Do AI music detectors work? They catch Suno and miss open models
I ran 29 songs through three AI music detectors you can download. The best caught every Suno song and no human ones, but passed 8 of 10 ACE-Step songs as human.
One of the three AI music detectors I tried caught all 9 Suno songs I gave it and flagged none of 10 human recordings, but called 8 of 10 songs from ACE-Step, a free open-source music model, human. The other two missed most of the Suno songs. I ran 29 songs in all: 9 Suno songs from 2026, 10 I made with ACE-Step 1.5, and 10 human recordings from before AI music existed.
That's a small test, so read the counts as what happened, not as accuracy rates. I didn't test the web checkers, which don't say what's inside them. What the counts do show is that a verdict cuts one way. “AI” from the best detector was real evidence, and “human” mostly meant the song hadn't come from a generator the detector had met.

The three detectors, and what each learned from
The web checkers don't publish their models. The ones that say what they detect list services like Suno, Udio, MiniMax and Mureka, and no open model. So I used three whose models you can download:
- SONICS is a neural network trained on about 49,000 Suno and Udio songs from 2024. It listens to two minutes of a song.
- Fakeprint is a small model built on a method from Deezer's research team. It ignores style and looks only for artifacts the generator leaves in the audio. Its makers say it covers Suno up to v5 and Udio 1.5.
- AST is a general audio classifier fine-tuned on 10,000 AI songs, about half from Udio and only a few hundred from Suno, plus 10,000 human tracks.
To make sure a miss wasn't a broken setup, I first ran eight songs from SONICS's own 2024 test set. Every detector got at least seven of the eight right. My test songs were different: Suno songs that people posted under Creative Commons licenses between May and October, human tracks released under CC licenses between 2006 and 2015, and ten ACE-Step songs made with the same settings our free AI music maker uses.
The detectors only agree about the human songs
Each dot is one song, placed at its AI score
Fakeprint · looks for decoder artifacts
SONICS · trained on 2024 Suno and Udio
AST · trained mostly on Udio
All 29 scores
| Song | Fakeprint | SONICS | AST |
|---|---|---|---|
| To Give You The Day · Suno, 2026 | 100% | 6% | 0% |
| Moon Vending · Suno, 2026 | 100% | 6% | 0% |
| The Lily Pond · Suno, 2026 | 100% | 15% | 100% |
| Sip They Hate · Suno, 2026 | 100% | 98% | 99% |
| First Warm Day · Suno, 2026 | 100% | 21% | 99% |
| Million Shadows · Suno, 2026 | 100% | 23% | 0% |
| Nessuno · Suno, 2026 | 100% | 13% | 0% |
| Chhera · Suno, 2026 | 100% | 8% | 100% |
| It's TV6 Junior Time · Suno, 2026 | 100% | 8% | 3% |
| Synth-pop · ACE-Step 1.5 | 5% | 16% | 0% |
| Garage rock · ACE-Step 1.5 | 36% | 12% | 0% |
| Lo-fi beat · ACE-Step 1.5 | 91% | 30% | 0% |
| Boom bap · ACE-Step 1.5 | 32% | 27% | 0% |
| Folk ballad · ACE-Step 1.5 | 3% | 92% | 0% |
| Country · ACE-Step 1.5 | 8% | 55% | 0% |
| R&B slow jam · ACE-Step 1.5 | 35% | 98% | 0% |
| Melodic house · ACE-Step 1.5 | 90% | 31% | 0% |
| Lounge jazz · ACE-Step 1.5 | 1% | 96% | 0% |
| Phonk · ACE-Step 1.5 | 47% | 17% | 0% |
| Jonathan Coulton, Code Monkey (2006) · Human, pre-2016 | 0% | 7% | 0% |
| Kellee Maize, In Tune (2012) · Human, pre-2016 | 0% | 9% | 0% |
| Derek Clegg, Be Free (2011) · Human, pre-2016 | 0% | 4% | 0% |
| Pete Lund, Gypsy Cab (2014) · Human, pre-2016 | 0% | 4% | 0% |
| Tigerberry, Get Out (2015) · Human, pre-2016 | 0% | 10% | 0% |
| Maelstrom, I Dreamt For A Brighter Sky (2012) · Human, pre-2016 | 0% | 11% | 0% |
| ProleteR, April Showers (2012) · Human, pre-2016 | 0% | 6% | 0% |
| Boogie Belgique, Once Have I (2012) · Human, pre-2016 | 0% | 3% | 0% |
| ETHX, 8 Days (2009) · Human, pre-2016 | 0% | 5% | 7% |
| intouch, Heaven In Your Mouth (2014) · Human, pre-2016 | 0% | 5% | 0% |
None of the three flagged a human recording. That's the failure musicians worry about, and on these ten it didn't happen. It depends on the exact model, though: another SONICS variant scored Jonathan Coulton's 2006 “Code Monkey” as 96% AI.
On the AI songs they split three ways. Fakeprint caught every Suno song, all at 100%, and only 2 of the 10 ACE-Step songs. AST caught 4 of the 9 Suno songs and none of the ACE-Step ones. SONICS caught 1 Suno song and, oddly, 4 ACE-Step songs. Four of my ACE-Step songs got past all three.
Why the AI music detector that catches every Suno song misses ACE-Step
Most AI music models don't write sound directly. They build a small, compressed sketch of the song, and a decoder blows it back up into audio in several stages. Each stage leaves faint bumps at evenly spaced pitches, like the seams in a tiled floor, and they show up if you chart how much of each pitch a recording contains. The spacing depends on how the decoder is built, so each model family leaves its own pattern. Deezer's researchers worked this out in 2025, and it's what the fakeprint detector measures.
Faint bumps in the sound, by pitch, averaged over each group of songs
Pitch runs left to right. Tick marks are a fixed row: every 200 Hz for Suno, every 250 Hz for ACE-Step. Hover to line the panels up.
In my songs, Suno's bumps sit every 200 Hz. ACE-Step's land in different places, many of them 250 Hz apart. Human recordings have neither, because no decoder made them. Counting how strongly each song lines up with a 200 Hz row, all 9 Suno songs scored above all 10 human songs. With a 250 Hz row, all 10 ACE-Step songs did. I chose 250 after seeing where ACE-Step's peaks fell, so that shows the fingerprint exists, not that a detector built on it would work.
So the ACE-Step songs carry a fingerprint about as clear as Suno's, in a place fakeprint wasn't built to look. My guess is that a detector retrained on a few hundred ACE-Step songs would catch them easily. Someone has to do that, though, for every open model and every fine-tune that changes the decoder.
Synth-pop, made with ACE-Step 1.5
Detectors age along with the service they learned
SONICS shows the other way a detector fails. On its own 2024 test songs it got all eight right. On the 2026 Suno songs it flagged one in nine, probably because Suno has shipped several new models since 2024. The four ACE-Step songs it flagged were the four mellowest ones (folk ballad, country, slow jam, lounge jazz), which looks more like reacting to a sound than to anything about how the music was made.
Fakeprint didn't age: it flagged every 2026 Suno song at 100%, as it did the 2024 ones. A detector that learns a style goes stale with every model update. One that learns an artifact goes stale when the architecture changes, and whatever fakeprint reads in Suno's output hasn't changed since 2024. That's the first thing I'd want to know about any detector, open or commercial.
“To Give You The Day”, made with Suno in 2026 (CC0, from the Internet Archive, first 75 seconds)
Re-encoding hurts all of them. After a harsh 40 kbps MP3 re-encode, about the quality of a bad rip, fakeprint still caught 6 of the 9 Suno songs. SONICS and AST caught none.
Check the file before you run a detector
Six of the nine Suno files still had “made with suno” in their MP3 tags, along with a creation date and a song ID. The newest one also carried a C2PA credential, a signed provenance record, issued by Suno. The other three had no tags, probably because their uploaders re-exported them. None of the ACE-Step files said anything about where they came from.
So I'd go in this order: read the tags (exiftool or any tag editor shows them), run a detector, then weigh its answer by what it was trained on. An “AI” from an artifact detector was the strongest signal I got. A “human” on a song that could have come from an open model tells you very little.
That part won't fix itself. ACE-Step runs on a home GPU and anyone can modify it. A detector is a list of generators it has already met, and open models are the ones with no guarantee of being on that list.
More from the Blog
NVIDIA MotionBricks: The 15,000 fps Isn't the News
MotionBricks' 15,000 fps is mostly a counting choice. I ran it on a laptop. The news is fast, smooth and accurate at once, enough to replace game transitions.
ChatGPT Images 2.6 GIFs Act Like Sprite Sheets: They Keep the Look, Lose Count
ChatGPT Images 2.6 GIFs behave like one sprite sheet cut into frames, so they keep the look and lose count. In 18 juggler runs, none kept three balls.