Microscopic software fingerprints can now spot AI-generated music with nearly 100 percent accuracy.
April 20, 2026
Original Paper
ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics
arXiv · 2604.16254
The Takeaway
ArtifactNet detects AI music by looking for the forensic residuals imprinted by neural audio codecs. These physical artifacts are tiny distortions that occur during the generation process and do not exist in human recordings. Instead of trying to analyze the melody or the lyrics, the system treats the audio file like a physical crime scene. This method is much more effective than current AI detectors that try to understand the music style. Copyright holders and platforms can now reliably flag AI content even if it sounds perfectly human.
From the abstract
We present ArtifactNet, a lightweight framework that detects AI-generated music by reframing the problem as forensic physics -- extracting and analyzing the physical artifacts that neural audio codecs inevitably imprint on generated audio. A bounded-mask UNet (ArtifactUNet, 3.6M parameters) extracts codec residuals from magnitude spectrograms, which are then decomposed via HPSS into 7-channel forensic features for classification by a compact CNN (0.4M parameters; 4.0M total). We introduce Artifa