Detectors Say 87%. What Does That Mean?

It is tempting to imagine a detector as a very good listener — a machine with an ear for the uncanny quality of synthetic vocals. That is mostly not what is happening. The strongest signal a detector has is not aesthetic at all: it is a machine-readable identifier that the generator deliberately embedded in the file at export.

Detection generally draws on four layers, in descending order of reliability. First, embedded identifiers — Google's SynthID watermark and C2PA content credentials, which are designed to be found. Second, spectral fingerprints, structural traces a generator leaves in the audio itself. Third, statistical regularity— synthesised audio tends to be more uniform in timing, tuning, and dynamics than a room full of people playing. Fourth, catalog matching, checking whether the track resembles something already known.

That ordering explains nearly every strange result people report. Strip or degrade the first two layers — through processing, re-encoding, or editing — and the detector falls back on statistical inference, which is far shakier. It also explains the false positives: a meticulously produced human track, quantised to a grid and tuned to the cent, looks statistically a lot like a machine made it.

Search for experiences with AI music detectors and the dominant theme is inconsistency: the same file scored 12% by one tool and 91% by another, human-made tracks flagged as synthetic, obvious AI output waved through. None of that is surprising once you know the layer model. Different tools weight the four signals differently, so a file that has lost its embedded identifiers lands wherever each tool's statistical guesswork puts it.

Three honest limits are worth stating plainly. Published accuracy figures are largely unverifiable — a percentage means little without the test set and the false-positive rate beside it, and those are rarely shared. The binary framing is wrong for most modern music: a track with an AI-generated instrumental, a human vocal, and conventional mixing is not meaningfully "AI" or "not AI," yet detectors must return one number anyway. And the target keeps moving — detectors are tuned against current generator output, while generators keep changing, so accuracy measured last year describes a different problem than the one you have today.

The practical reading: use a detector as one input among several. A high score on a raw export tells you something real. A high score on a processed, hybrid, or unusual track tells you mainly that the tool is uncertain.

If you generate music you own the rights to and intend to release it, the useful takeaway from all of this is not "how do I beat a detector." It is that a raw export is an unfinished file. It carries embedded identifiers, sits outside platform loudness targets, and may overlap with existing recordings — three separate reasons automated screening holds a release, only one of which is about AI at all.

Preparing the file addresses all three. Clean the export so the embedded artifacts are gone, master it into platform spec, and check it for fingerprint collisions before you distribute rather than after. For the cleanup step we recommend Undetectr, which handles the artifact pass and a mastering pass in one run and includes a fingerprint check — our full review covers it in detail. The AI audio cleanup guide explains the order of operations regardless of which tool you use.

Where the line sits is worth saying directly: this is preparation for music you made and own, and it does not extend to impersonating real artists, uploading at spam volume, or hiding AI involvement where a platform asks you to declare it. Those are the behaviours detection exists to catch, and they are the ones that get accounts terminated rather than tracks rejected.