How the Detector Handles a Track

001 / Detail

What AI music is, and how it gets made

An AI-generated song comes from a model trained on huge quantities of recorded music. You give it conditioning information — a prompt describing genre and mood, often lyrics, sometimes a reference clip — and it samples a new piece of audio from the patterns it learned.

Nearly every system in use today works in a compressed token space rather than directly on waveforms. A neural codec turns audio into discrete tokens, a transformer or diffusion model generates a token sequence that fits the conditioning, and a decoder converts those tokens back into sound. Most consumer services then run loudness normalisation and limiting, so the file you download already sounds mastered.

Every one of those stages leaves a trace in the finished audio. Detection is the work of hunting for those traces — and of admitting plainly when they simply aren't there.

002 / Detail

What is known—and not known—about the model

We neither train nor own the provider's model, and its internal feature weights are not disclosed here. Published synthetic-audio research discusses spectral, temporal, dynamic, stereo and vocal features, but that literature must not be presented as a description of this provider's implementation.

• Known: the provider returns a primary category, component percentages, confidence bands and sometimes windows.
• Unknown here: its training corpus, feature weights, calibration and internal decision process.
• Consequence: public research can explain the field, but not reverse-engineer this particular result.

003 / Detail

What this application actually does

Everything above describes the field in general. This section describes the product, and the two shouldn't be conflated: a technique used across the industry only counts here if it appears in the list below.

We don't train, own or run a detection model ourselves. When you press analyse, the complete file goes over HTTPS to our server, which forwards it to one specialist third-party AI music detection service and passes back what that service reports. There's 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 other audio tools on this site do run entirely in your browser. The detector can't work that way — classification needs the complete file at a dedicated remote service.

• Implemented: upload, a single 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 to a SHA-256 hash of the file, so an identical repeat upload reuses them rather than re-running the analysis.
• Fair use: a salted one-way hash of your network address caps use at 12 analyses per 24 hours. The address itself is never stored.
• No accuracy figure is published, because no controlled evaluation of the current system has been completed. The protocol and the status of every study are at /research.

004 / Detail

How to read the returned fields

The report gives one primary category plus separate vocal and instrumental percentages supplied by the provider. The application does not average those percentages into an overall score. Any windows shown also come from the provider, not a second pass by us.

Confidence is a separate provider field shown as Low, Medium or High. The interface preserves an inconclusive outcome when the primary category cannot be mapped safely; it does not infer a hidden answer from the other values.

We publish no accuracy figure. No controlled evaluation of the current system exists yet, so precision, recall and error rates are unmeasured. Any number we quoted today would be marketing dressed up as measurement.

005 / Detail

Why detection is never perfect

The two categories overlap acoustically, and the overlap keeps widening. Loudness-maximised human masters share most of their measurable profile with generated output, because generators trained on loudness-maximised masters to begin with. Hybrid workflows — generated backing under a human vocal, a human performance run through generative tools — never fit cleanly into either category in the first place.

Re-encoding degrades everything it touches. A track that has passed through a video platform, a messaging app and a screen recording carries far more information about those encoders than about how it was originally made. Short clips are worse still, simply because there aren't enough windows to compare.

And detection has a shelf life. Signals that were dependable two model generations ago have since been engineered away, which means any accuracy claim only means something against named generators, named versions and a named test set.

That's 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.

006 / Detail

What to do with a result

Treat the report as one input among several. A high-confidence lean toward generated is a reason to ask questions, not a conclusion in itself. An inconclusive result is genuine information too: the file simply doesn't carry a clear enough signature either way.

Provenance settles what acoustics can't. 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 let the detector inform the conversation rather than replace it.

007 / Detail

What happens to your file

The complete file travels over HTTPS — not an excerpt, and not a set of measurements taken locally in your browser. Uploads are capped at 25 MB and at recordings between 5 seconds and 15 minutes long.

Our server holds the audio in memory for the length of one request and forwards it to the third-party detection service. We never write it to disk, object storage or a database. Retention and handling inside that service follow the provider's own terms; we don't send a name, email address or account identifier along with the audio.

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