How accurate are AI music detectors, really?
Every AI music detector on the market advertises accuracy above 98%. The peer-reviewed literature, including work published by the research teams behind these products, describes something considerably messier. This page collects what has actually been measured.
The one thing to take away. A published accuracy figure describes performance on a named list of generators, in clean conditions, on the date it was measured. Move any of those three variables and the number moves a lot.
Why headline accuracy numbers do not survive contact with real audio
Deezer's research team published a detector at ICASSP in January 2025 reporting 99.8% accuracy on its own test set, then measured what happened to it under ordinary audio processing:
| Condition | Detection of genuinely AI-generated tracks |
|---|---|
| Clean test set | 99.8% |
| Pitch shift of two semitones | 0.0% |
| Added white noise | 2.3% |
| Opus re-encode at 64 kbps | 14.6% |
| MP3 re-encode at 64 kbps | 19.6% |
| AAC re-encode at 64 kbps | 23.3% |
Source: Afchar, Meseguer-Brocal and Hennequin (Deezer Research), "AI-Generated Music Detection and its Challenges", IEEE ICASSP 2025.
Read the direction of that failure carefully, because it is almost always reported backwards. These are false negatives: AI tracks going undetected. The same paper reports that transfer to an unseen decoder family gave results of "almost always" 0%, and the authors conclude that "there will always be an unseen manipulation or generation method".
What these detectors are actually measuring
The same Deezer team published the physical mechanism in June 2025. Neural vocoders and codecs upsample audio by inserting zeros, which periodises the spectrum and replicates the energy peak at 0 Hz at regular intervals. The detector reads those periodic spectral peaks.
So a detector is not answering "did a human write this song". It is answering "did this audio pass through a neural vocoder". Those are different questions, and the gap between them is where every argument about AI detection actually lives.
The false positive problem
Research from KTH Royal Institute of Technology tested detectors against 30,000 tracks and found something uncomfortable: the models lean on encoding correlations rather than musical properties. Suno renders at 48 kHz and 192 kbps, Udio at 48 kHz and 320 kbps, while their human reference corpus was mostly 44.1 kHz at variable bitrates. The authors note these compression artefacts could create spurious correlations, which is a polite way of saying the detector may be identifying a file format.
They also found that high-pass filtering at 8 kHz or above caused one commercial detector to label every excerpt as AI, human tracks included. That test used very few samples, so it demonstrates that the failure mode exists rather than how often it occurs.
A 2026 preprint measured false positive rates on 94,654 human tracks across several open detectors and found a range from 0.1% to 14.7% depending on the model. The same work found those detectors caught only 44.8% to 74.2% of genuinely fully-AI tracks, while catching 92.5% to 100% of AI tracks that had been mixed or mastered by humans.
Sources: Cros Vila, Sturm, Casini and Dalmazzo (KTH), Transactions of the ISMIR, June 2025. Go and Kim, "HAIM", arXiv preprint, June 2026, not peer reviewed.
That last pair of numbers is the one nobody advertises. Missing between a quarter and a half of fully AI-generated tracks is the normal state of this technology, and it is the exact opposite of what the marketing on every detector's landing page implies.
What the platforms themselves say
Spotify stated publicly in September 2025 that detection systems are imperfect, produce a lot of false positives, and are "not helpful for policy enforcement", saying they prefer a declarative approach through metadata. Deezer, which built its whole labelling programme on detection, publishes research from its own team describing the field as a cat-and-mouse game.
Those two positions are not contradictory. Detection is useful as a signal at scale. It is weak as a verdict about any individual track, which is exactly why the industry is moving toward artists declaring AI use rather than machines inferring it.
How to read any accuracy claim, including ours
Three questions dismantle almost every number you will see advertised:
- Measured against which generators? A detector trained on Suno and tested on Udio drops sharply in the published literature. On a generator it has never seen, detection rates in the 3% to 24% range have been measured.
- Measured on what audio? Clean studio exports or the compressed, re-encoded, streamed audio you actually get in the wild. Five detectors that all scored above 0.98 F1 in clean conditions dropped to 0.19 on real broadcast audio in one 2026 preprint.
- Measured when? Every new generator release invalidates the previous measurement. A figure without a date and an engine version is decoration.
We hold ourselves to the same standard, which is why you will not find a single headline accuracy percentage on this site. What a score from our engine means is: the audio was compared against a set of forensic signals associated with known generation chains, and this is how strongly they were present. A low score means no generation signature was found. It is not proof that a human wrote the song, and no detector on the market can give you that proof today.
If you want the practical consequences of all this, we cover what to do when a track is wrongly flagged and what Deezer's labelling system actually does.