Sony Music, Warner and thirty-three other publishers just sued Anthropic: tens of thousands of songs, allegedly pulled through pirate sites and scraped off licensed lyric sites, used to train Claude. Odd lawsuit at first glance—no audio comes out of Claude, this is lyrics as text, feeding a language model the same way books do.
The case leans on a distinction that hardened last year, when Anthropic settled a separate suit over pirated books for $1.5 billion: training on lawfully acquired work can be fair use, but that doesn't clean up how the copy was obtained. Not that interesting to argue who's right here—this will probably settle like the book case did.
For my part, I find it more interesting to zoom out a few decades, to the moment the industry now suing Anthropic first got built.
In 1908, music publishers filed an earlier version of the same lawsuit, over an earlier machine—a maker of player-piano perforated rolls, accused of reproducing their sheet music without permission. They lost. The Supreme Court ruled that a roll wasn't a copy of the music, only an encoding of it: holes a machine could read and no human could, no more the song than a negative is the scene it captured. Congress stepped in a year later, not by giving composers the control the Court had denied them, but by inventing a mandatory license instead: two cents a roll, owed regardless, while stripping publishers of any power to block the format—the direct ancestor of the mechanical royalty publishers still collect. The industry now asking a court to stop Anthropic outright has made this argument once before, and lost.
The question over encodings kept resurfacing. MIDI files caused a near-identical fight decades later—not audio, just instructions a synthesizer could follow. What a model keeps after training isn't the lyric either, just a trace of one spread across billions of numbers—the latest, least legible encoding yet.
Zoom out further and it stops being a story about copies at all. Printing industrialized reproduction, the internet industrialized distribution, generative AI industrializes generation—and at each shift, whoever controlled the prior infrastructure had reason to call the next one theft. None of that makes what Sony and Warner allege illegitimate. It just means the publishers pressing it are themselves a machine built, over a century, to aggregate and monetize somebody else's creative output—no version of this fight has really been creators against a capturing industry, just industries taking turns.
Eventually, this isn't so much artists against AI, perhaps another round of a match that's been played more than once before, over who gets to tax human creation on its way to an audience.
