How Much AI Is in This Track? The Detector That Says Human While Half the Song Is Machine

A detector that identifies AI-generated music with 99.97% accuracy will still call a track human while nearly half its sound came out of a machine. Researchers from Pompeu Fabra and BMAT built 21,212 music mixtures with a known AI share to prove it, then changed one thing, the question the model was asked, and recovered the missing information: average error dropped from 0.289 to 0.076. They also found the machine's fingerprint sits in the upper frequencies, so a generated drum loop gets caught and a generated bassline walks straight past. The caveat is the authors' own: these are codec reconstructions rather than Suno output, and they call the work a foundation, not a product.