Detector reliability: our results and our limits

A detector that claims to be “100% reliable” is lying. We prefer to publish our results and our limits, so that everyone knows how to read a score.

The figure to remember

In November 2026, we measured the tool's complete verdict, as you see it (all readings combined), on 545 real tracks under free licences that it had never heard: jazz, classical, piano, orchestra, electronic. 28 were classed “Likely AI” or higher, i.e. 5.1% false alarms, including 8 “Very likely AI” (1.5%). In other words, about one real track in twenty is wrongly flagged. We are working to bring this figure down; meanwhile, an “AI” score is a clue to check, never proof.

Our tests

We tested the tool on human recordings (studio masters, demos, a cappella choir, live recording), on tracks generated by Suno, and on hybrid tracks made for the occasion.

October 2026 update. We tested the tool on 288 excerpts of real music: masters and demos, live bands, well-known artists, backing tracks, drum lessons and isolated instrument tracks. The old rule wrongly classed nearly a third of them as “AI” or “hybrid”, because of the isolated vocals, which sometimes reach 100% on a real voice. The isolated vocals are now shown for information but no longer used for the verdict, and there is no longer an “undetermined” zone. On these same excerpts: no false alarm on 70 complete mixed tracks, 1 out of 98 backing tracks, and 4% overall, mostly on single-instrument tracks (drums, brass). The trade-off we accept: a generated voice placed over real music is not flagged.

This is a starting sample. We keep expanding it and will publish updated results on this page.

Recent generators: our in-house model

Public detectors were trained on Suno and Udio. Faced with more recent generators, such as ACE-Step (free and open), they see almost nothing: our base detector caught only 2% of these tracks. Since October 2026, an extra reading, “Recent generators”, relies on a model we trained ourselves.

So that it does not mistake real music for AI, we trained it with about 1,750 human tracks: tracks released before generative AI arrived, under free licences that allow this use (jazz, piano, orchestra and string quartets, electronic), studio and concert recordings entrusted to us by musicians, and tracks uploaded by contributing artists. Then we tested it on tracks it had never heard:

We set this model to speak up only when it is very sure: better to let an AI track through than to wrongly accuse a musician.

What is reliable

• The instrumental only: this is the most discriminating reading (2 to 10% for human, up to 98% for AI).
• Suno and Udio tracks from the versions the models were trained on (up to 2024). Recent versions (Suno v6…) are still to be measured.
• Location in time: the timeline shows the passages where the clues concentrate. Since November 2026, a passage that strongly resembles AI (90% or more for about ten seconds) is flagged even when the rest of the track sounds human, without changing the verdict. On 545 real tracks never heard before, this note appears on 7 of them (1.3%): it is a passage to listen to, not proof.

What is less reliable

• The isolated vocals: this is the weakest clue. They are shown for information but not used for the verdict.
• Unknown generators: each new version can change the traces left in the sound. Our in-house model covers ACE-Step; we extend it as we go.
• Reworked tracks: slowing down, speeding up, re-mastering or recompressing a track can blur the detection.
• Retouched AI tracks: some tools offer to “clean up” a generated track (softened highs, EQ, saturation, reverb, pitch shift). In November 2026, we tested this on 60 ACE-Step tracks that our model actually knows: 52 caught as they were, 16 after a retouch (EQ, saturation, reverb), none after a half-tone pitch shift. The retouch also erases the C2PA signature: 0 out of 20 Eleven Music tracks caught after retouching. Today, a retouched AI track can therefore pass for human. We are training a new version to withstand these retouches and will publish its results here.
• Platform links (YouTube…): the sound is recompressed there, which slightly reduces accuracy.
• Unusual styles: very synthetic electronic music and robotic vocals (vocoder, extreme Auto-Tune) can raise the score without any AI.

Spoken voice (beta)

For the spoken voice, we compared eight public detectors on a test bench made of real voices (audiobooks and readings in French, English and Russian, studio recordings, then in October 2026 more than 900 excerpts of amateur LibriVox audiobooks and 370 speeches at the European Parliament, in five languages) and generated voices (ElevenLabs including the v3 and v4 models, Microsoft neural voices, the open-source Piper and Kokoro engines). Six of them did barely better than chance. We kept the best one and complemented it with a layer trained only on real voices and open-source synthetic voices (Piper, Kokoro): ElevenLabs and Microsoft voices are used for testing only. Each generator is tested without having been seen during training:

That is 6 out of 10 ElevenLabs voices caught, with 2% of real voices wrongly flagged, although no ElevenLabs voice was used for training. Weak point: Microsoft neural voices currently go unnoticed. This is a beta module, still in training: a “Likely human” verdict does not guarantee that a voice is real, and the most recent cloned voices can go unnoticed. When the file carries a provenance declaration (C2PA), like those downloaded from ElevenLabs, it is read and takes precedence over the analysis.

How to read a score

A high score is a clue to check, not proof. Before any important decision, cross-check it with other elements: careful listening, the artist's history, session files, a conversation with the author. A low score does not guarantee either that no AI was used.