Why did I get a different AI score for the same song?

Almost always because a different part of the recording was analysed. Repeating an identical analysis on an identical file moved our readings by 0.0 percentage points in robustness testing, so the detector itself is deterministic. Changing which excerpt is measured moved them by about 4.4 percentage points on average, and further on recordings whose arrangement changes a lot between sections.

What is an excerpt stability check?

It is an optional extra pass in the detector. For recordings of sixty seconds or more, the tool re-runs the same analysis over two further windows of the same file and reports the spread between the readings. A small spread means excerpt choice is not driving the result. A large spread means the headline number should be read as the middle of a range rather than a fixed value.

Does a stable score mean the result is correct?

No. Stability measures agreement, not truth. A recording can measure consistently AI-like across every window and still be a human performance — loud, quantised, synthesiser-heavy music does exactly that. Stability tells you the number is dependable as a number; provenance tells you what actually happened.

Which files tend to be unstable?

Hybrid tracks where only part of the material was generated, compilations and edits stitched from different sources, live recordings whose room and dynamics shift, long arrangements that move between sparse and dense sections, and anything with heavy dynamic mastering applied unevenly.

Does the stability check upload my audio?

No. Every pass runs in your browser on the audio buffer already decoded on your device. Nothing is transmitted, nothing is stored, and the buffer is released as soon as the analysis finishes.

Why a score changes between parts of a song

A detection score is a measurement of an excerpt, not a property of a song. This page explains where that variance comes from, what we measured, and how to check it on your own file.

Last updated August 2026

Short answer

Why does an AI music detector give different scores for the same song?

Because it is measuring a specific window of audio, not the whole work. Repeating an identical analysis on an identical file moved our readings 0.0 percentage points, so the detector is deterministic. Changing the excerpt moved them about 4.4 percentage points on average. When two parts of a recording disagree strongly, the honest reading is a range, not a number — and that is usually a sign the arrangement or the source material changes across the track.

Measured robustness data →

The detector is deterministic; the audio is not uniform

Two different things get confused here. The first is whether a tool returns the same answer twice for the same input. Ours does: in our robustness testing, repeating an identical analysis on an identical file produced a spread of 0.0 percentage points. There is no random seed, no sampling, no server-side model whose weights change under you.

The second is whether every part of a recording measures the same. It does not, and it should not. An intro with one synthesiser pad and no drums has a different spectral balance, different transient behaviour and a different stereo field from the final chorus of the same song. The measurements that feed a detection score are describing that audio, so they move with it. Across our test set, changing which excerpt was analysed shifted readings by about 4.4 percentage points on average — small enough to ignore in the middle of the range, large enough to flip a borderline call.

What moves a reading, ranked

Arrangement density. Sparse sections give the analysis less to work with and tend to read as more synthetic than the full band section of the same track.
Vocal presence. Instrumental passages and sung passages measure differently. A track analysed over its instrumental bridge is a different measurement from the same track analysed over its chorus.
Mastering applied unevenly. Dynamic processing that clamps loud sections harder compresses the very micro-dynamics the analysis reads.
Edits and stitches. Compilations, remixes and hybrid tracks contain genuinely different source material in different places, so genuinely different readings are the correct output.
Encoding. A separate effect, and a smaller one: re-encoding the same excerpt moved our readings about 2.4 percentage points. See the compression study.

Running the check on your own file

For recordings of sixty seconds or more, the free AI music detector offers an optional excerpt stability check. Tick it before you analyse and the tool runs the same analysis over two further windows of the same recording, then reports the spread between all three readings. It takes roughly three times as long and, like everything else here, runs entirely in your browser with nothing uploaded.

Read the output like this:

Spread under about 5 pp — stable. Excerpt choice is not driving the result. Quote the headline number, still as a probability rather than a verdict.
Spread of roughly 5 to 15 pp — moves with the excerpt. Report a range. Give provenance more weight than the acoustic reading.
Spread above about 15 pp — unstable. Do not quote a single score for this file at all. Ask what changes across the recording, and whether parts of it came from different places.

Stability is not accuracy

This is the part people skip. A stable reading means the measurement is repeatable across the recording. It says nothing about whether the interpretation is right. Loud, quantised, synthesiser-led human music measures consistently AI-like from start to finish — in our electronic-music study, three of twelve human synthesiser renders read above 70 percent. The stability check would have called all three stable, and all three were wrong.

Use stability to decide how precisely to state a number, and use the limitations page to decide what that number can be used for. Provenance — stems, project files, dated drafts, platform disclosures — outranks every acoustic measurement on this site, including this one.

Questions

Why did I get a different AI score for the same song?: Almost always because a different part of the recording was analysed. Repeating an identical analysis on an identical file moved our readings by 0.0 percentage points in robustness testing, so the detector itself is deterministic. Changing which excerpt is measured moved them by about 4.4 percentage points on average, and further on recordings whose arrangement changes a lot between sections.
What is an excerpt stability check?: It is an optional extra pass in the detector. For recordings of sixty seconds or more, the tool re-runs the same analysis over two further windows of the same file and reports the spread between the readings. A small spread means excerpt choice is not driving the result. A large spread means the headline number should be read as the middle of a range rather than a fixed value.
Does a stable score mean the result is correct?: No. Stability measures agreement, not truth. A recording can measure consistently AI-like across every window and still be a human performance — loud, quantised, synthesiser-heavy music does exactly that. Stability tells you the number is dependable as a number; provenance tells you what actually happened.
Which files tend to be unstable?: Hybrid tracks where only part of the material was generated, compilations and edits stitched from different sources, live recordings whose room and dynamics shift, long arrangements that move between sparse and dense sections, and anything with heavy dynamic mastering applied unevenly.
Does the stability check upload my audio?: No. Every pass runs in your browser on the audio buffer already decoded on your device. Nothing is transmitted, nothing is stored, and the buffer is released as soon as the analysis finishes.