A listening test reported by MIT Technology Review found that participants struggled to separate AI-generated tracks from human-made music. The result was not a formal universal measure of human detection, but it showed how convincing short AI music examples could already be.
The test used twelve genres
MIT reporter James O'Donnell generated 30-second tracks in 12 genres with Udio, mixed them with human-made songs, and asked newsroom colleagues to identify which was which.
The average score was 46%. One composer scored 50%, while a creativity researcher scored 66%. Instrumental examples including classical piano, jazz, and pop were among the tracks participants found difficult to classify.
The broader MIT Technology Review site is available here.
The result challenges easy detection
The test suggests that listeners cannot always identify AI music from sound alone, especially in short clips. It does not mean every AI track is indistinguishable from human music or that a 46% score would repeat across every audience and genre.
The source also describes AI music models as systems that can start from noise and use prompts and training data to shape the resulting waveform.
Detection is separate from copyright
At the time, major record labels were suing Suno and Udio over alleged copyright misuse. The platforms said they used filters intended to prevent protected works from being reproduced.
That legal question is separate from whether listeners can hear that AI was involved. A track can be difficult to classify by ear while still raising questions about how the model was trained.
The full test is covered by MIT Technology Review.
