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 audio features are analysed

The detector decodes your file, samples up to five sections spread across the track, and runs a Hann-windowed FFT over each. Six measurements come out of that, and each one is a proxy for something rather than a direct observation of AI.

• Spectral ceiling — the frequency above which almost no energy remains. A hard, round-numbered shelf can indicate generation into a bandwidth-limited representation, but a lossy encode produces the same shelf.
• Cross-segment tonal agreement — how similar the tonal balance of one section is to another. Real performances drift; whole-track generation often does not.
• Crest factor — how far peaks sit above the average level, a measure of surviving dynamic range. Weakened badly by modern mastering on both sides.
• Spectral-centroid variability — how much the perceived brightness moves over time. Performances breathe; some generated material holds unusually still.
• High-band energy ratio — how much of the total energy lives in the top octaves, which can expose synthetic upsampling or its absence.
• Stereo correlation — whether the image looks captured by microphones in a room or constructed from separately placed elements.

How the probability and confidence score are calculated

Each measurement contributes evidence in one direction, weighted by how much that signal is worth on its own. The weighted evidence is combined into a single estimated probability, reported as a range rather than a false-precision figure.

Confidence is a separate axis, and it matters more than most people expect. It reflects how much the individual signals agree with each other and how much usable audio the file provided. Six signals pointing the same way on a clean three-minute WAV gives high confidence. Three signals pointing one way and three the other, on a forty-second re-encoded clip, gives low confidence — and if the evidence is genuinely balanced the result is reported as inconclusive rather than forced into a category.

A high probability with low confidence is not a strong result. Always read both numbers together, and read the per-signal findings underneath them, which say in plain English what each measurement actually observed.

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 YouTube, 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 sections 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 acoustic 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 spectrum analyser can. Ask for them first and use the detector to inform the conversation, not to replace it.

Privacy during analysis

Every step described above runs on your device. The file is read by your browser, decoded with the Web Audio API and analysed in memory; no copy is transmitted anywhere. There is no account, so there is no record of what you checked.

Deletion is automatic in the strongest sense available: the decoded audio exists only for the duration of the analysis and is discarded when you remove the file or close the tab.