Whose music it learned from
Why music moved first
Music has three features that made this the sharpest front.
Concentrated ownership. Three major labels control a large share of recorded music, which means a small number of parties can act together and afford to litigate.
Existing licensing machinery. The industry already has collecting societies, mechanical licences and sync licensing. There is an established answer to "how do you pay for using this", so "we could not have licensed it" is a weaker argument here than elsewhere.
Precedent for tight infringement standards. Music litigation has long treated short similarities seriously, which the next lesson covers.
What has actually happened
The pattern, described without picking a side.
Major labels brought infringement claims against leading music-generation companies in 2024, alleging that training on their recordings without licence was infringement at scale. Collecting societies in Europe brought parallel claims, and a German court ruled against a large model developer over song lyrics in 2025.
During 2025 several of those disputes moved toward licensing rather than judgment: label groups announced arrangements with generation companies covering training and, in some cases, artist participation. Other proceedings continued.
That combination — litigate, then license — is a familiar shape in music, and it happened before with sampling and with streaming. It is not a ruling on whether training is infringement. It is a commercial settlement of a question the courts did not have to reach, which leaves the legal position genuinely unresolved while the market moves on.
What "licensed" means, and what to check
"Trained on licensed data" is now a marketing claim, and it covers several different things:
- Fully licensed catalogue, with rights cleared for training and for output, and payment flowing to rights holders.
- Public domain and openly licensed material only, which is a real and verifiable position and constrains the sound.
- Licensed from a stock library whose own contributor terms may or may not have permitted training — a gap that has caught several providers.
- Licensed for training but silent about output, which leaves you exposed on the thing you actually care about.
The questions worth asking a vendor: what exactly was licensed, from whom, does the licence cover the output as well as the training, do you indemnify commercial use, and what are the conditions on that indemnity. The final module covers what an indemnity is actually worth.
The artists' position, stated fairly
Musicians' objections are not all the same objection, and conflating them makes the argument unproductive.
Some object to uncompensated use of their recordings as training input. This is a licensing dispute and it is the one that settles.
Some object to style imitation — a model producing tracks in a recognisable manner. Style is not protected by copyright almost anywhere, so this is a moral objection with no legal instrument behind it, which is exactly why it feels unanswerable to those making it.
Some object to voice cloning, which is a personality-rights question rather than a copyright one, and which several jurisdictions have legislated on specifically.
Some object to displacement — that generated library music takes the work that paid the rent, regardless of any rights question.
These need different answers and only the first is on a path to one.
Where this leaves you
If you use generated music commercially, the practical position today:
- Read the licence on the model and on the output. Several capable open models are non-commercial only.
- Prefer providers who state their training sources and indemnify output, and read the conditions.
- Keep a record of what you generated, with what, and when. If a question arises in two years it will be about a specific track.
- For anything high-value or long-lived, consider licensed library music or a commissioned composer instead. The cost is not as different as people assume, and the rights position is settled.
The unresolved part, plainly: whether training on copyrighted recordings without licence is lawful has not been decided consistently anywhere, courts in different countries have begun to answer adjacent questions differently, and the commercial settlements have removed several of the cases that would have produced an answer. Anybody who tells you this is settled is describing their jurisdiction, their preference, or their sales position.
The one thing to keep
The music industry's dispute over training data is further advanced than in any other medium, and it has produced both litigation and licensing, which makes the shape of the disagreement unusually visible.
Before you move on
Why does the wave of settlements between labels and music-generation companies leave the legal question unresolved?
Pick the one you would defend. Nobody sees your answer.