On 10 July, eight organisations announced they had agreed on how to label AI music. The IFPI, the RIAA, A2IM, WIN, IMPALA, the Grammys, SAG-AFTRA and the Human Artistry Campaign put their names to two tags: "AI-Generated" and "AI-Assisted."

Getting that many bodies to agree on anything is hard but there are a few things we should be aware of.  

When you read the rules closely and you find something odd. A track counts as AI-Generated if AI produced the lead vocal or the primary instruments. It counts as AI-Assisted if humans performed those parts and AI filled in around them. Lyrics and composition are not covered at all. A song written entirely by a machine stays un-labeled, as long as a human sings it.

This tells you exactly what part of our creative process the industry chose to solve. Currently we are looking at performance. The labelling system won’t tell you anything about ownership.  

What the label tells a brand, and what it doesn't

Say a brand wants to use an AI-assisted track in a campaign. The tag reads "AI-Assisted." The brand now knows a human sang the lead.

It still does not know who wrote the song, whether the model was trained on work someone else owns, whether the uploader had the right to license it, or who to send the money to. Every question that has to be answered before a licence can exist is untouched by the tag.

Compare that to a conventional release. A song has writers, publishers, performers and a master owner. The splits are registered somewhere. The data is often wrong and the registries often disagree, but a chain exists and you can follow it. That chain is what makes clearance possible at all.

AI music arrives without one. A track can be generated from a prompt, uploaded, and monetised with nothing attached to it but a filename. The model behind it was trained on work that had owners, none of whom appear anywhere in the output. The chain of custody does not break. It never gets created in the first place. ‍

The volume is what makes it structural

Deezer is the only major platform publishing its numbers, and in April it reported receiving around 75,000 fully AI-generated tracks a day. That is 44% of everything uploaded to the service.

Spotify, Apple Music and Amazon have not published equivalent figures. Apple's approach, launched in March, asks labels and distributors to declare AI content when they deliver it. Spotify is backing an industry disclosure method in the credits. Both put the burden on whoever is uploading, which means both systems inherit whatever that person chooses to say.

This is our honest take. Roughly half of new music arriving at the one platform that measures it has no verified human author, and the industry's answer is a voluntary tag that its own rules exempt if a human sings the top line...

Make that make sense.  

Why this costs real money

None of this would matter much if AI tracks stayed un-played. On Deezer they mostly do: AI music is 1 to 3% of streams, and 85% of those streams are flagged as fraudulent and demonetised.

The damage lands somewhere else. Every commercial use of an un-licensable track is a licence that never gets bought. A brand chooses an AI track precisely because it looks frictionless, and the rights holders who should have been paid are never in the conversation. CISAC and PMP Strategy put nearly a quarter of creators' revenue at risk by 2028, as much as €4 billion. That figure is not about AI making better songs. It is about AI making cheaper ones that nobody has to clear.

Our own audit puts the unlicensed commercial music market at £12 billion a year. AI does not create that gap. It widens it, by making the frictionless option more available than it has ever been.

What sits upstream

Labelling is downstream work. It describes a finished track and asks whoever uploaded it to be honest about how it was made.

The alternative is to record the process while it happens. If an AI tool captured the human input, the prompt, the model and the output at the moment of creation, that record would travel with the track. Not a description of the result, but a history of it. Connect that history to a licensing layer and the track becomes clearable, because you can finally answer the question the tag skips: who is responsible for this, and who gets paid when someone uses it.

A system that records provenance at the point of creation still depends on people telling the truth. They can still lie. A record that exists and can be checked is a different world from no record at all, and the same is true of every rights registry we already rely on. The whole points is to make the honest path the faster and easier one.

The choice ahead

The eight organisations behind the labels did the harder thing, which was agreeing at all. The tag is a real step and it will help fans. It was just never going to help a brand work out who to pay.

Knowing a track is AI tells you what it is. It does not tell you who stands behind it. Until something does, the value AI music generates will keep flowing to everyone except the people whose work made it possible.

The rules for who gets paid still hold. What is missing, again, is the infrastructure to apply them.

The chain of ownership has to start somewhere. That's what we're building.