AI music detectors fail on real TV broadcasts, new 40-hour dataset shows

AI-generated music is flooding broadcast media, raising urgent questions about transparency and fair compensation. But a new paper accepted for ISMIR 2026 reveals that current detection systems are far from reliable in real-world conditions. Researchers David López-Ayala and colleagues built BAMM (Broadcast AI-Music Monitoring), the first dataset of its kind: 40 hours of actual television recordings containing both AI-generated and human-made music. They then evaluated CNN-based detectors trained on clean audio versus broadcast-oriented data across three escalating scenarios: Clean Foreground Music (CFM), Synthetic TV Broadcast (STB), and Real TV Broadcast (RTB).

The results are sobering. Both models achieved near-perfect accuracy on clean clips (CFM), but performance collapsed under synthetic broadcast conditions (STB). Adding broadcast-conditioned training improved robustness, yet scores on real TV broadcasts (RTB) were dramatically worse, with substantial overlap between AI-generated and human-made music. The authors argue this exposes a critical domain gap: current CNN training approaches remain insufficient for reliable AI-generated music detection in broadcast monitoring. As AI music proliferates across television, radio, and streaming, these findings underscore the need for more sophisticated detection methods and better real-world datasets to protect royalty systems and content authenticity.