The headline is GEMA had a good day: a cease-and-desist, disclosure of some Suno earnings, and damages to be figured out later—generally, its claims were granted.
Eight days ago I wrote that a Munich courtroom was about to do something no American court has done yet and I promised we’d grade the court together when it did. It did. So here goes.
Quick caveat: the written judgment is not public yet. We’re going on what we read in the news, on LinkedIn, some courtroom reporting, and best of all that, same-day notes from eyes and ears in the room. For now, we’re looking at the court’s reasoning through that reporting, beginning with the main points. Also, I know these nano banana images are gonna age poorly. But right now tho?
The court said lots of “yes” to GEMA. It found Suno pulled songs, namely Daddy Cool, Rasputin, Forever Young, Big in Japan, Atemlos, and Mambo No. 5, from YouTube by stream-ripping, circumventing the “rolling cipher” that tries to prevent downloading. Not a great beginning, and we haven’t gotten to the training part yet.
If there’s a steamroller, it’s that the court held the six songs were “memorized.” Fixation, they figure, doesn’t require an identical stored file. As long as the work produced is perceptibly similar (my words) through simple prompts to anyone with full access, that was memorization. The evidence was the outputs. The similarities between the originals and what Suno generated were, in the court’s reported words, “too extensive and specific to be coincidental,” across melody, harmony, and rhythm. That one finding did most of the work in this case. Once because the works were effectively, if not literally, retained rather than merely analyzed, and again, the rolling cipher thing—that Suno’s access to them in the first place is iffy. You have a new finding, not quite the training questions as I’ve seen them to this point.
Next, since the actual training took place in America, the German court sought to apply American law, and not in Suno’s favor. Again, the idea of “memorization” based on the outputs does a lot of work. Prior US decisions finding AI training to be fair use because the work is transformative get flipped on their head if the material is deemed memorized and reproducible. Then the fourth fair use factor falls similarly, necessarily—the reproduced work is a substitution in the market. And that all sounds logical, but relies on the “memorization” premise.
The judgment is not final, and Suno, naturally, says it disagrees and will evaluate its options, including an appeal.
GEMA’s CEO called it “a verdict of global significance.” Not remotely alone. There’s plenty of rejoicing going on by “training is theft” pundits who finally have a major victory. And maybe that’s a step in a good direction. I’m reserving judgment.
The Munich court’s entire case against Suno stands substantially on that GEMA prompted each song the way a plaintiff seemingly should: original lyrics, a style, a title. No melody specified, and the melodies came back anyway! If the melody comes out and nobody put it in, the melody was in the model — “Memorized,” the court called it. That foundational finding cascades, beating Germany’s text-and-data-mining exception, undercutting America’s fair-use rulings, and handing GEMA nearly every claim it brought.
But what if I tell Suno something along the lines of, “anticipate the last word of each couplet on the and of four. Strongly favor static arcs, and open intervals, avoiding thirds of accompanying chords, and restrict yourself to an octave.” I just described the actual behavior of Forever Young’s melody, and I never wrote a note. Neither instruction touches “no melody specified.” Both move toward handing the melody over anyway. We don’t need note names to allude to what to do with the limited universe of possibilities. Now imagine I keep sharpening my prompts in that vein, round over round, until I get what I want. To my knowledge, nobody outside the case has seen what GEMA’s “style” field actually said. (Correct me, please.)
I’m not claiming that’s what GEMA did. I’m wondering why nobody seems to have ruled it out. I do not speak German but Suno’s pleaded defense evidently called the plaintiff’s prompts complex and iteratively refined. The court seems to disagree, and I might too if I’m shown the work.
One bit of pushback I’d expect is that the style field is a suggestion box, not quite gospel. Suno doesn’t obey. But nobody needs the field to obey if we’re patient with it. We just need it to lean. Move the other variables: A description that merely raises the odds of anticipated phrase endings and certain melody arc particulars in prose let’s say, run enough times, with me selecting the keeper and running my selection and convergence machine. I hang a “Humble steering, large sample, hard filter” sign above my computer.
A same-day account of the hearing says GEMA’s own argument put numbers on the sampling: four prompts for Mambo No. 5, a hundred twenty-four for Big in Japan, a hundred seventy-six for Atemlos. Secondhand until the judgment text confirms it. But if those numbers hold, the unknowns stop being separate quibbles. Style field, generation count, selection. Each one multiplies the other two. And they don’t all point the same way. Four pulls is no sample at all, and if Mambo No. 5 really came back in four, that’s the number that should worry Suno. But nothing about a hundred seventy-six dissuades me from wondering what happened here, nor should it comfort GEMA.
And Suno can build on its own output, saving a previous generation’s melody and structure while you change other parameters, in a chain sharpening your style prompt each round. A hundred seventy-six fresh pulls would be more of a slot machine and function of patience and knowing when to quit. A hundred seventy-six progressive refinements is navigating to a known destination.
I’m being belligerent and hypothesizing. (duh) Accounts I’ve seen don’t tell me whether any or all generations in a hundred seventy-six were built on a previous one. The two produce evidence of completely different weight. Nobody has said which this was.
GEMA’s published exhibits were evidently prepared by its own staff musicologist. I could be wrong here too, but that staffer is the only musicologist I know to have been involved. A firsthand account of the March hearing describes the chamber playing the originals and the Suno versions in full, in the courtroom, and judging them by ear. Suno’s lawyers objected that the works should be compared as pure piano versions. And that’s smart, but also a musicologically informed objection. I almost certainly would’ve recommended it, but unless it was really well delivered, it may have read as a dodgy debating move. It apparently went nowhere and the biggest AI-music ruling so far was decided by judicial listening, unaided. Courts are courts, and musicology is musicology. Different lanes. Only one of them evidently occupied. Old territory for Musicologize reviews. But caveat time: written submissions aren’t public, so an expert opinion could be in there somewhere.
I DID read the exhibits. Here’s what holds.
I pulled GEMA’s own published exhibits, the side-by-side leadsheets of AI output against original, and went through a few; Daddy Cool, Forever Young, Big in Japan.
Don’t let me give the impression they’re not close, closer than mere shared vocabulary and style comfortably explains. Elements more or less track bar for bar, the same phrase shapes landing on the same words in the same places. Individually, plenty of it is trite and I can highlight so much that’s generic and belongs to nobody. But I’m not. A long series of trite gestures, reproduced in the same order, stops being trite as a series. So, I don’t want think the mp3’s and transcriptions aren’t similar. If two human composers showed me this much correspondence, phrase after phrase, it would hardly be interesting. They’re very much the same. That’s meaningful only as far as it goes. With a little scrutiny, it frays.
Where the versions diverge, and what that’s worth forensically
Divergence is where I look after considering the similarities and their forensic value. For example, at the end of an eight-bar phrase in Forever Young, the original as shown by the plaintiffs, goes to Am where Suno goes to G. A real difference but not very persuasive at that structural spot in this idiom. Daddy Cool’s chorus goes from Cm, a disco-flavored minor dominant option, whereas Suno went with C7, the more answer instead. The model is reaching for the expected chord over the one that would point to memorization. Forever Young is fairly reliant upon a diatonic Dm that in Suno is thirdless, open, neither explicitly major or minor, but distinctly leaning major where the key calls for and expects minor. It would sound jarring to a Forever Young afficionado. What is that? Guessing at Suno’s mechanism is just part of my hypothesis; it’s not analysis. I’d just like to know a lot more before I invest in this ruling.
As things are, taking the evidence and the accounts, I’m only a square or two from where I was a week ago. Output similarity proves memorization only if the similar material couldn’t have come from vocabulary. These six songs make the test harder than it looks. These aren’t complex works to begin with, and their production elements aren’t unique to their masters. They get covered, performed, and even sampled. A session player who’s never heard the original can still nail the riffs because the riffs left the recordings a long time ago. The correspondence can be real and the memorization inference still wrong.
And indeed, that’s a lot of forensic musicology and originality clearance. It’s the stuff of influence in the absence of either memorization or copying.
On August 23rd I wrote that a very viral Johnny B. Goode video proved less than its maker thought, because the system run long enough will eventually hand you something recognizable. I did not expect to be pointing that sentence at the winning side’s evidence so soon after. But recognizable still isn’t the bar, and wow, a hundred twenty-four pulls is a long stay at my slot machine.
GEMA’s lawsuit page says it “selected some of the audio content created for the test case.” Let’s be skeptical again. Selected. Its own examples page publishes nine leadsheet pairs, not six: the six it sued over, plus Cheri Cheri Lady, Brother Louie, and You’re My Heart, You’re My Soul, evidently tested but not taken to court. Something about those three didn’t clear whatever bar the six did?
The ruling could be right. None of this makes the ruling wrong. This court wasn’t naive. Suno’s lawyers made the cherry-picking argument, the chamber heard it and rejected it. But I’m skeptical. The style field, the generation counts, and perhaps progressive selection. Is it so hard to imagine the court was reasonably and legitimately convinced, on a slice of evidence a scientist wouldn’t accept as proof? We’ll see what the inevitable appeal says.
Musicologists, have you tried…?
Nevermind. It would take too long to write the “have you tried” list.
How much of last week’s pt. 1 held up?
Most. Let’s jump to what was a surprise. I expected that a GEMA win would establish something significant in the EU and “nothing especially impactful about US cases.” Well, this ruling isn’t binding on an American court certainly, but I didn’t expect getting into US doctrine at all, and a written opinion distinguishing fair use findings in the US, is exactly the reasoning American plaintiffs will be waving around. Edward Lee at Santa Clara had his analysis up within a couple of hours under the title “Throwing Comity to the Wind,” asking whether a Munich chamber should be grading American fair use at all, when the EU itself argued for territorial restraint in Abitron v. Hetronic. A German court reaching for America’s own law to judge the American conduct makes sense to me. But I’m not here to referee comity. Chat GPT Is Eating The World will do it better. I’m here because when a court says a model memorizes six songs, somebody should ask the musicology.
At any rate, if it survives appeal, every model trained the way I imagine them all to have been then has a Germany problem. They have one now. And every brand and platform using AI-generated music has a precedent to think about. This, though- what I flagged on the 23rd as surviving any outcome- does: when an AI output resembles a real song, is what’s shared protectable expression or common vocabulary? Munich resolved the case on retention, albeit with the gaps I propose remain. Understandably, creators have waited a long time for a ruling that reads like rage vindication. Next week’s reaction is bound to be pretty tribal, a creators’ side and a tech side.
Musicologize has, for over a decade on this site been telling plaintiffs their cases were silly and telling defendants their songs were copies, whichever way the notes pointed. That’s forensic musicology. We’re really NOT supposed to have a side. The best clients just want the facts. The notes here point in uncomfortable directions for everybody. The outputs are close enough that I’d certainly never wave them off, but produced by a process I’ve no faith was audited well.
Friendly arguments welcome. Especially to this: If I memorize something, it doesn’t take me one hundred and seventy-six attempts to reproduce something pretty close. Or really, even four. And don’t the implications of this whole thing turn on that?
