In 1908 America's Supreme Court ruled a piano roll is not a copy of a song — no human can read it. This week, 118 years later, a Munich judge looked at neural-network weights — billions of numbers no human can read — and ruled the opposite: if the song can be recovered from those numbers, they are a copy.

Between those two sentences sits the entire machinery of how music gets paid. This week it collided with Suno — a $5 billion company whose machine, by every appearance, learned from an enormous slice of recorded music — and the German royalties society won in court on almost every point. It is the fifth time in a little over a century that a new machine has broken the meter that bills for music.

GEMA — collecting royalties for 100,000+ composers — sued Suno over six songs, including Rasputin, Daddy Cool and Mambo No. 5; the court found Suno's outputs confusingly similar to the originals. The findings, in plain words: storing songs inside a model's weights, in a form they can be pulled back out, counts as making a copy — so training that swallows them whole, without permission, broke the law; and generating near-identical tracks breaks it again, as unauthorised publishing. Remedy: stop, open the books, pay — the amount to be set separately.

Two caveats, both important: this is a first-instance ruling, and Suno disagrees and is weighing an appeal; and it is not a ban on AI music — the logic is about licensing and payment. Still, the milestone is real: the first ruling in Europe that a generative AI company must license the music it learns from — and the second time this court has said something like it, after OpenAI in November. That case is on appeal too.

Every payment in music is triggered by a countable event: a sheet sold, a record pressed, a radio play logged, a stream ticked. Collecting societies are essentially meter readers — they watch the counters and split the money. And every counter meters the same underlying atom: the copy. Three hundred years of copyright law, refined into machinery for counting reproductions.

Now watch a learning model eat the catalogue. It doesn't file songs away as files; it reads millions of them and nudges billions of internal numbers, each by a hair. Nothing is pressed, played, or streamed — the meter reads zero, while the machine has, in a real sense, drunk the entire catalogue. The value clearly moved; the atom the law knows how to count never occurred. Not a villain — a meter built for one physics, watching a machine that runs on another.

When the atom breaks, courts stretch old words over new objects. Munich's stretch: Europe's 2019 rule allows machine reading of text and data unless the rights holder opts out — and the court accepted that AI training, as such, can fall under it. But it drew a line at memorisation: a song recoverable from the weights is stored, and storage — by a definition first written in the tape-recorder era — is a copy. Even the opt-out has a story: GEMA raised its hand, and the hand took a year of paperwork to become legally visible. No standard registry, little way to police it — protection mostly on paper.

Meanwhile the same machine gets different answers elsewhere: American courts are split over whether training counts as fair use — their doctrine that some copying is allowed without asking — and Japan, that same year, wrote the opposite rule on purpose: training broadly free, liability only at the output. Munich: infringement. America: contested. Tokyo: legal. One global machine, three legal realities — which rounds to none.

This has happened four times before, and the fix was never to ban the machine. The piano roll broke the meter first; after 1908's "not a copy," Congress answered within a year: the compulsory license — anyone may record any song at a fixed 2¢. Radio broke it next — you cannot count copies in a million living rooms — and the blanket license moved metering from the copy to the doorway.

Then the tape recorder — and this chapter belongs to Germany and the same plaintiff. GEMA asked courts to make recorder makers identify buyers; the court refused: privacy. So in 1965 parliament invented the levy — a small charge on every recorder sold, into a fund for rights holders. First in the world; dozens of countries copied it. The sampler broke the meter again in the nineties; the answer was clearance. Four repairs — every one an institution for counting collectively what could no longer be counted individually.

One license for the whole catalogue (radio's playbook): societies sell AI companies the right to train on everything they represent, paid as a revenue share before training starts. The cost hides in the split: a song's share of the weights can't be measured, so money divides by market share — biggest catalogues first, small songwriters last, non-members never; and societies are national while models are global.

A levy on the machine (Germany's 1965 invention, re-aimed): a fixed slice of every generative service's revenue into a common fund, split by rule. The cost is bluntness: the same slice whether you trained on licensed data, scraped data, or no music at all; distribution still runs on proxies; and levies are famously immortal.

Charge only the outputs (Japan's line): training free by statute; a bill arrives only when a generated track resembles a specific song. The cost sits at the entrance it waives: the catalogue's value transfers free, forever — and the better the filter, the smaller the bill. It rewards hiding the resemblance, not licensing the source.

The EU's current opt-out belongs in the defect column, not this list: a right with no standard, no registry and no enforcement is why six songs took eighteen months in court — in a world of a hundred million songs.

Two courts, a bit more than a century apart, looked at a music machine no human could read. One said: not a copy. The other said: a copy. Neither answer ever settled anything on its own — the real work always happened afterwards, in parliaments, when someone designed the next meter. Munich has said — for now, and an appeals court may yet disagree — that the machine owes something. It has not decided how the bill gets computed for everyone: one license for the whole catalogue, a levy on the machine, or a charge on the outputs.

So when the next machine breaks a meter — and there is always a next machine — you already know the question that matters. Not "is this theft, or is this progress?" — that argument has been shouted at every machine since the player piano. Ask instead: what does the new meter look like, and who gets to read it?