Contents
A flag is not an explanation. It means the track carried a trace that resembled the generated songs a detector was trained on, scored past a threshold, and nothing more. It does not tell you which part of the audio moved the decision, and the music-detection research keeps landing on an uncomfortable point: the trace lives at fixed frequencies where compression cuts, and a full-mix detector can end up reading the production rather than the voice.
What does a “flag” actually mean?
A detector outputs a score for how closely a track matches the traces it learned to tie to generation, and a threshold turns that score into a label; how a detector arrives at that score is set out in how AI music detectors work. Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025) showed that the deconvolution stages common to music generators stamp systematic peaks at fixed frequencies, an artifact set by the model’s architecture rather than its training data. The tell is precise enough that a roughly 10,000-parameter model reading those peaks matches a 1.6-million-parameter network. But a flag built on that signal tells you only that the track looked synthetic to one model at one threshold, not that the model was right or what it keyed on.
Is the detector reading the voice, or the file around it?
Often the file around it. A codec reshapes audio in ways a narrow model can confuse with a generator’s fingerprint, and most songs have been through at least one codec before you hear them. Afchar, Meseguer-Brocal and Hennequin (2024) measured a detector that scores about 99.8 percent on clean audio falling to 73.8 percent under MP3 at 64 kbps and 58.4 percent under AAC, which means the same processing that degrades a real recording can also drag its score toward the fake side. A track that was only ever compressed, never generated, can drift toward a flag purely because the codec left traces the detector associates with its synthetic training set.
Could it be reading the backing, not the vocal?
Yes, and this is the music-specific trap. A detector that scores the full mix can key on the arrangement and mastering rather than the performer. Desblancs, Meseguer-Brocal, Hennequin and Moussallam (ISMIR 2024) found that singer-identification models built on full mixtures encode the instrumental background more than the voice, and degrade sharply on cloned singers. So a flag on a full-mix track may be reacting to the production, the accompaniment, or the codec, none of which tell you whether the vocal was cloned. The one case that defeats every deployed family is the partly-AI song, a human take over a generated backing or the reverse, where no tool reliably says which part is synthetic.
What would a trustworthy flag be reading?
A trace grounded in the music rather than the container. The strongest research systems start from a large self-supervised speech model and add a graph back-end, and their learned representations generalize across generators better than a raw spectral fingerprint (Zang, Zhang, Heydari, Duan, ICASSP 2024). Music-aware models go further: SingGraph adds a music-understanding model for pitch and rhythm alongside the speech model for lyrics, and Chen, Wu, Jang and Lee (2024) report relative equal-error-rate improvements of 13.2 percent on seen singers and 24.3 percent on unseen singers over a strong baseline. A different family drops the signal fingerprint entirely and reads the transcribed lyrics, holding around 85 to 90 percent on an unseen generator where a spectral detector collapses to 56.9 percent (Frohmann, Epure, Meseguer-Brocal, Schedl, Hennequin, 2025). A flag resting on musical structure or the words is worth more than one that could be reading the master, though even these fall on a generator they never saw.
Why do two music detectors flag different tracks?
Because they read different traces and trained on different generators. The gap shows the moment a song comes from an unfamiliar model: the same simple detector that scores 100 percent on the versions it trained on falls to 39.83 percent on an unseen version (Afchar et al., 2025), and the SingFake system rises from a 4.62 percent equal error rate on seen singers to 42.77 percent on an unseen language or musical style (Zang et al., 2024). And a deployed commercial detector was “easily fooled by simply resampling audio to 22.05 kHz,” a change no listener would notice (Cros Vila, Sturm, Casini, Dalmazzo, 2025). So one tool can flag a track another passes, purely because of what each was fed and where each draws its line. Why that happens in detail is covered in why do two music detectors disagree.
So how should you treat a flag?
As a first-pass signal, not a verdict. Ask what the detector could have keyed on besides a cloned vocal: heavy compression, a resample, a dense production it is reading instead of the voice. Test the original file rather than a re-encode, since each re-compression strips more of the evidence, a failure mode covered in does MP3 compression defeat AI music detectors. Run a second method, ideally one that does not read the spectral fingerprint, such as a lyrics-transcript detector, and read agreement between independent readers as confidence rather than proof. A flag says “this resembled my generated examples.” Turning that into “this song was made by a machine” is your job. Putting that judgment to work on one track is the practical question in is this song AI-generated.
Sources
- Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025). A Fourier Explanation of AI-Music Artifacts. ISMIR 2025.
- Afchar, Meseguer-Brocal, Hennequin (2024). Detecting Music Deepfakes Is Easy but Actually Hard.
- Desblancs, Meseguer-Brocal, Hennequin, Moussallam (2024). From Real to Cloned Singer Identification. ISMIR 2024.
- Zang, Zhang, Heydari, Duan (2024). SingFake: Singing Voice Deepfake Detection. ICASSP 2024.
- Chen, Wu, Jang, Lee (2024). Singing Voice Graph Modeling for SingFake Detection. Interspeech 2024.
- Frohmann, Epure, Meseguer-Brocal, Schedl, Hennequin (2025). AI-Generated Song Detection via Lyrics Transcripts. ISMIR 2025.
- Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1).