A song is rarely one thing. Pull apart a finished track and you get stems: the isolated vocal, the drums, the bass, the guitars, each recorded or programmed separately before an engineer blends them into a mix. That modular structure is exactly what makes today's music-AI question harder than it first appears. A track is not simply human or machine-made. A producer can write and sing the melody themselves, then ask a generative model to fill in the accompaniment. What comes out is a hybrid, and until now nobody had published a study on how to detect one.
François Rigaud, Gabriel Meseguer-Brocal, Benjamin Martin and Romain Hennequin take on that gap in a paper accepted at ISMIR 2026, the main international conference on music information retrieval. Their starting point is a finding from earlier work: AI music detectors do not really recognize style or musicality. They latch onto artifacts left by the decoder, the component in a generative audio system that converts a compressed internal representation back into a waveform. Those artifacts are subtle, consistent, and largely inaudible, and they work like a manufacturing stamp on the audio.
The obvious worry is that a stamp on one stem gets smeared or buried once that stem is mixed underneath a human-recorded one. So the team built a controlled test. They used MUSDB18-HQ, a standard research database of multitrack music, and simplified each song to two stems: vocals and accompaniment. To simulate a hybrid without needing an actual AI songwriter, they passed one stem at a time through a neural codec, an autoencoder that compresses audio and reconstructs it. The reconstruction sounds close to the original but carries the same kind of decoder fingerprint that a fully generated track would. Mix that processed stem back with the untouched one, and you have a synthetic hybrid whose ground truth you know exactly.
Separating first turns out to be the wrong move
The intuitive approach is a two-step pipeline. Run a source separation system, the technology behind vocal-isolation and karaoke tools, to pull the mix back apart, then run an AI-music detector on each recovered stem. The team tried it, and it did not hold up. Generic separation systems, they report, do not reliably recover the artifacts tied to the AI-generated stem. That makes a certain sense once you consider what a separator is trained to do: reconstruct a clean-sounding voice or backing track. It is a generative process of its own, and it imposes its own fingerprints while erasing the faint ones the detector was hunting for. The evidence gets destroyed in the act of retrieving it.
So the authors restructured the problem. Instead of putting separation upstream of detection, they run the two in parallel and give separation a much narrower job: estimate how much of the mixture's energy belongs to each source. Not what the vocal sounds like, just how loud it is relative to everything else. That number is robust in a way an isolated waveform is not.
The detector then works on the mix as a whole, chopped into short chunks. For each chunk, a small binary classifier receives two inputs: the mix-level prediction of whether AI-generated audio is present, and the relative energy of the stem in question. Each stem gets its own classifier. The reasoning is that these two signals together are informative in a way neither is alone. A strong AI signal in a passage where the vocal dominates points differently than the same signal where the vocal is barely there. Averaging the chunk-level verdicts across a whole song gives what the authors call encouraging track-level results.
Why it matters
The practical questions around AI in music are mostly hybrid questions. Streaming platforms are introducing disclosure labels, rights organizations are arguing over how royalties should split when a machine contributed, and producers are already using generative tools for a bassline here or a pad there. A detector that only answers "was this whole track generated" cannot help with any of that. One that can point at a specific stem is a different kind of instrument.
It is worth being careful about what this paper actually establishes. The hybrids were simulated by autoencoding real stems, not produced by a music generation model, so the artifacts are a proxy for the real thing rather than the real thing. The setup used two stems in a single database. The authors describe their track-level numbers as encouraging and frame the work as showing potential, and the paper is a preprint of a conference submission, not a deployed system. Nobody should read it as a tool ready to adjudicate a royalty dispute.
What it does offer is a useful negative result and a reasonable design response. Detection built on unmasking hidden fingerprints breaks when you try to unmix the audio first, because unmixing is itself a rewriting of the signal. Working from the intact mixture, and asking separation only for a number rather than a sound, keeps the fingerprints where the detector can still read them.