How accurate are AI music detectors?

The state of the evidence

Detection vendors — Deezer's in-house detector, Vobile's AI Song Detector, and others in our Tools index — publish accuracy claims without the apparatus that would make them verifiable: no public evaluation corpus, no stated class balance (how many human vs. AI tracks in the test), no confusion matrix (how often human music gets flagged as AI), and no independent replication. We store each number with the label it deserves: PARTY CLAIM.

The asymmetry that matters to you: a detector that catches 99% of AI tracks but misflags 1% of human tracks would, at streaming scale, wrongly flag enormous numbers of human-made songs — and platform enforcement increasingly runs on these detectors. The denominator problem is not pedantry; it's your track.

What's coming

Detection accuracy is a measurable question, and measuring it independently — a public corpus of licensed human recordings, disclosed AI generations and hybrids, run against every commercial detector, with the full methodology published — is on this site's benchmark roadmap. Until someone does that (us or anyone else), treat every detector accuracy stat as marketing, and keep your project files as evidence of human authorship in case a false positive lands on you.