On September 18, Sony Music and Universal Music Group filed a new lawsuit against Suno, the AI music company, in a federal court in Massachusetts. It is not the first time. In June 2024, they had already sued alongside Warner Music over the same issue, training AI models on protected recordings without a license.
What changes now is the core argument, and that argument has direct implications for any organization that manages copyright or related rights.
Suno released v6 on September 9, the latest in a model line that started with v4, launched as a collaboration deal with Warner Music, BMG, and Believe.
Sony and Universal are responding with a different argument than the one they used in the first suit. They are no longer arguing only about which recordings trained v6. They are arguing about how that training was built.
According to the complaint, v6 did not start from scratch. It was built using what Suno calls "user interactions" with its earlier models, that is, the choices millions of people made when the platform offered two versions of a generated song and they picked one. Those choices became the training material for the new model.
The problem, according to the plaintiffs, is that those earlier models were themselves trained on protected recordings, without a license. If the new model learns by observing what the old models produced, it inherits those same capabilities, even though it never had direct contact with the original files.
The complaint sums up the idea with a line already being quoted across coverage of the case, training a new model on the outputs of an infringing model does not eliminate the infringement, it transforms it into something else that is less visible.
To put it in terms that don't require technical knowledge, it is like a student learning to paint by copying other artists' work without permission, and then a second student learning to paint by looking only at the first student's paintings. The second student never touched the originals, but the style they reproduce still comes from there.
Why this matters for a collective management organization
CMOs exist to make sure that using a work generates payment for whoever created it. That link between use and payment depends on being able to identify the origin of what was used, and that is exactly what this case puts in question.
Until now, most of the debate around generative AI focused on whether a company used a specific catalog to train its system. This lawsuit suggests that question can go stale the moment the technology moves one generation forward. If a model trained irregularly becomes the learning base for the next one, and that one becomes the base for another, at some point it gets hard to trace which work produced which output. And without that trail, it gets hard to determine who is owed what.
This hits directly at the logic behind rights distribution systems. An AI model does not declare its sources the way a radio station declares which songs it played. Traceability, the foundation of any royalty split, is exactly what this kind of chained training puts in doubt.
The cost of infringement
US copyright law allows claims of up to $150,000 per work in cases of willful infringement. Across the 60,202 recordings Sony and Universal identified this time, using an audio fingerprinting company called Audible Magic, the theoretical maximum runs to about $9 billion, though the plaintiffs themselves note that figure is only a fraction of what they claim was infringed. This second lawsuit exists because a judge had already rejected, in August, adding these recordings to the original case. On top of that, there is an additional claim over the use of tools to download audio from YouTube while bypassing its protections, which could add another $150 million.
These are figures that will probably not be collected in full. Warner, BMG, and Believe already chose to settle on a license rather than keep litigating, Warner in November 2025, BMG this past August, and Believe one day before v6 launched. Sony and Universal, instead, chose the opposite path.
What's still unresolved
Suno argues that training models on protected material falls under fair use, the US legal doctrine that allows certain uses of protected works without permission, under specific conditions. The earlier Sony and Universal lawsuit, still ongoing, is the one that will decide whether that argument holds. Discovery in that case closes on September 30, though a final ruling could still take months. Whatever gets decided there will set the terrain for every case that follows, including this one.
What this second case adds to the conversation is that winning that first legal battle might not be enough. If a court finds that training on unlicensed material infringes rights, the next question is what happens to everything that model generated afterward, and to the later models that learned from those outputs. No court has a firm answer to that yet, and this case will likely be where one starts to take shape.