How Much AI Is in This Track? Detecting AI-Generated Content in Hybrid Music Mixtures
Speaker at Audio Developer Meetup Berlin | August 2026 | Ableton AG, Berlin
Presented at August Meetup: Interactive Audio, on a shared bill with Robyn Pacetti (Protagonists!). A talk on my MSc thesis work.
🎥 Recording coming soon. The video isn’t published yet, so check back here.
Talk Description
AI-generated music is reshaping the music industry. By early 2026 it made up around 44% of daily uploads on Deezer, and studies suggest most listeners can’t tell it apart from human-made recordings. This has driven a major research effort on automatic detection, almost all of which asks one binary question: is this track AI or human? State-of-the-art detectors now hit over 99% accuracy on that question.
But that isn’t how AI actually shows up in modern production. Producers use it as a modular tool: an AI drum loop combined with a synthetic bassline and a human singer’s voice. The result is hybrid material that’s neither fully human nor fully AI, exactly the case the binary framing can’t describe.
This talk covers my work reframing detection for that reality. Instead of a yes/no label, I treat it as a regression problem: estimating a continuous AI energy ratio, the fraction of a track’s acoustic energy that comes from AI stems. I’ll explain why these detectors work at all (they exploit subtle spectral fingerprints left by neural audio codecs), and show how a model trained to estimate that ratio reveals patterns a binary detector misses entirely.
The broader takeaway: near-perfect binary accuracy is not evidence that a detector is ready for realistic, mixed-content music. As AI becomes a normal part of production, the question worth asking isn’t whether a track is AI, but how much.
Key Topics Covered
- Why binary AI-music detection falls short of real production workflows
- Neural audio codec artifacts as the signal detectors actually exploit
- Reframing detection as regression: the AI energy ratio
- Building synthetic hybrid mixtures with known AI proportions
- Per-instrument detectability: why drums and guitars give AI away and vocals don’t
- What near-perfect benchmark accuracy does and doesn’t tell you