Is This Song AI Generated?

001 / Detail

A responsible way to answer the question

Treat it as three separate lines of evidence, weighted in this order. Provenance carries the most weight: project files, stems, dated drafts, a consistent back catalogue, a straightforward explanation from the artist. Context comes second — how the track surfaced, who uploaded it, how fast, alongside how much other material, and whether that release pace is realistic for a human.

Acoustic analysis comes third. It's the only piece you can run yourself in thirty seconds, which is exactly why people give it too much weight. Use it to nudge your suspicion up or down, never to close the case.

• Provenance: session files, stems, drafts, release history — carries the most weight
• Context: upload pattern, output volume, account age — moderate weight
• Acoustic analysis: the detection service's reading — supporting evidence only
• Gut feeling: how the track 'feels' — weakest signal, most prone to bias

002 / Detail

What happens when you run a check

The full audio file travels over an encrypted connection to our server, which hands it to a specialist third-party AI music detection service. That service runs its own models and returns a verdict along with separate vocal and instrumental probabilities.

No detection model runs locally in your browser. Audio sits in memory only for the length of the request and is never written to storage on our end — we keep just the numeric result, and only for up to 14 days.

003 / Detail

Why the reading is wrong in both directions, often

Human-made music trips false positives all the time. Loudness-maximised electronic tracks, template-built pop, tightly quantised programming and low-bitrate uploads all produce the same measurable footprint as a lot of generated audio.

Generated music slips past just as often. Newer models leave fewer artefacts to find, and virtually any human touch afterward — re-recording, remixing, layering in a live instrument, running it through analogue gear — wipes out most of what remains.

This isn't a shortcoming unique to one tool. It's where the whole field stands right now, and any product that doesn't tell you that is selling confidence rather than information.

004 / Detail

A step-by-step way to check a song

Start with the best lawful source available, run the classification, inspect metadata, review provenance and release history, and listen critically. Do not convert those observations into a claim stronger than the underlying evidence.

A screen recording, phone capture or low-bitrate stream changes the audio before classification. Prefer the original file when authorised. The product accepts recordings from 5 seconds through 15 minutes, but it does not promise a particular confidence level for clip length or format.

Metadata often reveals more than the sound itself. Encoder strings, creation timestamps, tool names and comment fields buried in ID3 or Vorbis tags can carry traces of how a file was made or moved. It's trivially easy to edit, so its presence is suggestive and its absence proves nothing. A track carrying Content Credentials or a C2PA provenance manifest beats any detector outright, because that's cryptographic proof rather than a statistical guess.

• Cryptographic provenance, or a direct admission — settles it
• Stems, dated drafts, session files, live recordings — very strong
• Release-history oddities: a dozen albums in a month, no live footage, unrelated genres mastered identically
• Metadata traces — suggestive, easy to fake
• Detector output — supporting evidence
• Gut impressions from listening — weakest of all

005 / Detail

Which acoustic signals actually tell you anything

This is a measurement problem, not a lookup table. Early generators left obvious fingerprints — smeared transients, a hard ceiling on frequency, metallic reverb tails, repeating micro-textures — and those were the first things developers patched out. What's left now is statistical: distributions of energy, movement and structure that differ, on average and only slightly, between generated and performed recordings.

'Slightly, on average' is the phrase to hold onto. It means detection performs better across a large batch of tracks than it does on any single one — which is the opposite of what someone checking one song actually needs.

• Spectral artefacts: where energy cuts off, how sharply, and how the noise floor behaves above it — badly confounded by lossy encoding
• Structural regularity: uniform tonal balance and repeated section shapes — also describes template pop
• Dynamic behaviour: crest factor and micro-dynamics — flattened by ordinary mastering
• Timbral movement: how far the spectral centroid drifts over time — depends heavily on genre
• Stereo behaviour: correlation and width consistency — as much a mixing choice as an origin clue

006 / Detail

Why general-purpose AI audio detectors miss songs

Tools branded as "AI audio detectors" were nearly always built for synthetic speech, and aiming one at a mixed, mastered song is the wrong tool for the wrong job. Speech detection has advantages music simply doesn't offer: one dry source, a narrow bandwidth, predictable phonetics, and large labelled datasets of both real and synthetic samples.

A song combines many sources and layers of processing. A speech-oriented service and a music classifier therefore answer different questions. This product submits musical audio to a specialist third-party music-classification service; it does not perform speaker identity or voice-clone attribution.

007 / Detail

Before you accuse anyone of anything

The legal status of AI-assisted or generated music depends on jurisdiction and facts. A detector result is not grounds by itself for a takedown, failing grade, disciplinary action, terminated contract or public allegation, and it is not forensic evidence.

When classification and provenance disagree, do not invent an explanation from the output. Review the source quality, request production records and record the matter as undetermined unless stronger evidence resolves it.

If it genuinely matters, ask the artist for the project files. That one request settles more cases than any acoustic tool ever will.