A new arXiv study says AI music detectors are pretty good at catching machine-made songs - just not as good at ignoring ordinary remixes.
Researchers built a YouTube-sourced dataset pairing AI-generated, edited, and original versions of the same songs, then trained a binary classifier to separate AI tracks from edited ones. The system fed 10-second raw-waveform clips into a pretrained PaSST spectrogram transformer, with data split by original song to prevent leakage between train and test sets. At the video level, the final detector reached 0.811 balanced accuracy. Clip-level results split unevenly: AI-generated clips scored an F1 of 0.836, while edited clips lagged at 0.720.
That gap matters because edited audio - pitch-shifted, re-encoded, remixed - is exactly what shows up in real uploads, not pristine studio files. A detector that struggles to tell a remix from a synthetic track will misfire on ordinary human-made content once it leaves a controlled test set. The researchers also ran Grad-CAM visualizations to check whether confident predictions were keying on specific time-frequency regions of a clip rather than reacting to the whole thing.
Detection tools keep getting graded on tidy datasets; real platforms are messier, and messier is where these numbers will actually be tested.