Flagged as AI on a Track You Made Yourself

You made the record. A detector says a machine did. Depending on where that verdict landed, you may now be facing a rejected release, a demonetised catalogue, or an accusation in front of a competition panel.

This is not rare and it is not your fault. It is a documented, systematic property of how these systems currently work.

First, understand what probably happened

Detectors do not have a concept of "made by a person". They have a statistical model of what generated audio tends to look like. Certain legitimate production choices land squarely inside that model:

• Everything made in the box. No microphone, no room, no noise floor. That absence is one of the strongest signals detectors use.
• Heavy quantisation. Grid-locked drums and perfectly aligned parts read as machine timing.
• Lo-fi with looped samples. Repetition plus limited dynamic variation.
• Autotune or vocoded vocals. At least one detector's own documentation concedes its vocal metric fires on heavy autotune.
• Virtual instruments. Orchestral and ambient built from sample libraries has no acoustic origin to detect.
• Low-bitrate or re-encoded files. One published audit flipped a detector's verdict simply by resampling to 22.05kHz.

In one public thread, musicians testing commercial detectors reported analogue-era recordings, VST synth sounds and a rough strummed guitar take all being called machine-made, with two detectors disagreeing with each other on the same file.

What to gather, in order

Assemble this before you write to anyone. It is much more persuasive arriving all at once.

• The project file. Your DAW session with its full edit history and timestamps. This is the single strongest artefact you have, and almost nobody asks for it because almost nobody thinks of it.
• Stems and raw takes. Unmixed multitrack, especially anything with visible performance imperfection, count-ins, false starts, talkback.
• Timestamps. File creation dates, cloud backup history, version history from your collaboration tool.
• Session evidence. Studio bookings, invoices, session musician contracts, photos or video from the room.
• Correspondence. Messages with collaborators discussing the track while it was being made.
• A second opinion. Run the file through another detector. If the answers differ, say so explicitly and name both.

Then write, to the right desk

Keep it short and factual. State that the track was flagged, that you are the author, that you have the session evidence, and that you are asking for a documented review rather than sympathy. Attach the project file or link to it. Ask specifically: what threshold was applied, which detector produced the finding, and what the published false-positive rate is in that band.

That last question matters more than it looks. Most platforms will not be able to answer it, and their inability to answer it is itself useful to you.

What we are building for this

The gap here is not another detector. It is an evidence pack: a dated finding that scores vocal and backing separately, reports what it cannot determine, states the error rate in the band it landed in, and can be forwarded to a distributor as a document rather than a screenshot.

That is the part of the product we care most about, and it is why the verdict itself will always be free.