Is This Song AI Generated?

How to answer the question responsibly

Think of it as three independent lines of evidence, weighted in this order. Provenance is the strongest: project files, stems, dated drafts, a coherent back catalogue, and a straightforward explanation from the artist. Context is second: how the track appeared, who uploaded it, how quickly, alongside how much other material, and whether the release pattern is plausible for a human working at that pace.

Acoustic analysis is third. It is the only one you can run yourself in thirty seconds, which is exactly why it gets over-weighted. Use it to raise or lower your suspicion, not to settle the matter.

• Provenance: session files, stems, drafts, release history — strongest
• Context: upload pattern, volume of output, account age — moderate
• Acoustic analysis: what the detection service reports — supporting evidence only
• Vibes: how the track 'feels' — the weakest and most bias-prone signal of all

How detection works

The complete audio file is sent over an encrypted connection to our server, which forwards it to a specialist third-party AI music detection service. This service performs the classification using its own models and returns a verdict along with vocal and instrumental probabilities.

We do not run any detection models locally in your browser. Audio is held in memory for the duration of the request and is never written to storage by us; we store only the numeric result for up to 14 days.

Why the answer is often wrong in both directions

Human music produces false positives routinely. Loudness-maximised electronic tracks, template-driven pop, tightly quantised programming and low-bitrate uploads all reproduce the same measurable signature as much generated audio.

Generated music produces false negatives just as routinely. Newer models leave fewer artefacts, and any human post-production — re-recording, re-mixing, adding a live instrument, running the file through analogue gear — removes most of what remains.

This is not a flaw in one tool. It is the current state of the entire field, and any product that hides it from you is selling confidence rather than information.

How to check if a song is AI generated, step by step

Short answer: get the best copy of the audio you can, run a signal analysis, read the file's metadata, check the artist's provenance and release history, listen critically, then weigh those five inputs together — in that order of strength, with the detector score fourth rather than first.

Audio quality decides everything downstream. A screen recording, a phone capture or a 96 kbps stream rip strips out most of what analysis depends on, so ask for the original file and aim for 45 seconds or more of continuous music. Run the detector twice on different excerpts: agreement across excerpts is worth far more than any single reading.

Metadata is often more revealing than acoustics. Encoder strings, creation dates, tool names and comment fields in ID3 or Vorbis tags sometimes carry traces of a generation or distribution pipeline. Metadata is trivially editable, so its presence is suggestive while its absence proves nothing. Where a track carries Content Credentials or a C2PA provenance manifest, that beats any detector, because it is cryptographic rather than statistical.

• Cryptographic provenance, or a direct admission — decisive
• Stems, dated drafts, session files, live recordings — very strong
• Release-history anomalies: twelve albums in one month, no live footage, unrelated genres with identical production
• Metadata traces — suggestive, easily faked
• Detector output — supporting evidence
• Subjective listening impressions — weakest

Which acoustic signals actually carry information

Detection is a measurement problem, not a lookup. Early generators left obvious traces — smeared transients, a hard frequency ceiling, metallic reverb tails, repeating micro-textures — and those were the first things generator developers fixed. What remains is statistical: distributions of energy, movement and structure that differ slightly, on average, between generated and performed recordings.

"Slightly, on average" is the crucial phrase. It means detection works better across many tracks than on any single one, which is the opposite of what someone checking one song wants.

• Spectral artefacts: where energy stops, how sharply, and how the noise floor behaves above it — heavily confounded by lossy encoding
• Structural regularity: uniform tonal balance and section-to-section similarity — also describes template-driven pop
• Dynamic behaviour: crest factor and micro-dynamics — destroyed by ordinary mastering
• Timbral movement: how much the spectral centroid wanders over time — genre-dependent
• Stereo behaviour: correlation and width consistency — a production choice as much as an origin signal

Why general AI audio detectors fail on songs

Tools marketed as "AI audio detectors" were almost always built for synthetic speech, and pointing one at a mixed, mastered song is a category error. Speech detection has advantages music does not: a single dry source, a narrow bandwidth, predictable phonetic structure, and large labelled corpora of both real and synthetic examples.

A song is a deliberate composite — many sources, heavy processing, wide bandwidth, and mastering applied on top of everything. The artefacts a speech model looks for are either absent or buried under production. That is why this service uses a dedicated music-detection model, and why a low or high score from a speech-oriented tool tells you almost nothing about a track.

Before you accuse anyone

AI-assisted and AI-generated music is legal. Plenty of musicians use generative tools openly and say so. A detector score is not grounds for a takedown, a failing grade, a disciplinary process, a contract termination or a public callout — and this result is explicitly not forensic evidence.

A high score with a solid provenance trail usually means the detector is reacting to production style, not origin. A low score with no provenance at all means very little, because post-production removes detectable artefacts. In both cases the honest output is "undetermined".

If the question genuinely matters, ask the artist for the project files. That single request resolves more cases than any acoustic tool ever will.