AI opt-out custodianship · live demonstration

Let’s watermark a track.

Duona files and maintains the machine-readable “do not train” record for every master in your catalog. Search a title, an artist, or paste your own ISRC, and watch one resolve.

The state of play

Most catalogs have nothing on file. The training already happened.

<1%
of media carries provenance

Fewer than 1 percent of news images and video ship with provenance credentials, per the Reuters Institute figure. Audio trails even that.

Filed vs. not
publishing moved, music didn't

Machine-readable opt-outs are standard practice across European publishing, per WIPO. In recorded music they are still the exception.

Both of them
flagship AI music services, sued

The majors sued Suno and Udio over training on unlicensed recordings. The copying came first; the court dates came after.

An opt-out only covers the catalogs that filed one. Everything else reads as fair game to a crawler. Filing yours is the narrow, mechanical move this whole page demos.

From this demo to your catalog

Every master you own gets a do-not-train record, filed and kept current.

You send masters and a metadata export. The Duona Opt-Out Ledger Service fingerprints each recording, files the machine-readable record on the industry registry, and keeps it current every month. You get the ledger back. Getting a catalog on file becomes a signature and an invoice, not a project.

$$$
What it costs

Priced per master, not per project.

  • Solo artists: about $10 a month for a handful of tracks
  • Catalogs: $1.50 to $3 per master we file, $500 monthly minimum
  • Early pricing, locked in writing when you onboard
MP3
What survives

It still matches after the file changes.

  • Measured from the sound itself, not a tag anyone can delete
  • Matched through a 30 percent tempo change on our bench
  • Runs live above, on whichever record you pick
What it says

Your name and your terms, inside the sound.

  • Carries title, artist, who stamped it, and a short note
  • Rides inside the audio, so a stripped copy still says whose it is
  • Rewrite it any time: change the wording, stamp again, read it back
Write your own receipt
Why now

Since 2 August 2026 the European Commission can fine AI model providers that ignore machine-readable reservations.

Up to 3 percent of worldwide turnover or 15 million euros, whichever is higher. That fine lands on the model provider, not on you. Your side of it is narrow and mechanical: the reservation has to exist in a form a machine can read, on every recording you want covered. That is the record we file and maintain, so your catalog is already on file when enforcement starts.

Check the rules yourself: AI Act, Article 53 · the TDM opt-out, Directive 2019/790

What we do not claim

Filing a record does not decide a legal question, and no registry can prove what a model was trained on. A stamp is a label, not a lock: it survives ordinary delivery changes one at a time, and a one-key pitch shift breaks it. We hand you evidence and the people who run your takedowns decide what to do with it.

Vocabulary

Two pairs of words this industry mixes up.

Recognize it, or label it
Fingerprint · recognizes it

Nothing is added to the song. We measure the sound and get a code back. That code is how a registry knows which recording you mean.

Stamp · labels it

A name tag sewn inside the jacket. Invisible from outside. Open it and it says whose it is and what the rules are.

The recording, or the song
The recording

The master you own, the thing with an ISRC, and the thing a model actually trains on. This is where we file.

The song

The words and music underneath, carrying an ISWC, a separate right with its own registries. We do not file there.

Do both to one song, start to finish

Bring one ISRC or the whole catalog.

Scope and timeline are sized on the call, per catalog. Nothing here needs an account, and the demo above runs on the real index.

Stamp a song first
Real catalog intake

Start the conversation

Tell us who you are and roughly what you manage. A human replies.

Prefer email? [email protected]

[email protected]