As AI-generated music moves into professional workflows, producers often mix synthetic stems—drums, basslines, or vocals—with human-performed instruments. Existing AI-music detectors treat this as a binary choice: a track is either fully AI or fully human. A new paper from researchers including Fernando Garcia de la Cruz and colleagues, accepted at ISMIR 2026, reframes the problem as estimating a continuous "AI energy ratio" between 0 and 1.
The authors built mixtures from a multi-track dataset, combining human-performed stems with AI-reconstructed versions obtained via a neural audio codec. They first showed that a CNN-based binary detector with over 99% accuracy on pure tracks behaves poorly on mixed content: its output rises with the AI stems' energy contribution, but it is noisy and miscalibrated as an estimator. They also found that detection sensitivity varies by instrument, with drums and guitar showing strong codec-artifact signatures while vocals and bass are harder to detect.
Based on these insights, the team trained a similar CNN to regress the AI energy ratio directly. On held-out mixtures from the same pipeline, it achieved a mean absolute error of 0.076 and an R² of 0.85. The authors describe this as an initial but promising step toward AI-music detection that matches real production conditions, where hybrid mixtures are increasingly the norm.