Compression and AI music detection: what survives an MP3

Why this is the most important variable

Popular explanations of AI music detection lean heavily on the frequency spectrum: look for a hard ceiling, look for unnaturally smooth high-frequency texture, look for missing air. Every one of those cues lives in the region a lossy encoder attacks first.

An encoder’s psychoacoustic model spends its bit budget where human hearing is most sensitive and throws away what it believes you cannot hear. Unfortunately, “what you cannot hear” and “where the synthesis artefacts live” overlap almost perfectly. The result is that a 128 kbps MP3 of a human orchestra and a 128 kbps MP3 of a generated track can look far more similar to each other than either does to its own lossless original.

This is not a caveat at the bottom of a results page. It is the dominant source of error in real-world AI music detection, ahead of model quality.

Pre-registered expectations

The table below is our stated prediction, published before the evaluation runs. If the data contradicts it, the table stays and the correction is published next to it.

Study design

A single corpus of tracks, half human-produced and half generated, each rendered from a lossless master into the ladder above using a standard encoder at fixed settings. Every file is analysed by the same engine build, and results are reported per source condition rather than pooled.

• Fixed corpus. The same recordings across all conditions, so any change in the reading is attributable to the encoder and nothing else.
• Per-feature reporting. Not just the final probability, but each component’s output, so we can say which feature failed rather than only that the score moved.
• Confidence calibration. The question that matters most is not whether accuracy drops at 128 kbps — it obviously does — but whether the confidence level drops with it. A detector that stays confident while its evidence disappears is worse than one that is simply wrong.
• Second-generation encoding. A separate arm re-encodes already-lossy files, because that is what actually happens when audio moves between platforms.

How the engine already compensates

Waiting for results is not an excuse to ignore the problem, so the current build already treats encoding evidence as a modifier rather than a signal:

• The spectral ceiling feature is down-weighted when other measurements indicate a lossy source, because a codec ceiling and a model ceiling are not distinguishable from the file alone.
• Evidence of codec artefacts lowers the confidence level rather than pushing the probability in either direction. Compression is not evidence of generation.
• Segment agreement survives compression better than any single spectral measurement, so it carries more of the reading on degraded files — and when segments disagree, the outcome is reported as inconclusive.
• Dynamics measurements such as crest factor are largely preserved by lossy encoding, but they are also the features most easily mimicked by ordinary mastering, so they cannot be leaned on alone.

What to do with a compressed file

Get a better copy if one exists — a purchased download, a distributor file, or the original export from whoever made it. If not, analyse what you have but read the confidence level first and the probability second. And never convert a lossy file to WAV in the belief that it improves anything; it changes the container, not the content.

Above all, remember what the whole exercise is worth. Provenance evidence — project files, upload history, someone who can explain their own arrangement — beats every acoustic measurement on this page, at every bitrate.