Sony and Warner have filed suit against Anthropic, and the language in the complaint is not the usual careful corporate hedging. The labels describe a "brazen campaign" of intellectual property theft, and call it "one of the largest and most blatant ongoing thefts of intellectual property in history." Those are not the words companies use when they expect a quiet settlement. They are the words companies use when they want a jury, and every other rights holder watching, to understand exactly how they intend to frame this.
What the lawsuit actually says
The core allegation is simple even if the legal machinery around it isn't: Anthropic trained its models on copyrighted music catalogs without a license. The labels aren't arguing that AI music tools are bad, or that Anthropic shouldn't build language models. They're arguing that the specific material used to build those models was theirs, and that nobody asked.
Anthropic has spent the last two years building a reputation as the safety-conscious alternative to OpenAI. Its marketing to enterprise buyers has leaned hard on that reputation. Agencies that switched clients onto Claude, or built internal workflows around it, often did so partly because "the careful one" felt like a defensible choice to put in front of a client's legal team. This lawsuit doesn't prove that reputation was false. But it does mean the reputation and the training data are two separate questions, and agencies have been answering only the first one.
Why "safe" was never the same question as "licensed"
A model can be careful about the outputs it refuses to generate and still have been built on inputs nobody had the right to use. Those are different layers of the same product, and vendors have strong incentives to let buyers conflate them. "Safe" usually means the guardrails on what the model will say back to you. It says nothing about what went into it in the first place.
This is the same company whose Claude Code limit change was described as a raise on paper but a cut in practice, once the fine print of what usage actually meant to customers came out. The pattern worth noticing isn't that Anthropic is uniquely dishonest, it's that the terms a vendor advertises and the terms that actually govern your usage can diverge, and you only find out which one you're living under after you've built workflows on top of it.
Anthropic is also the company pushing to extend its Model Context Protocol from software into physical hardware, meaning its footprint in how agencies and their clients connect tools together is growing, not shrinking. The more central a vendor becomes to your stack, the more expensive it is to discover after the fact that its foundational claims don't hold up.
The vendor stack question agencies now must ask
There's already a precedent for what happens when a vendor's history catches up with a customer relationship: OpenAI cut off Cursor after its acquisition by SpaceX, citing Musk's history of breaking contracts. The lesson there wasn't really about Cursor. It was that access to a foundation model is not a permanent fixture of your workflow. It's a relationship that can be revoked based on facts that have nothing to do with your own conduct, and everything to do with who you're associated with or what your vendor decides matters.
Apply that logic to the Sony and Warner suit. If a court eventually finds that Anthropic's training data included unlicensed catalog material, the exposure doesn't stop at Anthropic. Agencies that have shipped client work built on Claude, embedded it in deliverables, or sold it to clients as part of a "responsible AI" pitch now have a paper trail running back through a vendor that is, at minimum, in active litigation over exactly that question.
- Ask any AI vendor directly whether they can produce a statement about training data provenance, not just a safety policy.
- Check the actual contract language around indemnification if a vendor's training practices are later found unlawful.
- Track litigation against your AI vendors the way you'd track a supplier's financial health, because the exposure runs in the same direction.
What this changes about buying decisions
None of this means agencies should panic or rip out tools they've built processes around. It means the due diligence conversation that used to end at "which model performs best" or "which one feels safest for client-facing use" now has to extend one layer further back. Whose work trained this. Was it licensed. What does the contract say if the answer to those questions turns out to be no.
Vijay Pande, reflecting on years of venture betting, put it plainly when he said his firm wasn't "doing 30 bets a year." The point applies here too: fewer, better-understood bets beat a portfolio of tools adopted because they were convenient. Agencies don't need to evaluate every AI vendor's entire training pipeline. But they do need to know that "trained on the open internet" and "trained on licensed catalog material" are not interchangeable claims, and that the difference is now the subject of active federal litigation, not a hypothetical.
The paper trail Sony and Warner are pointing to didn't appear because Anthropic did anything unusual for the industry. It appeared because two of the largest rights holders in music decided to make it visible. Other catalogs, other plaintiffs, and other models are watching the same case for the same reason agencies should be: to see exactly what "unlicensed" costs, and who ends up paying for it.