How to Screen a Music Catalog for AI Tracks

Why scale changes the problem

A single-track check has a natural safety valve: you look at the score, you listen again, and you form a judgement with all the context in your head. That valve disappears the moment you are working through a spreadsheet. At a hundred rows you stop listening. At a thousand you stop reading the confidence column. The failure mode of catalogue screening is not that the detector is wrong — it is that a range with a stated uncertainty gets flattened into a yes or a no by whoever exports the CSV.

The second thing scale does is amplify base rates. If your true incidence of generated material is two percent and your screening step has even a modest false-positive rate, most of the rows you flag will be human tracks. That is not a defect in the tool, it is arithmetic, and it is the single most important number to keep in front of you. A screen that flags fifty tracks out of a thousand has not found fifty AI tracks. It has found fifty tracks worth a human minute each.

So the goal of a batch pass is deliberately modest. You are building a queue, ordered by how much a human should look at each item, and you are recording enough context that the ordering can be defended. You are not producing verdicts. Every design decision below follows from that framing.

Preparing the files so the results mean something

Most bad catalogue results trace back to inconsistent inputs rather than the analysis itself. Before you queue anything, get the material into a state where rows are comparable to each other.

• Use the highest-quality copy you have. Screening the 128 kbps preview when the master is on the same drive throws away the high-frequency evidence and widens every range unnecessarily.
• Keep the format consistent within a batch. Mixing WAV masters with heavily compressed MP3s in one pass means the confidence column is partly reporting your file management.
• Drop clips under roughly thirty seconds into a separate batch. Short excerpts produce genuinely wider ranges, and mixing them in makes the whole table look noisier than it is.
• Exclude the obvious non-music: spoken intros, stems, sound effects beds, silence-padded deliverables. They will produce scores, and those scores mean very little.
• Preserve the original filenames or IDs. The row is only useful if you can trace it back to a catalogue entry without guessing.
• Note upfront which tracks you already know the provenance of. Ten known-human and a couple of known-generated tracks in the same batch give you a free sanity check on the whole run.

Running the pass

Our batch tool processes locally in your browser, one file at a time. That design has consequences worth understanding rather than working around. Nothing leaves your machine, which is what makes it usable for unreleased catalogues and for material you are contractually not allowed to upload anywhere. In exchange, throughput is bounded by your own CPU and by how much audio the browser will hold, so files are decoded, analysed and released sequentially rather than in parallel.

Practically: expect a few seconds per track, keep the tab in the foreground since browsers throttle background work, and split very large catalogues into runs of a hundred or so. Between runs, export the CSV. There is no server-side record, so a closed tab is a lost run, and that is the deliberate trade for privacy.

Each row you get back carries a probability range rather than a point estimate, a confidence level, an audio-quality note and the analysed duration. Read those four together. A midpoint of sixty-two with low confidence on a 128 kbps two-minute file is a much weaker statement than the same midpoint with high confidence on a lossless master, and any triage that treats them identically is discarding the most useful part of the output.

A triage order that survives contact with a real catalogue

Sort by something more useful than the score. What you want at the top of the queue is not the highest number but the highest product of signal strength and consequence. In practice that means working through four tiers.

• Tier 1 — high midpoint, high confidence, good source quality. These deserve a human listen first, because the measurement is as strong as this method gets and the ordering cost is lowest.
• Tier 2 — anything with commercial exposure regardless of score: sync placements, playlist pitches, releases with disclosure obligations. A middling score on a track about to be licensed matters more than a high score on an unreleased demo.
• Tier 3 — the wide, low-confidence middle. Do not fight these one by one. Re-run them from better source files if you have them, and only escalate if the range narrows.
• Tier 4 — low midpoint with high confidence. Leave them. Chasing these is where screening programmes burn their credibility and their hours.

Where the false positives cluster

Batch work makes patterns visible that single checks hide. After enough runs the same categories of human music keep surfacing near the top, and knowing them in advance saves an enormous amount of wasted listening.

Programmed electronic music is the big one. Quantised drums, cloned loops, sustained synth pads and heavily limited masters produce exactly the low-variance, high-repetition profile that generative output also produces — because both are machine-steady by construction. A techno catalogue will read higher on average than a folk catalogue, and that difference is about production method, not origin.

Beyond that: library and production music written to brief, which is intentionally generic and loop-based; bedroom recordings mastered with aggressive presets that flatten dynamics; anything reconstructed through a stem separator, which leaves smeared transients that look like synthesis artefacts; and re-encoded uploads whose band edge has been shaved, which reduces the evidence available and pushes the range outward rather than down.

The mirror-image error matters too. Generated material that has been through a human pass — re-sung vocals, live overdubs, real-room reverb, a mastering engineer — often reads lower than fully synthetic output, because the measurable behaviours the detector relies on have been partly overwritten. A clean low score is not a certificate. On a catalogue where the incentive to hide provenance is high, that asymmetry is the thing to state out loud in your report.

What to do with a flagged row

The batch pass ends at the flag. Everything defensible happens afterwards, and it is a short, repeatable sequence.

• Re-run the single track from the best available master, on the main detector, so you get the full explanation of which behaviours moved the estimate.
• Check the file's own metadata. Encoder strings, software fields and comments occasionally name a generator outright; empty tags prove nothing either way.
• Look at a spectrogram for repeated blocks, edit points and flattened vibrato — findings about construction that a number cannot express.
• Run the stability check on two or three excerpts. A midpoint that swings twenty points between sections tells you the track is heterogeneous, which is itself informative.
• Ask. For anything with money or a contract attached, a request for session files, stems, project dates or a writer credit resolves more cases than any measurement, and it is the only step that produces evidence rather than inference.

Writing it down so it holds up later

The output of a screening programme is a document that someone else will read without you in the room. Write it as if that reader is sceptical and entitled to be. Record the date, the tool version, the source file quality and the analysed duration alongside the range and confidence — a bare number in a spreadsheet cell is worthless six months on, because nobody can reconstruct what it was measuring.

Use language that matches the evidence. 'Flagged for review: estimated 68–82% consistent with generated material, high confidence, from a lossless master' is a sentence you can stand behind. 'AI: yes' is not, and if it ever reaches an artist or a rights holder it will be the sentence you have to defend. Keep the human decision in its own column, separate from the measurement, so it is obvious which parts of the row are arithmetic and which are judgement.

Finally, keep the base-rate note at the top of the report. Stating plainly that a screen of this kind produces more false flags than true ones at low incidence is not a weakness in the work — it is what stops the work being misused, and it is the difference between a triage tool and an accusation engine.