Anthropic sued over unlicensed song training

The lawsuit reads as a straightforward case of theft at scale: Sony Music Publishing and Warner Chappell allege Anthropic trained its Claude models on tens of thousands of copyrighted songs without permission, and they want billions for it. One notices that both plaintiffs are publishers, not labels - they own the words and the composition, not the master recordings. With that detail load-bearing, the case reads differently than the framing suggests, and not in Anthropic’s favour, but in a way the “AI steals music” headline actually undersells.

Publishers suing over lyrics is not a music case dressed as a copyright case. It is a text case dressed as a music case. A label suing over unauthorized use of a recording is arguing about audio - waveforms, samples, something a model would need to have processed as sound. A publisher suing over lyrics is arguing about strings of text that could have entered a training corpus the exact same way a New York Times article did: scraped, aggregated, sitting inside some web crawl of lyric sites, forum threads, or fan annotations that quoted the songs in full. That is a much easier claim to build and a much harder one for Anthropic to shrug off, because it does not require proving anything about audio pipelines at all. It requires proving the text existed somewhere in the training set, which is the same fight OpenAI has already had with the Times, just wearing a different genre’s clothing.

Here is the detail the framing keeps at the edge, because it is inconvenient for both sides: the complaint’s own contested point - whether the songs were actually used to train Claude, specifically - is not a factual claim yet. It is an inference from output. Nobody sued Anthropic because they found a training manifest listing forty thousand copyrighted lyric files. They sued, most likely, because Claude could be prompted into reproducing recognizable lyric passages, and reproduction was taken as proof of ingestion. That inference is usually right. It is not always right. A model can regurgitate a verbatim stanza it never saw as a discrete training example, because the stanza appeared quoted inside a thousand other pages - reviews, forum arguments, annotation sites like Genius, blog posts about “songs that changed my life” - and the model learned the pattern of the quotation, not the song. The distinction matters enormously for damages and barely at all for public perception, which is exactly why nobody in the coverage is dwelling on it.

So the plain question is not “did Anthropic use copyrighted songs.” Almost certainly some copyrighted text passed through some crawl that fed some training run, because that is how every large language model built before licensing regimes matured was assembled. The plain question is whether Anthropic can produce, or be compelled to produce, a dataset provenance record specific enough to show which songs entered training as discrete examples versus which ones the model absorbed as fragments inside other people’s commentary. If that record exists, this becomes a scoping exercise - count the songs, price the license, settle. If it does not exist, the case becomes a referendum on whether “we built a model that can do this” is itself the standard of liability, regardless of what anyone can prove about the pipeline, and that is a precedent that will outlast this particular defendant.

The people executing these training runs were not thinking about publishing rights when they built the crawler; they were thinking about token counts and dedup thresholds and whether the corpus was big enough to matter. That is not an excuse - it is the actual reason nobody kept the kind of manifest a court is now going to want. Building a system fast and building a system whose provenance survives a subpoena are two different disciplines, and almost no one doing frontier training in the last several years was rewarded for the second one. Sony and Warner Chappell are not really suing a music thief. They are suing a company that treated the internet’s text as ungoverned raw material and is now being asked, for the first time at this scale, to produce the paperwork that says otherwise. There usually isn’t any. That absence is the whole trial.