NEW YORK — Automated systems designed to identify artificial-intelligence-generated music may perform well on clean recordings yet become substantially less reliable when the same kind of audio appears inside an actual broadcast.

A newly published research paper introduced a 40-hour dataset of real television recordings containing both human-made and AI-generated music. The researchers found that detector performance deteriorated as testing moved from clean foreground recordings to synthetic broadcast material and then to genuine broadcast audio.

Most importantly, scores assigned to AI-generated and human-made music overlapped substantially under real-world conditions.

Broadcast audio creates a harder test

Television and radio rarely present music as a clean, uninterrupted master recording. Music may appear briefly beneath speech, pass through broadcast processing, be edited, compressed or mixed with ambient sound.

An earlier study from members of the same research group found that detection performance fell below a 60% F1 score when music was short or appeared in the background. That measurement combines precision and recall; a declining score indicates that the system is making more classification mistakes or missing more relevant examples.

An automated score should begin a review—not end one

The findings support a basic due-process safeguard for distributors, broadcasters, streaming services and music-research companies: an automated result should remain provisional.

Before a recording is publicly labeled, rejected, removed or financially penalized, the creator or rights holder should receive notice and an opportunity to respond. A meaningful review should preserve the detector and version used, the audio examined, the confidence score and the reason the system raised the flag.

If the creator disputes the classification, a qualified human should examine the recording and supporting production documentation before a final decision is issued.

That safeguard does not prevent platforms from investigating synthetic music. It recognizes that a probabilistic tool is evidence—not an infallible verdict.

Sources: Assessing AI-generated music detection in real-world broadcast monitoring; AI-Generated Music Detection in Broadcast Monitoring.