Screening new uploads for AI is the easy half: every track already passes through a gate, so you add a scan to the gate. The back catalogue is the hard half. You may have tens of thousands of tracks already live, already earning, and under contract, and you cannot pause delivery to check them. Here is a workflow that runs quietly in the background.
Screen beside delivery, not inside it
Run the catalogue through a detection API (application programming interface) as a separate batch job, not through your live delivery path. Nothing holds, nothing is pulled automatically, and no earning track stops while you work. You are building a risk list, not a blockade.
Sort by score and start at the top
A detector returns a likelihood for each track. Sort the whole catalogue by that number, highest first, and begin listening at the top. The point of a per-track score is that a handful of AI tracks cannot hide inside a 50,000-track catalogue; each one carries its own value, so the worst offenders surface on their own. Our explainer on how AI music detection works and how to read a verdict covers what that score is, and is not, telling you.
Triage in bands, never mass-delete
- Top band: listen, check the metadata and any AI disclosure, and decide case by case.
- Middle band: spot-check a sample before you trust the batch.
- Low band: leave it alone.
Never bulk-remove on a score alone. Detectors misfire on heavily processed or older recordings, as we saw when detectors flagged a human composer’s album and 1980s recordings. A mass delete turns one false positive into a pulled release and a furious artist.
Record the scan with the release
Store the score, the model version and the scan time against each track as you go. When a store later queries a catalogue for AI, you already hold the basis for every call you made. That record is also the backbone of the policy playbook every distributor should write down before the questions start arriving.
Start where the risk is highest
You do not have to scan the whole catalogue at once, and you should not treat every track as equally likely to be synthetic. Begin with the releases that carry the most risk: recent intake, high-volume uploader accounts, and any catalogue you acquired in bulk rather than signed track by track. Those are where undisclosed AI tends to concentrate. A detector that scores new tracks without a reference copy lets you work through them in priority order, a batch at a time, instead of freezing the whole library behind one enormous job. As detection models improve, a cheap re-scan of the top risk bands catches what an older model missed, without touching the tracks that already passed clean.
Pick a tool built for volume
Catalogue screening needs a specific shape of detector: one that scores brand-new tracks without needing a reference copy on file, accepts the formats your masters are stored in, and runs over an API you can point at thousands of files at once. SEAM, built by The InterSpace Group (which also publishes InterSpace Daily), needs no copy of the song on file and scans most tracks in a few seconds, and its first 1,000 scans are free, which is enough to pilot the workflow on your highest-risk releases before committing. Whichever tool you choose, the test is the same: can it run beside delivery, at catalogue scale, without stopping a single earning track?
