AI Music Is Over Half of New Uploads. What That Does to Your Royalties
Machine-made tracks now make up more than half of the new music arriving at a streaming service every day. Whether that costs you anything depends less on the machines than on the payment model underneath your royalty statement: on a shared pool you are sharing with all of it, and on a per fan model you are not.
The number
On 21 July 2026 Deezer published the first figure anyone has put on this properly. It is receiving roughly 90,000 fully AI-generated tracks a day, which at June's peak was more than half of everything delivered to the platform. In 2025 alone it detected and tagged 13.4 million of them.
Deezer is the only major platform publishing this, because it is the only one that built the detection tool. Everyone else is receiving the same flood and not counting it out loud, which is worth remembering when you read that this is a Deezer problem.
The listening numbers are the interesting half. All that volume accounts for somewhere between 1% and 3% of actual streams on the platform. Almost nobody is choosing to play this stuff. Up to 85% of the streams that did land on fully AI-generated tracks in 2025 were fraudulent, against 8% fraud across the entire catalogue. So the tracks are not really being uploaded to be heard. They are being uploaded to be counted.
Why upload volume touches your money at all
Most streaming services pay pro rata. Subscriber money goes into one pot, the pot is divided by every qualifying stream on the service, and you are paid for your share of the total. Your fan's money does not reach you. Their listening is averaged into everyone else's.
That denominator is the whole exposure. Anything that adds streams without adding revenue makes each stream worth slightly less, and a track uploaded purely to be farmed adds streams by design. When the fraud is caught the streams come out of the pool and nothing reaches the farm. When it is not caught, that money was yours and someone else has it.
Nobody publishes how much slips through, which is the part that should bother you. Deezer's own answer is that on Deezer the effect is minimal, because detected AI tracks are pulled out of recommendations and playlists and fraudulent streams are stripped from the royalty pool before anyone is paid. I believe them. It is also an arms race, run by the platform, scored by the platform, on numbers only the platform can see. You are being asked to trust an enforcement process you have no way to audit, forever, at a volume that keeps climbing. The CISAC and PMP Strategy study Deezer cites puts nearly a quarter of creators' revenues at risk by 2028.
The same three events, two payment models
| What happens | On a shared pool | On a per fan model |
|---|---|---|
| 90,000 machine tracks arrive today | They join the denominator your streams are divided into | Nothing moves. Your fans' money is split by what your fans played |
| A farm streams AI tracks all night | Money leaves the pool unless the platform catches it | The farm has to buy subscriptions first, and a subscription costs more than an account can extract |
| A fan plays nothing but you all month | Their subscription is averaged across the whole service | Their subscription follows their listening |
That is the argument for attention-based payment, and this is the clearest illustration of it anyone has produced. RAAYDR splits each subscriber's own money across the artists that subscriber actually played. There is no shared pot, so there is no denominator for anyone to flood. A fan who gives you a fifth of their listening is worth 71p a month. Earning that from Spotify instead takes around 237 plays. A fan who plays nothing but you is worth £3.56, or around 1,187 plays. That is 55% of a £9.99 subscription after VAT, publishing royalties and card fees. Actual earnings depend on your share of each fan's listening.
It is not a fraud cure. A dishonest artist could still buy subscriptions and stream themselves, and we will police that like everyone else. What changes is the economics of trying: on a pool, a fake stream is close to free and pays out of your money. Here, it costs a subscription and pays out of the fraudster's own.
The other half of this, which is your masters
There is a second question underneath the upload flood, and it is the one artists actually ask me about. If the models are getting better, what were they trained on, and was any of it yours.
The precedent worth knowing is SoundCloud. In early 2025 the platform quietly updated its terms with a clause saying uploaded content might be used to "inform, train, develop or serve as input to" artificial intelligence, spotted by Ed Newton-Rex and reported by TechCrunch. After the backlash they changed it again, to an opt-in requirement for generative models that copy a voice or a likeness. Nothing about your upload changed in either direction. The paragraph governing it changed twice in a year.
So the thing to read is not the platform's blog post about respecting artists. It is the licence you grant when you upload, and how easily it can be amended.
Look at the scope of the licence. A streaming platform needs permission to stream your music to its subscribers. That is the deal. If the wording also covers using your recordings to develop products, ask what that covers, because model training sits comfortably inside it.
Look for the words "improve our services". It is the most common home for a training permission that nobody wanted to write plainly.
Check whether the promise is in the contract or on the poster. Plenty of platforms have a nice tick on an artists page saying your music will never train AI, while the terms grant a machine learning licence with a carve out. Both can be technically true at once. Only one of them is enforceable.
Check who can change it, and whether you get told. A promise you can be opted into by a version bump is not a promise.
Where we stand on it
RAAYDR has four written pledges on the artists page. Your music stays yours. It never trains AI, and we never license it to anyone who would. Uploading is free and always will be. The artist money moves monthly and we do not absorb what goes unplayed.
Our terms page currently says full terms of service arrive before Day One memberships open, which is true and also means the pledges are not yet load-bearing in a contract. When those terms land, all four have to be in them, in language no looser than what you just read. That is the test to hold us to, and it is a fair one to hold any platform to.
The listener side of this is simpler than the industry makes it sound. When Deezer ran a blind test through Ipsos with 9,000 people across eight countries, 97% could not tell an AI track from a human one. Four out of five wanted them labelled anyway. People are not asking to be able to hear the difference. They are asking to know, which is a different thing, and it is roughly what artists are asking for too.
FAQ
- Is AI music actually reducing what I earn per stream?
- On a pro rata platform the mechanism is real: more streams divided into the same revenue means each stream is worth less, and undetected fraud takes money out of the pool. The size of the effect is not published by anyone, including the platforms that would know. Deezer says its own detection and demonetisation keeps the impact minimal there.
- Which platforms are using uploads to train AI?
- Terms vary and they change. SoundCloud added a training clause in 2025 and narrowed it after artist pressure. Bandcamp and Tidal both state they will not train on artist uploads without permission. Many services simply say nothing about it, and silence is not a promise. Read the licence you grant at upload rather than the marketing page.
- Does RAAYDR train AI on music uploaded to it?
- No, and we will not license it to anyone who would. It is one of four pledges on the artists page and it has to appear in our terms before memberships open.
- Why does attention-based payment make fraud harder?
- There is no pool to draw from. Money is split per subscriber, across the artists that subscriber played, so inflating plays only ever moves money inside subscriptions you paid for yourself. Nothing on RAAYDR is playable without a subscription, which is what makes a fake account cost more than it can earn.
- How many AI tracks are being uploaded per day?
- Deezer reported around 90,000 fully AI-generated tracks a day in June 2026, more than half of all new uploads on peak days, with 13.4 million detected across 2025.
Figures on AI upload volume, streaming share and fraud are Deezer's own published data as of July 2026 and describe Deezer's platform. RAAYDR is pre-launch, so figures for RAAYDR are modelled from published rates rather than observed.
