The limits of artificial intelligence
A study by the Hanover University of Music, Drama and Media shows: Artificial intelligence is clearly inferior to human solutions when it comes to melody continuation tasks.
According to the study, the results of two AI systems are rated significantly worse aesthetically than human solutions. The research group led by Reinhard Kopiez is critical of the current public debate about the musical and creative potential of AI and sees a tendency to overestimate it.
The Hanover-based music psychologists selected the opening bars of a melody from a largely unknown piece of music in the style of film music as the stimulus. Based on a melody continuation paradigm, a total of 111 compositions were created using ChatGPT and Google Magenta Studio. Music students generated a total of 57 continuation variants.
In a randomized blind test, 71 participants with above-average musical experience rated the aesthetic qualities of the melodies. The results astonished the music psychologists: The human solutions were rated significantly better than the AI versions on all scales, with the older Google Magenta Studio system still lagging behind ChatGPT.
Original article:
Schreiber, A., Sander, K., Kopiez, R., & Thöne, R. (2024). The creative performance of the AI agents ChatGPT and Google Magenta compared to human-based solutions in a standardized melody continuation task. Yearbook of Music Psychology, 32, Article e195. https://doi.org/10.5964/jbdgm.195
