What does an AI music-detector score mean?

An AI-music detector score is a number that says how strongly a track matches the generated examples a model was trained on, read at an operating point you usually cannot see. It is not the probability that the song is AI. Research reports the behavior as equal error rate, the threshold where false accepts and false rejects balance, a single figure for a whole benchmark rather than a verdict on your file, which is why the same score can mean very different things from one track to the next.

What is the detector actually outputting?

A single number for how closely your track resembles the synthetic examples the model learned from, mapped through a threshold into a label. It is not a measurement of reality. Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025) showed the thing most detectors key on is a production fingerprint, systematic spectral peaks that the deconvolution stages in music generators leave at fixed frequencies, so specific that a roughly 10,000-parameter model reading those peaks matches a 1.6-million-parameter network. The score is a reading of that fingerprint, not a musical judgment.

What is equal error rate, and what is an operating point?

Equal error rate is the threshold where false-accept and false-reject rates are equal, a way to compare detectors without picking a bias toward one kind of error. A deployed tool has to choose an actual operating point, and away from the equal-error one the two errors diverge: a threshold set to catch more fakes will also flag more real songs. The costs are asymmetric, a false “AI” label on an independent artist’s own recording harms a real person, while a missed fake pollutes a catalog. A score without its threshold, its false-positive rate, and its test condition is incomplete evidence, and the vendor rarely shows you which point they chose.

Does “94 percent AI” mean a 94 percent chance the song is fake?

No. The output is a similarity-to-training score passed through a function, not a calibrated probability. Reading “94” as “94 percent likely AI” assumes the tool is calibrated on audio like yours, and it usually is not. The gap shows the moment a track comes from a model the detector never saw: the same simple detector that scores 100% on the generator versions it trained on (Suno v3.5, Udio-130) drops to 39.83% on an unseen version, Udio-32, worse than guessing (Afchar et al., 2025). The same displayed confidence can sit on either side of that collapse.

Why does the same score mean different things on different tracks?

Because the error rate behind the score depends on the track’s origin and handling. The SingFake system rises from a 4.62% equal error rate on singers it trained on to 13.62% on an unseen codec and 42.77% on an unseen language or musical style (Zang, Zhang, Heydari, Duan, ICASSP 2024), and off-the-shelf speech anti-spoofing models collapse to roughly 50%, a coin flip, when pointed at singing at all. A “90 percent” on a clean file from a known generator and a “90 percent” on a compressed clip from an unfamiliar one are not the same claim.

What can make a high score misleading?

A high score can be reading something other than synthesis. A full-mix detector can key on the arrangement rather than the voice: Desblancs, Meseguer-Brocal, Hennequin and Moussallam (2024) found singer-identification models built on full mixtures encode the instrumental background more than the vocal, and degrade sharply on cloned singers. And a commercial tool’s confident number is not a benchmark number. Cros Vila, Sturm, Casini and Dalmazzo (2025) found a deployed detector “easily fooled by simply resampling audio to 22.05 kHz.” In our internal testing, a processing cascade drove a deployed commercial music detector to 0% AI on tracks it would otherwise flag; the method stays in-house, but the finding is ours to report.

What can make a low score misleading?

A low score can mean the track is real, or that the detector never learned this generator, or that compression stripped the fingerprint. Because the evidence sits at fixed frequencies, lossy codecs cut where it lives: a detector scoring about 99.8% on clean audio falls to 73.8% after MP3 at 64 kbps, 58.4% after AAC, and 66.6% after a two-semitone pitch shift (Afchar, Meseguer-Brocal, Hennequin, 2024); how far that goes is the subject of does MP3 compression defeat AI music detectors. A low score on a heavily compressed song is not a clean bill of health. The hardest case for a low score is the partly-AI track, a cloned vocal over a real backing, which no deployed tool reads whole.

How should you read a score, then?

As a similarity reading at an operating point you are usually not shown, and as a prompt to ask what file you tested, which generator might have made it, what threshold was applied, and what the false-positive rate is on audio like this. One family fails differently and is worth pairing: Frohmann and colleagues (2025) transcribe the lyrics and classify the words, holding around 85 to 90% where an artifact detector falls to 56.9% on an unseen generator. Preserve the original file, run more than one method and read agreement as confidence rather than proof, and look for provenance where it exists. Putting that reading 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.
• Zang, Zhang, Heydari, Duan (2024). SingFake: Singing Voice Deepfake Detection. ICASSP 2024.
• 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.