About
Somebody had to count.
We are not here to argue about whether AI music should exist. We are here so that when it turns up undeclared, somebody can say so with a number attached.
Why this exists
In January 2025 one streaming service was receiving about 10,000 fully AI-generated tracks a day. By June 2026 it was near 90,000, over half of everything delivered to it. Nobody uploading those tracks is required to say what made them.
That creates a very ordinary problem. A playlist curator cannot tell. A competition judge cannot tell. A label A&R listening to two hundred submissions a week cannot tell. And an artist who spent four months and eleven hundred pounds in a room with a drummer is in the same royalty pool as a prompt.
So we are building the boring thing: a detector that gives a number, names the model, splits the vocal from the backing, and publishes its own error rates so the number can be argued with.
We are deliberately not an advocacy organisation. Plenty of good records use generative tools somewhere in the chain, and major labels have signed licensing deals with the generators. Our job is to establish what a recording is, not to tell you how to feel about it.
Deezer newsroom, 20 Apr 2026 and 21 Jul 2026What we will not do
Keep your audio.
Analysed in memory, discarded on completion. Not stored, not sold, not training data.
Hide the error rate.
Every threshold band published with its false-positive rate, including the ones that make us look worse.
Average away a weak spot.
Instrumental-only detection is the hardest class in the field for everyone. When we have a number for it, it goes next to the good ones, not in a footnote.
Put the truth behind a paywall.
The verdict is free forever. Paid plans buy detail and volume.
Claim a number we have not measured.
There is no accuracy figure anywhere on this site today, because there is not yet one worth printing.
How we intend to measure ourselves
The benchmark is fixed
Human recordings across many genres and generated output across every model we can obtain. Dated, frozen, and released to licensees so they can reproduce our figure instead of taking it on trust.
The index is versioned
Every finding carries the index version that produced it. A verdict from March can be re-litigated against the corpus that existed in March, which is the only way a finding survives a dispute.
Coverage is dated
A public, dated changelog of which generators reached the index and when, so the claim can be checked rather than asserted.
Where we are today
Pre-launch. Method published, benchmark in design, detector not shipped.
We are independent and not owned by a label, a distributor or a generator. If that changes we will say so on this page before you read it anywhere else.