Generated music, and three separate legal questions

How a music generator works, briefly

Modern music generators do not synthesise waveforms directly. They work in a compressed representation: a neural audio codec encodes sound into a stream of discrete tokens at a manageable rate, a large model is trained to predict those tokens (from a text description, from an audio prompt, from both) and a decoder turns the predicted tokens back into audio. Diffusion-based systems do something analogous in a spectrogram or latent space.

The consequences you can hear follow from this. The model has learned the statistics of a very large corpus, so it produces things that sound like plausible music in a described style. It has no representation of song structure beyond what fits in its context, so long outputs wander. It has no notion of a performer’s intention, so a solo is a stylistically appropriate sequence rather than an argument. And it reproduces the biases of its corpus: styles that are heavily represented come out well, and styles that are not come out as approximations of something else.

Three separate legal questions

Public discussion runs them together. They have different answers, different jurisdictions and different degrees of settlement.

1. Was training on copyrighted material lawful?

This is actively contested. In the United States, several suits involving music and audio generators are working through the courts, with fair use as the central argument, and the outcomes so far do not form a single coherent rule. In the European Union, the Digital Single Market Directive provides text-and-data-mining exceptions, with Article 4 allowing rightsholders to reserve their rights, which makes the question turn on whether an opt-out was expressed and how. The United Kingdom has consulted repeatedly on the question and has not settled it. Japan’s exception is comparatively broad. Other jurisdictions differ again.

Anybody who tells you this question is resolved is describing their preferred outcome.

2. Does the output infringe?

Separate from training. A generated piece that closely reproduces a recognisable melody, hook or arrangement raises the same substantial-similarity question any other piece would. Models can and do memorise, particularly for material that appeared many times in training. This risk is about the specific output, not about the technology.

3. Who owns the output, if anyone?

The United States Copyright Office has stated that material generated by a machine without sufficient human authorship is not registrable, while works combining human authorship with machine-generated elements may be protected as to the human contribution. Several other jurisdictions have reached similar positions. The United Kingdom is unusual in having a statutory provision for computer-generated works with no human author, giving a 50-year term, a provision the government proposed in March 2026 to remove, though no law has changed yet.

So in some places a purely generated track may have no copyright owner at all, meaning you may have nothing to stop anybody else using it.

What this means in practice

You cannot resolve the law and you can manage your exposure.

• Read the service’s terms. They state what rights you are granted in the output, whether commercial use is permitted, and on what tier. They vary substantially and they change.
• Look for an indemnity, and read what it excludes. Some providers indemnify business customers against copyright claims arising from outputs, with conditions.
• Keep records. Which tool, which version, which prompt, which date, what you did to the output afterwards. If a question arises in two years, contemporaneous records are the difference between a defensible position and a guess.
• Check the output. For anything prominent, search for the melody. Tools exist for audio matching, and a hook that feels familiar often is.
• Do not assume the output is yours to license to others. Delivering a generated track to a client with a warranty of ownership may be a warranty you cannot honour.
• Expect platform enforcement to differ from the law. Content ID and its equivalents are matching systems operated by private companies under their own rules; they can flag your generated track because it resembles something in their database, and being legally in the clear does not stop that.

The disclosure question

Separate from copyright, a growing set of rules concerns telling people that something is synthetic. The European Union’s AI Act contains transparency obligations for certain synthetic content; several jurisdictions have or are considering rules aimed at synthetic media in elections and advertising. These are new, phased, and not uniform.

Two honest positions coexist and neither is wrong: that generated background music is a production tool no more requiring disclosure than a synthesiser, and that audiences have an interest in knowing. Where a rule applies, follow it. Where none does, decide deliberately rather than by default.

The free path

Open-weight music generation exists (MusicGen and AudioCraft from Meta, and Stable Audio Open, among others) and their licences differ, with some explicitly restricting commercial use. Read each one.

The alternative free path is not generation at all: Free Music Archive, ccMixter, and Creative Commons repositories contain large quantities of human-made music with clear licences, and a clear licence from a named person is a much simpler legal position than a generated track of uncertain provenance. The next lesson covers reading those licences properly.

The limitation

This lesson describes a moving position and will date. The cases in progress will produce rulings, legislatures will act, and the picture in two years will differ from the one here.

Nothing above is legal advice, and it cannot be: the answer depends on your jurisdiction, your use, your contract and facts nobody in a course knows. What is offered is the shape of the disagreement, so that you can recognise which question you are actually asking and know that a confident answer from any single source (including a vendor’s marketing) is a claim rather than a settlement.