Does a detector mistake human electronic music for AI?

The question

Detection features are, for the most part, performance features. Micro-timing that drifts because a human played it, a room that colours the recording, a bass note fingered slightly differently the second time around, a stereo field built from real microphone placement. Generated audio tends to lack those, which is exactly why the measurements work at all.

Human electronic music also lacks them, by intent. A grid-locked techno arrangement programmed in a DAW with software instruments has no room, no performance variation and no acoustic source. So the honest question is not whether such a track can be flagged, but how often, and whether the engine at least admits uncertainty when it happens.

What was measured

The corpus is the same one behind the engine behaviour study: twelve tracks of 20–23 seconds at 44.1 kHz stereo, rendered procedurally by this site’s own sample generator across twelve genre presets. Every track is synthesiser-only, quantised, single-take material with no recorded performance and no third-party audio.

Each track was analysed under seven conditions — lossless WAV, MP3 at 320, 192, 128 and 64 kbps, a mono downmix, and a 10-second excerpt — twice each, for 168 detections on build ensemble-1. Because none of the material is AI-generated, every reading above 50% is a reading in the wrong direction.

Results

Three of twelve tracks read above 70% AI probability. Three of the twelve renders — all synthesiser-only, quantised, single-take material — produced high AI probabilities. Purely electronic human-authored music is the hardest control case, and this study confirms the risk rather than hiding it.

The engine abstained almost everywhere else. On single-render synthetic reference audio the engine declined to issue a verdict in almost every case. It prefers abstaining to guessing — which is the designed behaviour, and also a reminder that an inconclusive result is a normal outcome.

Those two findings belong together. The engine’s conservatism is what keeps the failure from becoming an accusation in most cases: below the moderate confidence band it reports inconclusive instead of a verdict. But conservatism is not immunity — in a quarter of these renders the confidence was high enough to produce a high-probability reading on music a human wrote.

How to read this

• A high score on electronic music is weak evidence. The features that drove it are the same features that describe ordinary in-the-box production.
• An inconclusive result is the normal outcome. It is not a fault, and it is not a soft yes.
• Genre changes the prior. Live acoustic material is the easy case; quantised synthesiser material is the hard case. Any detector that reports a single accuracy figure across both is hiding this.
• No file-only measurement settles authorship. Provenance — project files, stems, version history, someone who can explain their own arrangement — outranks every number on this page.

Limitations

• This is not a false-positive rate. That requires verified commercially released human recordings; this corpus is procedurally generated, so it is reproducible but not representative.
• The corpus contains no AI-generated tracks, so nothing here says anything about detection rate.
• Renders are 20–23 seconds, below the 30 seconds the tool asks for, which suppresses segment agreement and pushes results toward abstention.
• All figures apply to build ensemble-1. A configuration change invalidates them, and they will be re-measured and re-stamped rather than carried forward. See the changelog.

What we do about it

The planned control arms include sixty fully synthetic human-authored electronic tracks precisely because of this result, and the false-positive rate — not the detection rate — is the metric we treat as primary. A false accusation costs a musician more than a missed detection does anyone.