Contents
Two AI-music detectors can read the same song and return opposite verdicts, and most of the time neither one is broken. They are simply not measuring the same thing. Each was trained on a different set of generators, keys on a different trace in the audio, and draws its own line between real and AI, so disagreement is the normal outcome rather than a malfunction, and the split itself is worth reading.
They are reading different signals
A music detector is a family of methods that share an interface. The deployed norm scores the whole mix for a generator’s spectral-peak fingerprint (Afchar, Meseguer-Brocal, Akesbi, Hennequin, 2025). A research system starts from a large self-supervised speech model and fine-tunes it with a graph back-end (Zang, Zhang, Heydari, Duan, ICASSP 2024). A music-aware model adds pitch and rhythm cues alongside lyrics (SingGraph, Chen, Wu, Jang, Lee, 2024). A high-resolution model keeps the frequencies above 8 kHz that most detectors discard (Chen et al., 2026). And a lyrics-transcript detector ignores the signal entirely and classifies the transcribed words (Frohmann, Epure, Meseguer-Brocal et al., 2025). These tools can hear the same track and genuinely notice different things, because each was built to read a different trace. How each family works is covered in how AI music detectors work.
They were trained on different generators
Even two detectors of the same kind can split, because a detector mostly knows the generators in its training set, and each generator’s fingerprint is its own. The gap shows the moment a song comes from a model a tool never saw: the same simple detector that scores 100% on the generator versions it trained on falls to 39.83% on an unseen version (Afchar et al., 2025), and the SingFake system rises from a 4.62% equal error rate, the point where false accepts and false rejects balance, on seen singers to 42.77% on an unseen language or musical style (Zang et al., 2024). So a track from a brand-new version of Suno or Udio can read as synthetic to a detector that happens to cover it and clean to one that does not, purely because of what each was fed.
One of them may be reading the backing, not the voice
Part of the disagreement is not about synthesis at all. A detector that scores the full mix can end up keying on the production rather than the performer. Desblancs, Meseguer-Brocal, Hennequin and Moussallam (2024) found that singer-identification models built on full mixtures encode the instrumental background more than the voice, and degrade sharply on cloned singers. When one tool is effectively reading the arrangement, the mastering, or the accompaniment while another reads the vocal, they can land on different labels for reasons that have little to do with whether the song was generated.
They set the line in different places
Every detector ends in one decision: is the score high enough to call “AI”? That threshold is a choice, and different vendors choose differently. Research reports the threshold-neutral figure, the equal error rate, but a deployed tool has to pick an actual operating point, and away from that point the two errors diverge. A tool tuned to catch as many fakes as possible flags more real songs; a tool tuned to avoid false accusations waves more fakes through. Two tools reading an identical underlying score can therefore print different verdicts simply because their lines sit in different places.
Commercial and research tools disagree most
You cannot assume a consumer tool behaves like a published benchmark number. Cros Vila, Sturm, Casini and Dalmazzo (2025) tested a deployed commercial AI-music detector and found it “easily fooled by simply resampling audio to 22.05 kHz,” a change no listener would notice. Our own work points the same way: in our internal testing, a processing cascade drove a deployed commercial AI-music detector to 0% AI on tracks it would otherwise flag; the method stays in-house, but the finding is ours to report. A benchmark equal error rate is a poor predictor of how a commercial product will score any given song, so a research detector and a consumer tool can split with ease, and a vendor’s headline accuracy is not a measured reliability figure.
How to use disagreement instead of being confused by it
Disagreement is information, not noise to be averaged away. When independent detectors that read different traces all agree, that convergence is worth something, because they are unlikely to share the same blind spot. When they split, the split is telling you the song sits near the edge of what these tools can resolve, and no single score should be treated as the answer, least of all on a partly-AI track that no detector reads whole. Preserve the original file, run more than one method, look for provenance, and weigh the source. A verdict that survives two methods reading different traces, a signal detector and a lyrics detector, for instance, is far stronger than a confident number from one. For how far any of these tools can be trusted across media, see are AI detectors accurate.
Sources
- Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025). A Fourier Explanation of AI-Music Artifacts. ISMIR 2025.
- 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.
- Chen, Hu, Huang, Wu, Lee, Jang (2026). Joint Fullband-Subband Modeling for High-Resolution SingFake Detection.
- Frohmann, Epure, Meseguer-Brocal, Schedl, Hennequin (2025). AI-Generated Song Detection via Lyrics Transcripts. ISMIR 2025.
- Desblancs, Meseguer-Brocal, Hennequin, Moussallam (2024). From Real to Cloned Singer Identification. ISMIR 2024.
- Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1).