrawal, and Anthropic is defending rather than settling on day one. If you are building on Claude, nothing about your deployment changes this week.
In the medium term, three things are worth tracking:
1. Training-data liability gets priced in. Repeated nine- and ten-figure settlements across the industry become a cost of doing business, and that cost lands somewhere — in model pricing, in restricted training corpora, or in narrower model capabilities. A model trained on a legally defensible subset is a different product from one trained on everything.
2. Output filtering becomes your problem too. If a model can emit copyrighted lyrics verbatim and your product surfaces that to users, the vendor’s legal exposure is not the only exposure in the chain. Practical step: filter verbatim reproduction of song lyrics in any user-facing product, the same way you would filter other reproduction risks. Do not assume the vendor’s guardrails are sufficient or stable across model versions.
3. Individual-defendant naming is an escalation. Naming Dario Amodei and Benjamin Mann personally, rather than only the corporate entity, signals a strategy aimed at discovery into decision-making rather than a straightforward damages claim. That tends to produce more disclosure about how training corpora were assembled — which is the material that shapes the next round of cases across the whole industry.
The Broader Reading
The pattern through 2026 is that the acquisition question is proving far more dangerous to AI labs than the fair-use question. Courts have been genuinely uncertain about whether training constitutes transformative use. They have been much less uncertain about obtaining copies from unauthorized sources.
For anyone building models rather than on them, that is the operative lesson: provenance of the corpus is the liability, more than what you did with it afterwards.
Sources
- Sony Music Publishing and Warner Chappell sue Anthropic — Music Business Worldwide
- Sony, Warner Music sue Anthropic, saying it pirated songs to train its AI — Al Jazeera
- Anthropic hit with new copyright lawsuit from Sony, Warner — PCMag
- Anthropic sued over alleged theft of songs to train Claude — The Guardian