How AI Music Detection Works

What AI music is, and how it is generated

An AI generated song is produced by a model trained on large quantities of recorded music. You supply conditioning information — a prompt describing genre and mood, often lyrics, sometimes a reference clip — and the model samples a new piece of audio from the distribution it learned.

Practically every current system works in a compressed token space rather than on raw waveforms. Audio is encoded into discrete tokens by a neural codec, a transformer or diffusion model generates a token sequence under the conditioning, and a decoder turns those tokens back into sound. Most consumer services then apply loudness normalisation and limiting, so what you download already sounds mastered.

Each of those stages leaves a trace in the finished file. Detection is the business of looking for those traces — and the business of admitting when they are not there.

What a detection model is looking for

We do not train or own a detection model, so we cannot claim to know the exact features the service weighs. What the published literature on synthetic-audio detection consistently points at, and what the verdicts on this site behave consistently with, is the following family of evidence.

• Neural codec reconstruction traces — the characteristic way a decoder rebuilds high-frequency detail it never stored, which differs from how a microphone and a lossy encoder handle the same band.
• Cross-section consistency — how similar the tonal balance of one part of a track is to another. Recorded performances drift; whole-track generation often holds unusually still.
• Transient and dynamic behaviour — how far peaks sit above the average level and how percussive attacks are shaped. Heavily weakened by modern mastering on both sides.
• Stereo construction — whether the image looks captured by microphones in a room or assembled from separately placed elements.
• Vocal artefacts — formant behaviour, breath, sibilance and pitch micro-variation, which is why the service reports vocal and instrumental likelihoods separately.
• Temporal grid regularity — how exactly events line up to a metronomic grid across a whole arrangement.

What this application actually does

Everything above describes the field. This section describes the product, and the two should not be confused: an industry technique is only implemented here if it appears in this list.

We do not train, own or run a detection model. When you press analyse, the complete file is uploaded over HTTPS to our server, which forwards it to one specialist third-party AI music detection service and returns what that service reports. There is no in-house algorithm, no browser-side analysis, no stem separation, no watermark reader, no metadata inspection and no lookup against a database of known AI releases.

The audio tools elsewhere on this site do run entirely in your browser. The detector does not, and cannot: classification needs the full file at a dedicated remote service.

• Implemented: upload, one third-party classification call, and presentation of that service's verdict, vocal and instrumental probabilities, confidence and per-window values.
• Not implemented: generator attribution, stem separation, watermark detection, metadata or provenance checking, database matching, or any second opinion from a model of our own.
• Result caching: the returned numbers are kept for up to 14 days keyed by a SHA-256 hash of the file, so an identical repeat upload reuses them.
• Fair use: a salted one-way hash of your network address limits use to 12 analyses per 24 hours. The address itself is not stored.
• No accuracy figure is published, because no controlled evaluation of the current system has been completed. The protocol and the state of every study are at /research.

How to read the probability and confidence

The report gives one primary verdict, separate vocal and instrumental probabilities from the service's own labels, and per-window values showing how short spans of the track scored. Those windows are the service's output, not a second analysis by us.

Confidence is a separate axis, and it matters more than most people expect. A high probability with low confidence is not a strong result — it usually means the windows disagreed, the file was short, or it had been re-encoded enough to blur the evidence. Where the evidence is genuinely balanced, the result is reported as inconclusive rather than forced into a category.

We publish no accuracy figure. No controlled evaluation of the current system has been completed by us, so precision, recall and error rates are unmeasured. Any number we quoted today would be marketing rather than measurement.

Why detection is not always perfect

The two categories overlap acoustically, and the overlap is growing. Loudness-maximised human masters share most of their measurable profile with generated output, because generators were trained on loudness-maximised masters. Meanwhile hybrid workflows — generated backing with a human vocal, human performance processed through generative tools — do not belong cleanly in either category to begin with.

Re-encoding degrades everything. A track that has been through a video platform, a messaging app and a screen recording carries far more information about those encoders than about its origin. Short clips are worse still, because there are not enough windows to compare.

And detection ages. Signals that were reliable two model generations ago have been engineered away, which means any accuracy claim is only meaningful against named generators, named versions and a named test set.

This is why results here are always probabilities with stated confidence, never claims of certainty, and never something to act on alone in a copyright, employment or academic decision.

What to do with a result

Treat the report as one input. If it leans generated with high confidence, that is a reason to ask questions, not a conclusion. If it is inconclusive, that is genuine information: the file does not carry a clear enough signature either way.

Provenance settles what acoustics cannot. Project files, stems, alternate takes, rough mixes, session collaborators, a coherent back catalogue — these establish authorship in a way no classifier can. Ask for them first and use the detector to inform the conversation, not to replace it.

What happens to your file

The complete file is uploaded over HTTPS — not an excerpt, and not a set of measurements taken in your browser. Uploads are limited to 25 MB and to recordings between 5 seconds and 15 minutes long.

Our server holds the audio in memory for the duration of one request and forwards it to the third-party detection service. It is not written to disk, object storage or a database by us. Retention and handling inside that service are governed by the provider's own terms; we send no name, email address or account identifier with the audio.

We store three things: the returned result for up to 14 days, keyed by a SHA-256 hash of the file, so an identical repeat upload can reuse it; a salted one-way hash of your network address, to enforce a fair-use limit of 12 analyses per 24 hours; and the analysed duration, for capacity reporting. There is no account, no upload history and no shareable result link.