Every AI music detector does the same simple thing: it listens to a track and returns a score. The score is the easy part. What a distributor does with it, the thresholds, the review step, the record you keep, is where a catalogue stays clean or quietly fills with tracks that should never have shipped.
This is the policy layer that sits on top of any detector. A detector (a tool that scores audio for signs of AI generation) flags; a person decides. Here is how to write that down before you need it.
A score is a flag, not a confession
A detector returns a likelihood, not proof, and it can misfire. InterSpace Daily has covered detectors flagging a human composer’s album and genuine 1980s recordings: heavily processed or vintage audio can read as synthetic. Treat a high score as a reason to review, never as an automatic rejection that touches an artist’s release or royalties.
Set two thresholds, not one
One number is too blunt. Use two:
- A review line. Below it, the track passes without a human look.
- An action line. Above it, the track cannot move to any outcome that affects money or availability until a person signs off.
Between the two lines, a reviewer listens. Tie both lines to the score itself, not only the verdict word, because a 0.55 and a 0.97 can carry the same label and mean very different things. Our explainer on how AI music detection works and how to read a verdict covers that gap.
Decide what each band triggers
Map scores to actions in advance, so a reviewer is applying policy, not improvising:
- Pass and deliver.
- Request an AI disclosure from the uploader.
- Hold for a second listen.
- Label the track as AI-assisted where a store requires it.
Store rules decide how strict you have to be. Beatport bans fully AI-generated music and labels AI-assisted tracks, and Tidal tags and demonetizes AI-generated music. Your policy has to satisfy the strictest DSP (digital service provider) you deliver to.
Keep a record you can defend
For every track you screen, store the score, the model version and the time of the scan alongside the release. When an artist appeals a hold, or a store asks why a track was labelled, you can show the basis instead of reconstructing it. This is ordinary QC (quality control) hygiene, and it is the difference between a defensible decision and a guess.
Tell the artist what you found
When a track is held or asked to carry an AI label, say so plainly, and say why. An uploader who is told that a track scored high for AI-generated audio, and asked to confirm how it was made, can answer, correct the record, or withdraw the release. Silence turns a routine QC step into a grievance, and a clear, documented message doubles as your evidence if the dispute escalates.
Keep the tool advisory
The most important rule is the simplest: the detector never gets the final say. Build the workflow so a score opens a question and a human closes it. SEAM, the AI music detector built by The InterSpace Group (which also publishes InterSpace Daily), is advisory by design and never rejects, holds or approves a track on its own; whichever tool you use, keep that line. A distributor that auto-rejects on a score alone will eventually reject the wrong track, and the artist will be right.
Write the policy down, give reviewers two clear lines and a record to lean on, and the detector becomes what it should be: a way to point attention, not a judge.
