So, where did all that AI music come from?

Back in June, we wrote about who actually owns the copyright in royalty-free music. Near the end of that post, we mentioned AI-generated music and the question of what it’s trained on. We said we’d write a post about it, and although it took longer than expected, there’s more to say now than there was back in June.

What’s changed since then

In short, the big record labels have spent the past year signing agreements with AI music platforms, while some of them continue to pursue legal action against those platforms in court.

Universal Music Group settled with Udio and struck a licensing deal for a new AI platform. Warner Music Group did something similar with Suno. Sony hasn’t. Sony is still suing both Suno and Udio. One of those cases could decide what ‘fair use’ means for AI music, the argument AI companies have used to say training on copyrighted material without permission should be allowed.

Suno has also just announced download limits on its own platform, starting 3 September. Free accounts get up to seven trial downloads for life. Paid plans get a monthly cap. The company says it’s about stopping mass exports of music. Whatever the reason, it’s a sign that even the platforms leading this space are still working out what they’re doing, in public, as they go.

It also makes you think. Suno and Udio built huge platforms first and settled with rights holders afterwards, as the legal pressure mounted. If that route worked out for them, does it make the same approach worth the risk for whoever builds the next one? Nobody can answer that yet, but it’s hard not to wonder.

The bit that doesn’t get talked about much

Here’s what all those deals have in common: they’re about the major labels. Universal’s catalogue. Warner’s catalogue. Sony’s, eventually, probably.

No one negotiated on behalf of the independent artist who, let’s say, posted a track online in 2019 and has no idea whether it ended up training a model. Nobody sat at the table for the small production library whose catalogue might have been scraped before anyone thought to ask permission.

Some of them are trying to do something about it. Independent musicians have filed their own class actions against Suno and Udio, separate from the major label suits. At least one production music library has taken legal action of its own, over its catalogue turning up in an AI company’s training data without permission. Whether any of that goes anywhere is hard to say. But it tells you something: that the majors got a seat at the table and almost everyone else didn’t.

Not every AI training deal looks the same

Some platforms let composers opt in to having their music used for AI training, and pay them separately for it. If you don’t want your music used that way, you can say no.

That’s a different model from “we trained on it first and asked questions later”. Consent and compensation are what set it apart.

When it’s done with permission and fair pay, that’s the composer’s decision, and it’s a legitimate way for a platform to earn from its catalogue.

Two different calls

Licensing music for training pays the composers whose work it is. Some of those same platforms, though, won’t allow AI-generated tracks into their own marketplace. Keeping AI-generated tracks off the marketplace protects those same composers.

An AI can produce track after track in minutes, with none of the time or creative effort a composer puts into an original composition, and still compete for the same buyers.

There are two different calls here. One compensates composers for their work. The other protects them from being crowded out by copies of their own work.

An honest question

We’ve wondered about this ourselves. Music has to be heard somewhere to be found, whether that’s previews, streaming, or YouTube, which turned out to be one of the biggest sources these AI companies pulled from. Court filings and a source code leak this year both confirmed it, with millions of tracks scraped straight from YouTube to train some of these models.

Was a specific track, either yours or anyone’s, pulled into a set like that? There’s no way to know for certain. As far as we know, these companies haven’t published a full account of what went into their models, so there’s nothing for a small library or independent artist to actually check against. Unless the AI’s output ends up sounding enough like the original to give it away, which has already happened in some cases.

That’s not a claim that it happened to anyone’s music in particular. It’s just the honest, unresolved state of things right now, for anyone whose music has ever been public anywhere. And whose hasn’t?

Does it feed on itself?

Here’s something we keep wondering about. If AI-generated music starts appearing in libraries and gets used commercially, does any of it eventually get pulled into another company’s training data too?

This is a real, documented problem in AI research called model collapse. It happens when a model gets trained on data that includes earlier AI output rather than human-made work. Quality and variety drop with each generation, because the model keeps learning from a copy of a copy. It’s not specific to music. The mechanism doesn’t care what kind of content it’s applied to.

It doesn’t seem to be closely monitored for music yet, and maybe nobody can. It’s a fair thing to think about, though.

Little outfit, big outfit

If a huge company has trained its model on your music without asking, what can you do about it? Realistically, not much on your own. Litigation is expensive, and the companies with the deepest pockets are the ones with the least to fear from it.

The class actions exist because banding together is the only real leverage anyone without a major label behind them has. Whether that leverage adds up to anything is still an open question.

It’s also worth saying that independent music licensing looks different to how it did twenty years ago. There are far more players in this space now, some of them very large, and that makes it harder for smaller, independent outfits to compete on visibility alone.

One thing that hasn’t changed is what actually sits behind a smaller library. Twenty years ago, if you licensed a track, you were dealing with an actual person, whether that was the composer or someone who worked directly with them. That’s still true of several independent libraries today. It’s easy to lose sight of when so much of the conversation has moved on to what the technology can do.

Where does that leave you?

None of this is settled, and it probably won’t be for a while yet. But it’s a good reason to think about where the music you use actually comes from, and who agreed to be part of it.

That’s the real difference between AI-generated tracks built on scraped catalogues and a library where every composer agreed to include their music. In a library like that, you can trace it back to a person. With scraped AI output, right now, you mostly can’t.