Sony and Warner Just Changed the AI Industry's Core Business Model

The suit that changes everything

Sony Music Publishing and Warner Chappell did not come to play. On August 28, they filed a 48-page complaint in the U.S. District Court for the Northern District of California that calls Anthropic’s conduct “one of the largest and most brazen instances of intellectual property theft in history.” The language is not hyperbole. It is a legal strategy.

This lawsuit is different from every music-AI case before it. Not because of the drama — though the drama is real — but because of the scale, the specificity, and the direct challenge to the foundational assumption that powers the entire generative AI industry: that you can take creative work without paying for it.

Thirty-five plaintiffs. Up to $150,000 in statutory damages per song. A request for the company’s training data to be disclosed. A demand that the infringing copies be destroyed. If the court accepts the plaintiffs’ framing, this case does not just threaten Anthropic. It threatens the free-training-data model that every AI company from OpenAI to Google to xAI has quietly relied on.

Who is suing, and why now

The complaint names three defendants: Anthropic itself, CEO Dario Amodei, and co-founder Benjamin Mann. Naming the individuals matters. It signals that the plaintiffs believe these decisions were not accidental — they were directed at the top.

The lawsuit draws heavily on findings from a previous case, Bartz v. Anthropic, where a federal judge already acknowledged that Anthropic acquired vast quantities of pirated books. That earlier settlement — $1.5 billion, finalized in July 2026 — is treated here not as closure but as evidence of a pattern. “It is apparent that Anthropic views this settlement amount as merely a cost of doing business,” the complaint states. That framing turns a financial loss into a legal argument about willful infringement, which unlocks higher damages.

The evidence: how the data was acquired

The complaint paints a detailed picture of systematic acquisition. Mann downloaded more than 5 million books from the pirate site LibGen using BitTorrent in June 2021. The plaintiffs argue this violates not just reproduction rights but distribution rights, because BitTorrent turns each downloader into a distributor — a nuance that expands liability.

In July 2022, another Anthropic employee pulled over 2 million books from PiLiMi, a second pirate site. Among the material: songbooks containing the lyrics to hits like Bon Jovi’s “Livin’ On a Prayer,” Michael Jackson’s “September,” and Leonard Cohen’s “Hallelujah.” These are the works owned or controlled by the 35 plaintiffs.

Beyond piracy, the complaint lists four additional channels: scraping from licensed lyric sites MusixMatch and LyricFind; purchasing used books, scanning them, and then destroying the physical copies — what the filing calls “destructive scanning”; and ingesting data from Common Crawl, The Pile, and the Books3 dataset.

Each channel is a potential vector of liability. Taken together, they form an architecture of circumvention.

The smoking gun: intentional deletion of copyright notices

Perhaps the most consequential allegation in the 48-page document is not about what Anthropic took — it is about what it erased.

The complaint asserts that Anthropic deliberately stripped copyright management information (CMI) from the data it ingested. CMI includes things like author names, publisher credits, and copyright notices embedded in web pages and documents. Federal law imposes separate penalties for the removal or alteration of CMI — up to $25,000 per violation, on top of the per-work damages.

Internal communications cited in the complaint reveal that Anthropic rejected a scraping tool called jusText precisely because it preserved footer copyright notices. The company instead adopted a tool called Newspaper, which the internal discussion described as more effective at removing copyright information entirely. This is the difference between negligence and intent, and in copyright law, that distinction is everything.

Claude reproduces lyrics verbatim

The complaint does not stop at training data. It points to output. According to the filing, Claude reproduces copyrighted song lyrics verbatim when prompted — not paraphrases, not summaries, but exact text. The plaintiffs note that Anthropic introduced safety guardrails after prior lawsuits, but argue these can be circumvented by simply rephrasing the prompt.

This matters because it closes the loop between the alleged misconduct and the harm. The company took the data. It removed the attribution. It produced the output. The chain is complete.

The stakes: billions on the line

The plaintiffs have requested a jury trial and are seeking statutory damages of up to $150,000 per song for willful infringement, plus up to $25,000 per instance of CMI deletion or alteration. With tens of thousands of songs in the complaint’s scope, the total could reach billions of dollars. They are also asking the court to order Anthropic to disclose its training data and to destroy the infringing copies.

No AI company has ever faced a copyright case of this magnitude — not in scope, not in the specificity of the allegations, and not in the potential financial exposure. The $1.5 billion authors’ settlement was a warning shot. This lawsuit is the main event.

What happens next

Anthropic told TechCrunch it “cannot agree with the publishers’ claims and will contest them fully in court.” That is the expected posture. But the real question is what happens to the industry.

If the plaintiffs prevail — and particularly if a jury finds willful infringement and CMI removal — the precedent would apply to every AI developer that ingested copyrighted material without a license. The free-training-data model would face either catastrophic damages or a forced pivot to licensed data. Both outcomes raise the cost of building frontier models.

The music industry, which has been litigating against AI companies for two years, now has a roadmap. Other publishers — book publishers, newspaper publishers, even visual artists — can follow the same framework: document the acquisition path, prove the CMI stripping, show the output, and demand per-unit damages on a scale that makes the math unmistakable.

Sony and Warner did not file this lawsuit to settle for a quiet payment. They filed it to draw a line. The question is whether the AI industry will stop at that line, or whether it will test what happens next.