The AI music lawsuits are no longer about the models. They're about the session players.

The first wave of AI-music litigation was straightforward to describe: rights holders versus generators, over whether training on copyrighted recordings required permission.

That fight is still running. As of August 2026 Sony remains the only major label litigating against both major generators, with the Suno case in Massachusetts scheduled for dispositive motions in April 2027 (Chartlex).

But the more interesting case in 2026 does not have a generator as the defendant.

The union sued the labels

The American Federation of Musicians has sued Universal Music Group and Warner Music Group, alleging the labels licensed AFM-member session recordings to Suno and Udio for AI training without the compensation or consent required by the new-use provision of the union’s collective bargaining agreement (Chartlex).

Read that again, because the shape of it is unusual. This is not artists against AI companies. It is session musicians against the labels that employed them, over what those labels then did with recordings the musicians played on.

Why that shift matters

The training-data argument has been conducted almost entirely between two groups of large organisations. Model companies say training is transformative. Rights holders say it is infringement at scale. Both are well resourced and both are arguing about a corpus, not about people.

The AFM suit reframes it. It asks a narrower and much harder question: when a label licenses its catalogue for AI training, what is owed to the people who actually performed on those recordings?

Session players are typically paid a fee, sometimes with contractual provisions covering new uses of the recording. AI training is exactly the kind of use those provisions were written to catch, and nobody drafting them in previous decades had this in mind.

If that argument succeeds, licensing a catalogue for training stops being a decision a label can make alone. It becomes a negotiation with everyone who played on it — which is a very different cost structure from the one the current deals assume.

What this means for anyone making AI music

The instinct is to treat all of this as someone else’s problem: big companies, big lawyers, years of appeals.

That instinct is half right. The outcomes will not change what tools are available next month. But they will shape two things that matter to anyone building on this:

What the models are trained on. If licensed training data becomes the norm because unlicensed training becomes expensive, the models available in two years will have been built differently. That has consequences for what they sound like and what they cost.

Whether provenance becomes standard. Watermarking and labelling are being built for platform transparency, but the same infrastructure answers a legal question about where a piece of audio came from. Those two motivations are converging fast.

The position that ages well

There is no comfortable seat here. Anyone making music with generative tools is using something built on a corpus whose licensing status is genuinely unresolved, and pretending otherwise is not honest.

What is available is the same thing that is available on labelling: be clear about what you did. Say the music is AI-generated. Say what the human contribution actually was. Do not claim a process you did not follow.

That does not settle the training-data question — nothing an individual artist does settles it. But it means that when the rules land, whichever way they land, you are not discovered to have been misrepresenting the work. You were describing it accurately the whole time, and the rules simply catch up to a description that was already true.