If a given “Song B” copies, without license, protectable expression from an earlier “Song A” enough and in a way such that society should find it a trespass, that’s music copyright infringement.
On August 20, a judge refused to throw out Anthony Justice’s class action against Suno. He trimmed one weak count along the way, a mislabeled Tennessee consumer-protection claim, and let the DMCA claim ride the reasoning from his own UMG v. Suno order two days earlier. Neither turns on substantial similarity. The interesting part to me is what the plaintiffs never had to bring to the party. And I’m gonna harp on it a little, because it’s sustaining that silly reggaeton case too.
Justice is an independent country artist. He, his label 5th Wheel Records, and his publisher My Heartland Publishing say Suno pulled their catalog off YouTube, trained on it, and generates music substantially similar to their songs. A tortuous and meaningless sort of claim, but these days ordinary enough.
The amended complaint does not point to a single Suno output that sounds like a Justice song. Not one.
What it points to instead is a hundred or so supposed Suno outputs that imitate somebody else’s songs, and a quote from Suno’s CEO that the model can already produce outputs that “replicate real artists’ vocals and songs.”
The court held that’s enough at this stage. And maybe it should be? Enough for what?
Suno argued the plaintiffs had to identify the actual infringing work. Saylor said there’s no categorical rule requiring it, and borrowed the reasoning from Andersen v. Stability AI, the image case. It is a plausible inference, he wrote, that a model shown to replicate famous songs has likewise produced something substantially similar to these plaintiffs’.
I’m a musicologist, not a lawyer, and as a pleading matter, I don’t know how to process this. Is there a bar lower than plausibility? Suspicion? Is that next? There’s a musical assumption sitting in there, right? At what point do they need to produce it before there should be attorney’s fees? And do I have this straight? Suno was ruled to have memorized famous songs. In Germany at least. So Suno memorizes songs, and that by the way deserves an “in a sense” rejoinder. But it therefore plausibly memorized Justice’s.
How plausibly and to what degree and why do we care?
And it still doesn’t sit well that this middle step treats memorization as an agreed-upon property of the model when really it’s a property of how thoroughly a song saturates the training data and what gives the algorithm its dopamine at the outputs. Consider what it takes for a model to hand you a recognizable “Johnny B. Goode.” That song isn’t in there once. It’s in every cover, live version, karaoke track, backing track package, lesson video, and thousands of performances on the internet. Every one of them agrees with the others about what the song “is.” Meanwhile, a session player who never in their life heard the original record still nails the guitar riff, all four bars. It has a fame away from the original recording along with a duck walk.
Now consider an independent country record on a small label that maybe exists in the training data a little bit. No covers or karaoke libraries. The question this case eventually has to answer isn’t whether Suno regurgitates. We know it can, and pushed a bit, for the time being, still will. Do we not juxtapose that against whether it regurgitates this song? How plausible is it? What’s de minimis on plausibility?
This is a proposed class action. A questionable propensity for reproducing songs that are in the corpus a zillion times is the stand-in for however many independent songwriters’ catalogs, none of them individually examined I guess. I would question whether memorization is song-specific. If it isn’t, if it’s an ethereal condition across a class, I’m not sure what memorization means.
None of which is a defense of Suno really. Whether training on the catalog was lawful is an open and mostly separate question. Or it was until someone said “memorized” and muddied things.
Let me know when someone sits with a Justice track and a Suno output and does the truth-seeking work. I’d like to analyze what gets regurgitated when the time comes.
