Finding the Noise: Catching AI Music From Generators Nobody Has Named Yet

Deezer reported that 18% of the new music arriving on its platform in June 2025 was fully AI-generated, and less than a year later that figure was approaching half. Every detector built for this problem learns by example, which means a newly launched service or a fresh model version walks straight past it. Two Deezer researchers solved that by describing real music instead, treating it as statistical noise and letting the repeating machine patterns separate themselves out with no labels at all. The result catches Suno 99.9% of the time without ever seeing a labelled Suno track, tells model versions apart, and exposed one service quietly running another company's generator.