We tested an AI music detector
Before building this site we tested the best open AI-music detector we could find on the audio that is actually available for most artists: 30-second streaming previews. It did not work well enough to publish.
What we tested
SONICS (SpecTTTra), an open research model released under the MIT licence and trained on tens of thousands of Suno and Udio songs. Its authors report strong accuracy on full-length songs from those generators. We ran the 5-second and 120-second versions over five 30-second Deezer previews per artist, on 28 October 2026.
Results
The table shows the 120-second model's average and highest "AI" probability across each artist's five previews. Anything above 0.5 would be read as AI.
| Artist | Status | Average | Highest |
|---|---|---|---|
| Sleep Token | human | 0.80 | 0.98 |
| Morgan Wallen | human | 0.54 | 0.97 |
| Ghost | human | 0.31 | 0.97 |
| Taylor Swift | human | 0.38 | 0.92 |
| Slipknot | human | 0.22 | 0.84 |
| Kendrick Lamar | human | 0.40 | 0.80 |
| Noah Kahan | human | 0.27 | 0.71 |
| Chappell Roan | human | 0.43 | 0.67 |
| Bad Omens | human | 0.37 | 0.60 |
| Radiohead | human | 0.17 | 0.37 |
| Metallica | human | 0.19 | 0.33 |
| Sabrina Carpenter | human | 0.19 | 0.29 |
| Zach Bryan | human | 0.09 | 0.18 |
| Tyler Childers | human | 0.07 | 0.11 |
| Enlly Blue | reported AI | 0.89 | 0.96 |
| Breaking Rust | reported AI | 0.10 | 0.10 |
About one in four previews by well-known human artists scored above 0.5. Sleep Token averaged 0.80 and Morgan Wallen 0.54, while Breaking Rust, a project widely reported as AI-generated, scored 0.10. Across all previews, the model ranked a randomly chosen AI preview above a randomly chosen human one only about two times in three. The 5-second versions produced fewer false alarms but also missed most AI tracks.
What that means for you
Treat any single "AI score" from a website with care, especially one built on a short clip. Detectors are trained on particular generators and particular audio quality, and modern pop production can look synthetic to them. The catalog evidence our checker uses, release bursts, missing credits and no history, is less dramatic but much harder to fake at scale, and you can verify every part of it yourself.
Other ways to check, compared
| Tool | What it looks at | Good for | Watch out for |
|---|---|---|---|
| Is This Band AI? (this site) | The artist's public record: release dates, credits, label, MusicBrainz, Wikidata | Any artist on a streaming service, in seconds, with every piece of evidence shown | Reads the artist, not one recording; a careful AI project can look ordinary |
| Audio detectors (upload or link a track) | The sound itself | A full-length file from a generator the detector was trained on | Short clips and modern pop production cause false alarms, as in our test above |
| Deezer's AI tag | Deezer's own detector, run on uploads to Deezer | Albums distributed to Deezer, shown in its app | Only visible inside Deezer, and only where it has been applied |
| Artist disclosures and credits | What the artist or distributor says | The most reliable answer when it exists | Most AI projects say nothing |
| Community lists | Fans and researchers flagging acts | Spotting new suspects quickly | Mixed standards of evidence; check the reasoning |
Full-length audio and detectors run by the platforms themselves, like Deezer's, do better. If you need a verdict on one specific track and you have the full file, a platform-side tag or a disclosure from the artist is worth more than any public tool.