What happened in September 2026
The largest provider in the field announced a new generation of models on a single day, and retired every earlier one on the same day. That on its own is instructive: anyone who had built a workflow around the sound of one particular model lost it overnight.
The bigger news, though, is that the new models were trained from scratch on licensed material, under agreements with several large labels and distributors. In the history of music AI that is a turning point: until now models mostly learned from material gathered without permission, and that is what the lawsuits are built on.
The difference between a licensed and an unlicensed model
The difference is not in the sound, it is in who stands behind you when somebody asks. With a model trained on licensed data, the cleanliness of the training material follows from the provider's contracts. With an unlicensed model it is an open question, and courts are working on the answer right now.
There is an important footnote that rarely gets said out loud: the deal is not complete. Two large rightsholders did not hand over their catalogues, and are still litigating. So a "licensed model" does not mean that every rightsholder of every piece of music agreed — it means that some of them did. That is considerably better than nothing, but it is not the same as being fully in the clear.
Who owns the generated music
Four layers are worth separating, because "the music is yours" means something different in each one.
- The provider assigns you whatever it holds. That is a contract question, and at the large providers it is generally in place.
- But it does not warrant that any copyright arises at all. That is the key sentence, and it sits right there in the terms. If nothing arises, there is nothing to assign.
- The American office's position is unambiguous: a prompt on its own does not establish authorship. The same logic as with pictures — which is what Who Owns an AI-Generated Image works through in detail.
- Commercial rights depend on the plan and on the format. They typically cover only the downloaded track, not a version that stays inside the platform or gets remixed.
The plan that decides what you may do
In practice this matters more than the legal theory. At the large providers the free plan grants no commercial rights — music made on it may not go into an ad, into client work, or into a monetised video. Paid plans do grant them, but with a limited number of downloads: around twenty a month at the entry level, around sixty further up.
Two things follow from that. One is that attempts do not count against you — downloads do, so it is worth auditioning inside the platform and downloading only the final take. The other is that once a subscription lapses, the rights acquired in previously downloaded tracks generally survive, but check that in the specific terms, because it differs by provider.
What the market numbers show
Figures published by one large streaming service show clearly where this market stands.
- Roughly ninety thousand AI-made tracks arrive on the platform every day.
- That is more than half of all new uploads.
- Actual listening, however, gives them only one to three percent.
- Supply exploded, demand did not — volume does not turn into attention.
Those four numbers together explain why mass-uploading AI music is not a business plan, and why the same tool is a good one for scoring your own content.
The lawsuits still running
The claim brought by the two large rightsholders carries real weight: it lists tens of thousands of specific works, and the theoretical damages run into the billions. The evidentiary phase closes in the autumn of 2026, so there is no decision on the merits yet.
One thing from this matters to you: the lawsuit runs between the model maker and the rightsholder, not with you. Your exposure does not come directly from there. It comes from the fact that if a model later turns out to have been trained on infringing material, the content you made with its music can become risky too. That is the argument for the licensed route, even where it costs more.
What to do in practice
- Check what your plan includes. In most places music made on a free plan cannot be used commercially — that is the single most common mistake.
- Download only the final take. The download count is the bottleneck, not the generating.
- Keep the paperwork. Which provider, which plan, when, on what terms. If somebody asks, that is your answer.
- Where the music is the product, do not work with AI. For scoring, background and mood it is excellent. A released record is a different situation.
Common misconceptions
- "If the provider assigns the rights, my music is protected." It assigns what it has. If no copyright arose, there is nothing to assign — and the terms say so.
- "A licensed model is completely clean." Partly: two large rightsholders' material is not in it, and they are suing.
- "I can use the free plan as long as I credit the source." Crediting does not substitute for a commercial licence.
- "AI music is flooding streaming, so commercially it works." Half of the uploads are AI, but only one to three percent of the listening.
What to take away
AI music is a real, cheap answer today for scoring, background and mood — if you work from a paid plan, with a downloaded track, and you keep the paperwork. What not to expect from it: exclusivity, copyright protection, and the idea that volume on its own brings attention. On the labelling duties, Labelling AI Video is the post that goes through them.