Support

How do you decide a song is 85% AI?

The 85% figure comes from our AI Song Checker, a single likelihood score. Two models listen to your track:

• The spectral model turns the audio into a picture of its frequencies and looks for the texture patterns AI generators tend to leave behind.
• The temporal model follows how the music moves over time, checking whether the flow feels human or has the quirks AI tools produce.

The two blend into one number, weighted by how long the song is. For a short clip it's almost entirely the spectral read; for a full-length song it's mostly spectral with the temporal model filling in the rest. The score also folds "pure" and "processed" (hybrid) AI likelihood into that one percentage, so a heavily-processed AI track can clear 85% just like a fully-generated one.

A few things worth knowing:

• The song needs to be at least 30 seconds, and accuracy improves the longer the sample.
• The detector is independently measured at 99.4% accurate, but no detector is perfect. If we get it wrong, thumbs-down the result in the checker and we'll review.
• There's no setting or phrasing that moves the number. It's based on the audio itself.

For what happens when a track crosses the line, see What is your policy on AI Music?