Unreliable
"It sounds generic"
Plenty of human music is formulaic, and plenty of generated music is odd. Taste is not a detector.
Guide
Is this song AI generated?
How to check a track honestly: what you can listen for, what a detector actually measures, and why a confident yes-or-no answer is usually the least trustworthy thing you will read about a recording.
Check a file first
The free AI Music Detector analyzes audio in your browser — nothing is uploaded, no account is needed — and shows every measurement behind its estimate.
Method
How to check a song, step by step
- 01
Start with provenance, not the audio
Who released it, when, on which channels, and is there anything else in that catalog? A release history, credits, live footage or session material settles more cases than any acoustic test.
- 02
Run the file through a detector
Use at least thirty seconds of the highest-quality copy you have. Screen recordings and low-bitrate re-encodes strip out much of the evidence before analysis even starts.
- 03
Read the individual signals, not just the number
Bandwidth ceiling, tonal uniformity across sections, dynamics, brightness movement and stereo behaviour each lean one way. Agreement between them matters far more than a single percentage.
- 04
Accept inconclusive as a real answer
Short clips, heavy limiting, lo-fi recordings and AI-assisted production all sit in the middle. That is information, not a failure of the tool.
Signals
What people think gives it away, and what actually does
Most of the popular tells are unreliable. The useful ones are technical rather than musical.
Unreliable
"The lyrics are strange"
Lyrics are written by people constantly, badly and brilliantly. They say nothing about how the audio was produced.
Sometimes useful
Bandwidth and tonal sameness
An abrupt high-frequency ceiling and near-identical tonal balance across distant sections are measurable, and lean synthetic — but heavy processing produces both too.
Sometimes useful
Static dynamics and stereo image
Very little movement in level, brightness or stereo width across a whole track is another weak signal, easiest to trust when several others agree.
Sourced
What detection can and cannot prove
Published research is unusually consistent on this point: detecting generated music is easy in the lab and hard in the wild.
Work on music deepfake detection reports near-perfect scores against the generators a model was trained on, and a sharp drop against generators it has never heard. The same papers show that ordinary transformations — MP3 re-encoding, pitch shifting, time stretching, resampling — reduce measured performance, because they remove or move the very artefacts a detector relies on.
That is why no detector output, including this one, is evidence of authorship. A measurement can say a recording's acoustic profile resembles the profiles typical of synthesis pipelines. It cannot say who pressed which keys, whether a human performance sits underneath, or which service was used. Naming a specific generator would require a trained and validated attribution model, which this tool is not.
It also matters that authorship and copyright are decided by law and evidence, not by a spectrum plot. The U.S. Copyright Office's guidance on AI-assisted works turns on human contribution, not on how a file measures — so a detector result is at best a reason to ask a question, never an answer to one.
Primary sources
- Detecting music deepfakes is easy but actually hard (Afchar et al., 2024)
Reports strong in-domain accuracy alongside degradation on unseen generators and on manipulated or re-encoded audio.
- SingFake: singing voice deepfake detection (ICASSP 2024)
Benchmarks show detection performance falling on unseen singers, languages and codecs.
- ASVspoof — spoofing and deepfake speech detection challenges
The long-running community benchmark documenting how generalization to unseen attacks remains the central open problem.
- U.S. Copyright Office — Copyright and Artificial Intelligence
Official guidance on registering works involving AI, where the deciding factor is human authorship rather than any acoustic test.
This page describes an acoustic analysis tool and cites public research. It is not legal advice and it is not a certification of authorship.
Limits
Why nobody can give you a certain answer
Every detector is measuring a production pipeline, not a person. A human performance that passed through a limiter, a low-pass filter and a lossy codec can measure like generated audio, and a carefully produced generated track can measure like conventional music. Those two failure modes are not edge cases; they are the normal state of a music catalog in 2026.
The hardest and most common case is hybrid work: human writing and performance with AI mastering, stem separation, voice processing or generated one-shots somewhere in the chain. A binary human-or-AI question has no correct answer for that material, which is why a good tool reports the evidence rather than picking a side.
So treat a detector as one input. If the stakes are real — a claim, a takedown, a moderation or hiring decision — you need provenance, project files, session material or an admission, not a probability.
FAQ
Frequently asked questions
Train your sound. Generate your song.
If you would rather build with AI than argue about it: prepare your own music, train a Sound that stays private to your account, and generate from your own material.
Related