For labels, distributors and platforms
Establish origin before the delivery is accepted.
One request per track in front of your ingest queue. A 0 to 100 origin score, the generative model named and versioned, separate findings for vocal and accompaniment, and a declared-versus-detected flag your compliance team can act on.
The service is not live. This page sets out the intended contract so that procurement, legal and security can start reading it now rather than at the point of sale. Nothing here is a performance commitment yet.
Three places it fits in your chain
| Deployment | What happens | Where it sits |
|---|---|---|
| Ingest gate | Every delivery scored before acceptance. Undeclared synthetic tracks held rather than ingested, so they never enter the royalty pool and never need clawing back. | Pre-acceptance |
| Catalogue sweep | Backwards pass over what already shipped, producing a dated finding per ISRC for recovery claims and disclosure obligations. | Post-release |
| Review desk | Seats for A&R and rights staff to check individual submissions by hand, with exportable findings. | Ad hoc |
What your compliance team receives
Three fields that matter more than the score.
01 / Attribution
The model, by name and version
Different generators carry different licensing consequences, and two of the three major labels have signed deals with some of them. A binary AI or human flag cannot support either position in a dispute.
02 / Disclosure gap
Declared versus detected
The supplier's submitted metadata compared against what the audio shows. That gap is the enforceable part of your supplier agreement, and it is the field the whole industry's self-disclosure regime currently has no way to check.
03 / Provenance
Watermarks read separately
SynthID and C2PA reported apart from the acoustic score, so a signal and an inference are never conflated in the record.
The industry standardised on self-declared metadata through DDEX, and the largest platform to adopt it says plainly that "the absence of a credit doesn't mean AI wasn't used." Detection is the only independent check on a voluntary declaration.
Spotify newsroom, 25 Sep 2025Security, retention and procurement
Data handling · intended
| Audio retained | Never |
| Used for model training | Never |
| Findings retained | Contract term |
| In transit | TLS 1.3 |
| At rest | AES-256 |
Contracting · intended
| DPA | Standard, SCCs |
| Processing regions | EU / US |
| Index-version pinning | Available |
| Benchmark released | To licensees |
| Security questionnaire | On request |
Early access
Catalogue audits first, when we open.
Tell us your volume and where it hurts. Labels, distributors and platforms on the list get the first sweeps, before the consumer tool opens.
What procurement usually asks
Earlier is cheaper. A withheld delivery costs a webhook, a recovered royalty costs a legal letter. The three usual positions are an ingest gate before acceptance, a backwards sweep over what already shipped, and a review desk for staff checking submissions by hand.
The declared-versus-detected gap. Comparing what the supplier submitted in metadata against what the audio shows produces a difference that is contractual rather than aesthetic. That field is the enforceable one.
No. Audio is analysed and discarded. Findings are retained for the contract term because you need them for disputes, audio never is.
That is the intent. Every finding carries the index version that produced it, and pinning will be available so a catalogue is scored against one fixed corpus rather than a moving one.