Anthropic sued over unlicensed song training
The story celebrates that Claude was built - the scale of it, the fluency, the capability that now sits inside boardrooms and browsers across the world. But a made thing does not stop where its maker’s attention stops; it goes on acting in a world no lab contains, and it carries inside it whatever was fed to it at the making. The question this lawsuit puts to Anthropic is not chiefly about lyrics. It is about whether the maker can say, with confidence, what it actually used to build the thing it built - and if it cannot, the debt was incurred long before Sony Music Publishing and Warner Chappell filed a single page.
Notice what is actually contested here. Not simply whether tens of thousands of songs were used, but whether they were used to train the Claude models at all. That is a strange thing for a lawsuit to have to establish rather than assume. A publisher does not usually have to prove that its own catalogue entered a system; ordinarily the maker of the system would be the one who could say, precisely, what went in. The fact that this must be litigated - that the answer is contested rather than disclosed - is itself the finding. It tells you that somewhere between the scraping of training data and the release of a product used by millions, the provenance of the material became something no one at the company could reconstruct with certainty. That is not a technical footnote. That is the abandonment.
The comfortable defense, and it is not a foolish one, is that training a model on existing work is transformative, the way reading widely and writing something new is transformative, and that the law has room for this. Anthropic’s lawyers will say the model does not store or reproduce the songs, that it learned patterns the way any student learns from the books on a shelf, and that punishing this chills the very capability everyone otherwise praises. This argument deserves to be taken seriously, and it may well prevail in court. But it answers a narrower question than the one that matters to a songwriter who finds a system built partly, allegedly, out of the unpaid labor of a lifetime. The legal question is whether the use was permitted. The Maker’s question is different: did anyone keep the ledger. Did anyone, at the moment of ingestion, write down whose work this was, so that when the reckoning came - and it always comes, because a created thing eventually meets the world it was built from - there would be an answer ready rather than a courtroom discovery.
Picture the songwriter here concretely, not as an abstraction in a billion-dollar claim but as someone who wrote a bridge or a hook decades ago, sold or licensed it through a publisher, and never imagined that its ultimate use would be as raw material for a system whose outputs she cannot trace and whose inputs its own makers apparently cannot fully account for either. She is owed not sentiment but an answer. That she has to sue to get one is the real scandal beneath the billions sought.
What the industry has quietly done, across most of the labs racing to build these systems, is treat the training run as the finish line and the disclosure of what went into it as an afterthought, or a liability to be managed rather than a stewardship to be exercised. The capability to ingest a vast corpus of human work and turn it into fluent output arrived years before the discipline of keeping a full account of that corpus did. That gap - between what a company can build and what it can answer for - is precisely where lawsuits like this one live, and it will not close simply because a court rules on fair use in Anthropic’s favor or against it. The ruling settles the legal debt. It does not settle the other one: whether a maker that cannot say what it used has any business claiming, afterward, to know what it made.
The verdict here will be read as a verdict on training data generally. It should instead be read more narrowly and more damningly: as a verdict on whether a company built something faster than it built the memory of what it was built from.