Can AI companies use copyrighted music without paying

Sony Music and Warner Chappell are suing Anthropic for copying tens of thousands of copyrighted songs into training data without permission or payment, demanding up to $150,000 per work — a bill that could reach several billion dollars if courts award maximum damages. This looks like a copyright case. It is actually a test of whether the legal framework built for Google Books in 2015 survives contact with generative AI. Music labels are banking hard that it does not.

Google scanned 20 million copyrighted books without asking permission from a single author or publisher. The company never disputed this. In 2015, after years of litigation, the courts ruled the copying legal under fair use because the scanning itself was transformative: it enabled search and snippet retrieval that served readers in ways the original books did not. The copying was extensive, unlicensed, and permitted anyway.

Music labels watched that ruling and understood immediately what it meant for their territory. If training an AI system on copyrighted material counts as transformative use, then Anthropic and OpenAI and Meta can absorb entire catalogs without licensing fees, the same way Google absorbed entire libraries. The labels cannot afford that precedent to hold.

The weakness in the music companies' position is real: showing that an LLM memorizes and regurgitates exact lyrics is harder than proving a search engine returns book snippets. But the weakness in Google's precedent is equally real. Google's fair-use win turned on the claim that scanning served a fundamentally different purpose than the original books. If you can show the model outputs direct substitutes for the training material — if Claude can be prompted to write a Beyoncé song rather than discuss it. The transformative-use defense collapses entirely. The labels know this. They are betting that somewhere in Anthropic's outputs, that substitution exists in provable form.

What matters is not whether the lawsuit succeeds. What a court decides fair use actually means when the copied material itself is the asset. Every system that claims to transform input into something new is now hostage to that definition. Google got away with copying because search was genuinely different from reading books. AI companies will not have that protection if courts decide training on copyrighted material means the model itself becomes a substitute for licensing it. The industry's risk is not damages. The risk is that the precedent shifts from "copying is permitted if you change how it is used" to "copying is only permitted if you cannot extract the original back out."