Anthropic sued over unlicensed song training

Francis Bacon applies The Instrument of Progress.

We are told that to train a model on a song is already, in the full and settled sense, to use the song - that ingestion is appropriation, fixed as the melody itself, and no further inquiry is needed. This is the premise on which Sony Music Publishing and Warner Chappell now bring suit against Anthropic, alleging tens of thousands of copyrighted works were folded into the training of its Claude models without leave or payment. But notice the instrument that has actually arrived in the world: a system built by ingesting vast bodies of composed language and extracting from them not the works themselves but the patterns beneath the works - rhyme scheme, narrative device, the architecture of a bridge before a chorus returns. What it makes newly possible is generation from pattern rather than reproduction of text. The question is not whether to defer to the old frame, in which any exposure counts as theft, but what experiment would tell us whether that frame survives contact with what the model actually does.

This is precisely where the complaint’s own contested facts matter more than its rhetoric. The filing alleges tens of thousands of songs; it does not, on the record available, demonstrate that Claude was trained on them in any way that produces the works on demand. A number in a legal brief is not evidence of mechanism - it is an estimate of scale offered before the instrument itself has been examined. Bacon’s method does not ask which side’s press release to believe. It asks for the audit: open the training corpus, or failing that, interrogate the output. Can Claude, prompted plainly, return the second verse of a specific copyrighted lyric word for word, beyond the sliver fair use has always tolerated? That is a test with a real answer, obtainable now, and it settles more than either party’s characterization of “tens of thousands.”

The strongest version of the publishers’ case does not rest on the number at all. It rests on the claim that even pattern-extraction, at sufficient scale, does economic harm to songwriters whose craft is precisely those patterns - that a model fluent in the architecture of pop composition erodes the market for composers regardless of whether any single lyric is reproduced. This deserves to be taken seriously, because it names something closer to a demonstrable harm than the copying allegation does: a substitution effect in commissioned songwriting, felt by session writers and jingle composers long before any court rules. But notice what this argument concedes - it has quietly moved off the ground of copying and onto the ground of competition, and competition from a new instrument is not, by itself, a legal wrong; it is the oldest complaint every craft makes against the tool that learns its trade faster than an apprentice could. The remedy for that harm, if real, is a licensing market for pattern-training, not a ruling that pattern-training is theft.

Picture the actual person this case is fought over: not Anthropic’s lawyers, not Warner Chappell’s catalogue managers, but a working songwriter checking, alone at her laptop, whether a chatbot can hand back the chorus she wrote for a jingle three years ago and never got paid twice for. If it can, she has a grievance the law already recognizes, verbatim reproduction, and no talk of pattern or architecture rescues the defendant. If it cannot - if what the model gives back is only the shape of a bridge, the timing of a hook, the thing every songwriting class already teaches - then the suit is asking the courts to declare a method of learning illegal because it resembles, at industrial scale, what every apprentice does by ear.

Billions in damages is a claim about the size of an injury, not a description of how the injury occurred, and no sum stated in a complaint has ever been an experiment. The case will not be decided by the figure of tens of thousands, nor by Anthropic’s silence about its training data, but by whatever surfaces in discovery when someone finally sits down, prompts the model, and waits to see whether the chorus that comes back belongs to a songwriter who never signed anything at all.