AI passed Turing test in music

University of Minas Gerais in Brazil conducted an experiment. Participants were given pairs of songs, in each of which was one generated track. They needed to determine which one exactly. And the results were unexpected.

Researchers collected artificial songs from the open Suno library, from YouTube and from Reddit communities.

In the 5th control pair, the participant heard the introduction to Ludwig van Beethoven’s 5th symphony. Answers of those who didn’t recognize it were not counted in the study. In total 308 participants out of 653 finished. Average age – 31 years.

In pairs matched by genre, the overall share of correct answers was 60%, and in random pairs – 53%, “which is statistically indistinguishable from random guessing”, researchers note. In 35% of cases both songs, including the one written by a human, were considered generated.

Recognition probability increased if a person had musical education or instrument experience. Also the number of correct answers was higher among those who themselves tried AI instruments for music generation. The higher the participant’s age, the lower the share of correct answers.

A paradoxical picture emerges. AI passed the Turing test in music. But for musicians there’s also good news for now. The closer the pair was in style, vocals and sounds, the better listeners distinguished AI from live music.