Anthropic, the company that built its brand on being the sensible, safety-first lab in the AI race, is being sued again. This time the plaintiffs are Sony Music Publishing and Warner Chappell Music, joined by their affiliates for a total of 35 publishers, and the accusation is blunt. In a complaint filed in California federal court in late August, they call it “one of the largest and most blatant ongoing thefts of intellectual property in history.”
The songs at issue are not obscure. The complaint names, among tens of thousands of works, the Beatles’ “I Am the Walrus” and Leonard Cohen’s “Hallelujah”. Dario Amodei and Benjamin Mann, two of Anthropic’s co-founders, are named as defendants personally, alongside the company.
What the publishers say happened
The core allegation is about how the training material was obtained. The publishers claim Anthropic acquired lyrics and sheet music by “illegally torrenting, scraping and downloading copyrighted works on a massive scale”, pulling from pirate sources including Library Genesis and the Pirate Library Mirror, and scraping licensed lyric sites such as Musixmatch and LyricFind. On top of the copying itself, they allege Anthropic stripped copyright-management information from the works, a separate violation under US law.
The money being sought is eye-watering on paper: up to $150,000 for each infringed work, plus $25,000 for each instance of removing copyright information. Multiply that across “tens of thousands” of compositions and you get the multi-billion-dollar figure the trade press has attached to the case. They are asking for a jury trial and an injunction.
The publishers are careful to say they are not anti-AI. Their statement stresses that they “recognise the potential of ethical AI technology, and they have entered licences permitting the authorised use of their musical compositions”, the point being that a licensing market exists, Anthropic knew it existed, and (they allege) chose to take the works without paying into it.
What Anthropic says
Anthropic is not backing down. Its statement: “We disagree with the publishers’ claims and we intend to defend ourselves robustly in court.” The company leans on the argument that has partly worked for it before, that training AI models on copyrighted material can qualify as fair use. A spokesperson also tried to deflate the drama, noting “This is the third lawsuit from the same lawyers, recycling allegations from cases already before the courts.”
That last line has some truth to it. Universal Music Publishing, Concord and ABKCO sued Anthropic in Nashville back in October 2023, and filed a second suit in January 2026 covering more than 20,000 works and seeking over $3 billion. The Sony and Warner action is the newest front in a war that has been running for nearly three years.
Why this one is different, and why it might matter more
Here is the piece worth understanding, because it is easy to lump all these suits together. In September 2025, Anthropic settled a class action from book authors for a reported $1.5 billion. The judge in that case drew a distinction that now hangs over everything: training on copyrighted works was found to be potentially lawful, but obtaining those works through piracy was not. The legality turned less on the training and more on the sourcing.
That is precisely the ground the music publishers have chosen to fight on. Their complaint is built around the alleged pirate sources (LibGen, the Pirate Library Mirror) rather than the abstract question of whether an AI is allowed to learn from lyrics at all. If the earlier ruling holds, the sourcing allegation is the more dangerous one for Anthropic, because “we used a pirate library” is a much harder thing to defend than “our model learned patterns from public text.”
The bigger picture for AI music
Zoom out and a pattern emerges across the whole sector. The rights-holders have stopped treating AI as a threat to be resisted and started treating it as a market to be licensed, on their terms, with lawsuits as the negotiating lever. It is the same story we covered when Suno put download limits on the music its users generate: the free-for-all era of AI creativity is being replaced, one settlement and one licence at a time, by a permissions economy. The labels intend to be paid.
For Anthropic specifically, the reputational sting is sharper than for most. When your entire pitch is that you are the responsible one, a complaint alleging you torrented the Beatles lands differently than it would for a company that never claimed the moral high ground. Whether the fair-use defence holds or the sourcing allegation sinks it, the “ethical AI company” framing is going to take fire in open court.
For everyone building on or investing in these models, the takeaway is duller but more useful: the cost of training data is going up, and the biggest legal risk is no longer whether AI can learn from copyrighted work, but where the training set came from. That is a question every lab now has to answer with receipts.
This is a summary of active litigation, not legal advice; the allegations are unproven and Anthropic denies them.
Related on Top Tool Stack: Suno’s download limits and the licensed era of AI music.