AI Music Detector False Positives: How to Read a Score Before You Act on It
An AI music detector score is a statistical estimate, not a record of how a song was made. Before you block a release, reject a submission or hold royalties because of one, check a few things first: which file or link was actually analyzed, what the number means in that tool, and whether a second check agrees.
A case reported by Digital Music News this month shows why. The same 1980s recordings came back as "possible AI" from one link and as human from another.
What happened with the Twice a Man recordings?
Hamburg musician André Grafert, who records as Coloured Tears, released 23 covers of the Swedish post-punk band Twice a Man in June. He says he played, programmed and sang every track without generative AI. Ines Weidemann of International Artist Marketing (IAM), the PR firm working with him, ran all 23 through SubmitHub's detector, and every one was rated as potentially AI-generated. Other detectors she spot-checked mostly agreed.
She then tested Twice a Man's own recordings from the 1980s. Seven came back as "possible AI/hybrid," and two were matched to named AI platforms: "In My Mind" (1982) to "ElevenLabs Music" and "Across the Ocean" (1984) to "Treblo_v3."
SubmitHub pushed back in a follow-up DMN article on November 8. Those flags came from YouTube links. Checked by Spotify link, all seven originals came back human, including "In My Mind" at 91%. SubmitHub said its own test of the 1982 album scored all ten tracks between 85% and 95% human.
Why can the same song get different AI scores?
Detectors don't examine how a track was produced. They look for statistical patterns in the audio they receive, and research cited by DMN warns that they can lean on incidental features such as sample rate and bitrate. A YouTube stream and a Spotify stream of the same master are not the same file, so the input itself can change the result.
Results can also move without any change to the input. DMN reported that Let'sSubmit gave different verdicts on the same MP3 a week apart, moving one track from "Probably Human" at 61% to "Probably AI" at 65%. Genre may matter too: the detectors flagged machine drums, sequencers and processed vocals, the core sound of 1980s darkwave and post-punk.
What does a percentage like "93% human" actually mean?
According to SubmitHub, "93% human" does not mean 7% of the song was machine-made. It means the model is highly confident the track is human, because classifiers almost never return 100% even for recordings known to be human.
The trouble starts when a confidence score is used as a pass-or-fail line. SubmitHub's results page says tracks scored at 85% or above as likely AI are not accepted. DMN's point is that most people reading "82% AI/Hybrid" will take it to mean the song was probably made with AI.
Accuracy claims need the same care. SubmitHub cited a UMG-commissioned test that put its detector at 99.3% accuracy; the artist's campaign says that figure rests on confidential benchmarks. A peer-reviewed study in TISMIR found 4.7% false alarms on human music for a different commercial detector. Accuracy and false-positive rate are different metrics, and a high first number does not tell you the second.
What should you check before acting on a detector result?
The table below is our own checklist for anyone who reviews uploads, runs submissions or handles distribution.
Check | Why it matters | What the Twice a Man case shows |
|---|---|---|
Which source was analyzed? | Streams and files differ in encoding | YouTube link flagged, Spotify link scored human |
What does the score mean in this tool? | Confidence is not a share of AI content | "93% human" is confidence, per SubmitHub |
Is there a hard cutoff? | A cutoff turns an estimate into a verdict | 85% likely AI and above not accepted |
Does a second check agree? | Tools disagree, and can disagree with themselves | Same MP3 changed verdict a week later |
Can the creator respond? | Project files and stems can show the process | IAM says false flags can mean blocked releases and withheld royalties |
In our view, the safest place for a detector score is at the start of a review: it decides which tracks a person listens to again. It should not be the last step that decides what happens to a release.
Where does MusicScanner fit?
MusicScanner, MuseBlossom's AI music detection service, is in open beta. Its result is an AI analysis index meant to flag tracks for a closer listen, not proof of how a song was made. When we cite our own figure, we give it with its conditions: 201 of 207 tracks correct (97.1%) in an internal blind test. That figure does not show how a track outside the test set will score, and it should be read with the same care as any other detector's number.
What this case doesn't tell us
This is one artist's catalog, tested by the artist's own PR firm, so it does not measure how often any detector is wrong. SubmitHub disputes the description of false positives as widespread, and IAM says it knows of no study that has examined niche genres like darkwave. The link-dependence explanation also comes from SubmitHub's own retest.
FAQ
Can an AI music detector prove a song was made with AI?
No. Detectors estimate probabilities from audio patterns. They don't see the production process, so a flag is a reason to look closer, not proof.
Why would a YouTube link score differently from a Spotify link?
The audio each service streams is encoded differently, and detectors may react to features like bitrate and sample rate. If you can, test the original master file.
What should an artist keep in case of a false flag?
Project files, stems and recording notes. They show how the track was made when you appeal a result.