The lawsuits are piling up, but the real story is that asking permission was never part of the business model.
The Summary
- Sony Music and Warner sued Anthropic for allegedly torrenting music catalogs and scraping lyrics wholesale, with co-founder Benjamin Mann allegedly directing the pirating in company Slack channels
- All three major music publishers have now targeted Anthropic, after it already paid $1.5 billion to settle book piracy claims last year
- The pattern is clear: AI companies treat "fair use" as "take everything and defend later"
The Signal
The lawsuit claims Anthropic co-founder Benjamin Mann personally torrented music catalogs and discussed it openly in internal Slack channels. Not through intermediaries. Not through scrapers hidden behind corporate veils. The co-founder himself, coordinating mass copyright infringement like it was a college dorm project. This is not a compliance failure. This is the strategy.
Anthropic already paid $1.5 billion last year to settle claims over pirated books. Now it faces all three major music publishers. The company's response to Axios was five words: "we intend to defend ourselves robustly in court." Translation: we have $30 billion in funding and you have to prove damages in a system that moves slower than model training.
"AI companies will always take the most liberal view of 'fair use' when it comes to harvesting content for their models."
The economics explain everything. Training a frontier model costs hundreds of millions. Licensing content properly costs... also hundreds of millions, except you have to negotiate with every rights holder and some will say no. Settlement costs, even big ones, are just delayed licensing fees with better terms. You train first, fight later, and keep the model advantages you gained.
When OpenAI's then-CTO Mira Murati froze on camera last year trying to explain what data trained Sora, it wasn't because she didn't know. It was because the truth was legally toxic. The video model that wowed the internet was trained on YouTube videos, probably millions of them, without asking anyone. Sora is discontinued now, but the training data remains in the OpenAI knowledge base, ready for the next version.
Here's what content creators need to understand:
- "Don't scrape my content" robots.txt files are suggestions, not laws
- Licensing deals are signed by companies that already trained on your content unlicensed
- The legal system moves in years; model training happens in months
The Implication
Content owners have two options. Sue after the fact and hope for a settlement that might arrive years later, after the model advantage is already captured and monetized. Or verify everything upfront. Every agreement needs technical enforcement: allowlists, token tracking, audit rights on model training logs. The standard licensing contract written for Web2 is useless in Web4 when your content can be ingested, embedded, and regurgitated without your name on it.
The trust-and-verify model is dead. Anthropic just proved it again. Verify-and-enforce is the only framework that works when the incentive structure rewards piracy over permission.