Anthropic sued again, this time over what Claude hands back
On 28 August 2026 a lawsuit was filed against Anthropic, the company behind Claude, in federal court for the Northern District of California. The plaintiffs are Sony Music Publishing, Warner Chappell and other rights holders in musical compositions. Two of the company's founders are named as individual defendants alongside it, chief executive Dario Amodei and Benjamin Mann.
This is a complaint, not a ruling. Anthropic said it disagrees with the claims and intends to defend itself in court.
What the complaint says
It has two parts. The first concerns how the training material was gathered: torrenting, scraping licensed lyric repositories, and stripping the copyright information out of the files. The complaint puts the figure at tens of thousands of compositions.
The money follows from that. US law allows statutory damages of up to $150,000 for a work infringed wilfully, and up to $25,000 for each instance of removing copyright management information. Multiplied across tens of thousands of works, the sum claimed runs into billions.
The second part is the one that reaches the people using the model. The complaint says Claude returns protected text verbatim in its answers, and that the guardrails added after earlier litigation are bypassed easily: rephrase the request and ask again.
Why this is the business's problem, not only Anthropic's
A business that produces marketing copy with a language model is the one publishing the result. The interesting claim here is the mechanism rather than the sum. If a single refusal falls to a second phrasing, the refusal is not a check. It is a step.
That changes what you can read into the model's behaviour. A business that asked for text and got it without a warning did not receive confirmation that the text is original. It received text.
The copyright management information claim, which stands on its own in the complaint, lands on exactly this point. A file that arrives with its author and licence attached tells you what you may do with it. Once that information is gone, the text looks like any other text.
What to change in practice
Stop asking for something "in the style of" a named work or creator and then publishing the answer. A request like that steers the model straight at the material you are trying not to reproduce.
Check the output before it goes out. Searching a sentence or two of the text in a search engine exposes verbatim reproduction in seconds. That matters most for a slogan, a product name and an opening line, the short conspicuous pieces that tend to repeat.
Keep the prompt and the answer. If the origin of a piece of text is ever disputed, the record is what shows what was done and when.
Treat model output as a draft, not as copy. The rewriting you do afterwards is also what moves it away from any source it may have drawn on. The same check suits anything the business publishes under its own name, not text alone.
Background
This is not the first suit of its kind against Anthropic. Concord and Universal Music Group filed a similar one in January 2026, and a case brought by authors ended in a $1.5 billion settlement. The new complaint widens the front to the two largest music publishers and adds individual defendants.
For a small business using AI tools day to day, no rule changes today. What changed is that a court filing now sets out in detail what the model does with the material it trained on, and what happens when you ask a second time. The case will take months.