Contents
You often cannot tell by ear anymore, and a detector will hand you a probability rather than a verdict. The reliable way to check whether a song was made by Suno, Udio, or another generator is a short procedure: test the original file, read what the detector actually measured, look for provenance, and corroborate with more than one method. A high score is a reason to look closer, not a finding on its own.
What a detector is listening for
A music detector does not judge whether a song “sounds AI.” It measures a fingerprint the generator leaves in the audio. Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025) showed that the deconvolution stages inside music generators stamp systematic peaks at fixed frequencies, an artifact fixed by the model’s architecture rather than its training. That fingerprint is what makes detection possible, and, as the next steps show, it is also what makes it fragile; the full mechanism is in how AI music detectors work.
Test the original file, not a re-encode
The evidence is fragile because it sits at specific frequencies, and anything that moves frequency content erases it. Re-encoding a track to MP3, forwarding it through a messaging app, or capturing it from a video re-compresses the audio and strips much of what a detector reads. Afchar, Meseguer-Brocal and Hennequin (2024) measured a detector falling from about 99.8% on clean audio to 73.8% after MP3 at 64 kbps and 66.6% after a small pitch shift. Whenever you can, run the untouched file rather than a re-shared copy, and read a clean “not AI” result on a heavily compressed clip as carrying little information.
Read the score for what it is
A detector score says how strongly a track carries the traces the tool was trained on, not whether a machine made it. A high score means the audio resembles the synthetic examples in the tool’s training set; a low score means it does not. Neither is proof. The number is only as meaningful as the match between your song and the generator the detector learned, which is why the same tool can be confident and correct on one model and confident and wrong on the next.
Know where music detectors fall apart
The gap between benchmark and reality is the whole problem. A simple detector that scores 100% on the generator versions it trained on drops to 39.83% on an unseen version, worse than guessing (Afchar et al., 2025). The SingFake system rises from a 4.62% equal error rate, the point where false accepts and false rejects balance, on seen singers to 42.77% on an unseen language or musical style (Zang, Zhang, Heydari, Duan, ICASSP 2024). A commercial detector was fooled simply by resampling audio to 22.05 kHz, a change you cannot hear (Cros Vila, Sturm, Casini, Dalmazzo, 2025). So treat a confident verdict on an unfamiliar or processed track with suspicion.
Check for provenance and watermarks
Provenance beats detection whenever it exists, but it is narrow. 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 evidence a human made the song (SynthID check, what it can and can’t tell you). Most tracks in the wild carry no mark at all, and any provenance metadata can be stripped by a re-upload. When a credential or generator watermark is present, it is far stronger than a classifier score; when it is absent, it proves nothing.
Bring in a second, different method
The most useful cross-check reads something the signal detectors do not. Frohmann and colleagues (2025) transcribe a song’s lyrics with speech recognition and classify the words, and because lyrics survive the processing that destroys spectral artifacts, the method holds around 85 to 90% on an unseen generator where the artifact detector falls to 56.9%. It has its own weakness, it can be beaten by changing the transcribable words, but because it fails differently from a signal reader, agreement between the two is much stronger evidence than either alone.
What actually settles it
So how do you tell? You combine signals rather than trust one. Test the original file, run more than one detector and read agreement as confidence rather than proof, look for provenance, and weigh the source and context of the track. Remember the case no tool handles well: a partly-AI song, a real instrumental under a cloned vocal, sits in a genuine blind spot, because detectors can flag that synthesis is present but not whose voice it is. When the file is degraded, the generator is unfamiliar, and no provenance exists, the correct answer is “undetermined,” not a percentage dressed up as certainty. How far these tools can be trusted in the first place, and where they break, is the wider reliability question in are AI detectors accurate.
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.