Contents
Maybe, and the answer is that an AI-music detector is reliable only on the generators it was trained on. On a Suno or Udio version it has seen, it scores near-perfect; on a model it has not seen, or after a few seconds of ordinary processing, it collapses toward a coin toss. AI music is now a large share of what streaming services ingest, more than 10,000 tracks a day delivered to one platform alone (Frohmann, Epure, Meseguer-Brocal et al., 2025, citing Deezer), so the question is real, but a clean “not AI” result is not proof a song was made by a human.
What a music detector actually reads
A detector does not hear “AI-ness.” It reads a production fingerprint the generator leaves in the signal, as how AI music detectors work sets out. Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025) proved that the deconvolution stages common to music generators stamp systematic peaks at fixed frequencies, an artifact that depends on the model’s architecture rather than on its training data or weights. The tell is so specific that a roughly 10,000-parameter model reading those peaks matches a 1.6-million-parameter neural network, both clearing 99% accuracy on the open-source generators they were trained on. That is the reassuring number. What it hides is what happens off that training set.
Strong on known generators, near-chance on unseen ones
The same fingerprint that makes detection easy also makes it brittle, because each generator’s peaks are its own. On the versions it was trained on (Suno v3.5, Udio-130) Afchar and colleagues’ detector classifies AI music perfectly, but on Udio-32, a version it had not seen, it drops to 39.83%, worse than guessing. The pattern repeats in the singing-voice literature: the SingFake system reaches a 4.62% equal error rate, the point where false accepts and false rejects balance, on singers it was trained on, then rises to 13.62% on an unseen codec and 42.77% on an unseen language or musical style (Zang, Zhang, Heydari, Duan, ICASSP 2024). Even the winner of the first dedicated singing-deepfake challenge, which scored a 1.65% equal error rate on the contest set (Zhang, Zang, Shi et al., 2024), is a benchmark number, not a promise about the next model that ships.
A little processing flips the result
Because the evidence sits at fixed frequencies, anything that moves frequency content erases it. A detector that scores about 99.8% on clean audio falls to 73.8% after MP3 compression at 64 kbps, 58.4% after AAC, and 66.6% after a two-semitone pitch shift (Afchar, Meseguer-Brocal, Hennequin, 2024). Time-domain edits are the exception, surviving at 88.6% for time-stretching and 96.9% through added reverb, but most songs you encounter have already been re-encoded on the way to you. For the full picture of what compression does, see does MP3 compression defeat AI music detectors.
Commercial detectors, broken in testing
A vendor’s accuracy claim is not a measured reliability figure. 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 that does not alter how the track sounds. 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 behind that test stays in-house, but the finding is ours to report, and it matches the published result, a consumer detector is far easier to defeat than its marketing suggests.
The one robust exception: read the lyrics
Not every method reads the signal. Frohmann and colleagues (2025) transcribe a song with speech recognition and classify the resulting words, and because lyrics survive the processing that destroys spectral artifacts, the approach holds up where the artifact detector cracks. On an unseen generator their lyrics models stay around 85 to 90% recall while the artifact network falls to 56.9%; under a pitch shift the lyrics models hold near 89% while the artifact network drops to 59.0%. It has its own limit, it can be defeated by changing the transcribable words, but it fails in a different way than the signal readers, which is exactly why running both is stronger than trusting either.
What no tool can do
The case that defeats every commercial family is the partly-AI song, a real instrumental with a cloned vocal, or a human take over a generated backing. Artifact detectors can flag that a track contains synthesis but not whose voice it is; melody and fingerprint matchers identify the composition, not the singer; and singer-identification models degrade sharply on cloned voices, especially those that read the full mix and end up scoring the instrumental background rather than the voice (Desblancs, Meseguer-Brocal, Hennequin, Moussallam, ISMIR 2024). No deployed tool reliably answers “which part of this song is AI,” so a mixed track sits in a genuine blind spot.
So how should you read a result?
Treat a detector as a first pass, not a verdict. Test the original file rather than a re-shared or re-encoded copy, since each re-compression strips more of the evidence. Read a positive on a known generator as informative and a negative as weak, because absence of a fingerprint is not proof of a human. Check for provenance where it exists, but remember a watermark check only finds a watermark: a SynthID lookup detects Google’s own Lyria mark and returns “not detected” for Suno, Udio, and everything else, which is not the same as human-made (SynthID check, what it can and can’t tell you). Above all, corroborate, run more than one method, weigh the source, and read agreement between independent tools as confidence rather than proof. For the step-by-step version of this check, see how to tell if a song is 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.
- Zang, Zhang, Heydari, Duan (2024). SingFake: Singing Voice Deepfake Detection. ICASSP 2024.
- Zhang, Zang, Shi, Yamamoto, Toda, Duan (2024). SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge.
- Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1).
- 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.