Teaching an AI to compose: how this listening experiment began

By Yuval Shay-El

Can you teach an AI program to compose music by treating it like a human student?

That question was the starting point for this project. I began working with ChatGPT as a composition teacher might work with a student: introducing a musical language, setting exercises, discussing the results, and asking for corrections and revisions.

The aim was to compose music in standard notation, intended for a human musician to perform. The recordings let us hear realizations of those scores.

A shared musical starting point

The framework comes from the experimental dodecaphonic approach developed by the composer and music theorist Zvi Harry Nadel (1908–1987). It should not be confused with Schoenberg’s twelve-tone method, although there are aesthetic affinities.

Nadel’s approach also forms part of my own history as a composer. The human pieces in the listening experiment were my first compositions, written while I was studying with Nadel. They were recently brought together as Five Dodecaphonic Miniatures (1983–1984) for piano. In this project, the roles are reversed: as ChatGPT’s teacher, I have guided it through the same process of exercises, corrections and feedback that I experienced as Nadel’s student.

This gives the human and AI compositions a common starting point. The question becomes what each composer does within that musical language—and whether a listener can hear a difference.

From following rules to making music

Our work proceeded through teaching material, exercises, corrections and feedback. Understanding explicit rules was one part of the process. The more interesting question was whether the results could acquire a sense of musical style, including choices that are difficult to capture in a list of rules.

At this stage, I felt I had reached the first result that clearly went beyond rule-obedience. ChatGPT had begun to absorb not only the rules, but also the style and the “soft rules” of composition, and to use real creative imagination.

I had to encourage it explicitly to move beyond the “rules cage,” much as I would with a human composition student. Surprisingly, that worked.

That was my response as the teacher and listener. The listening experiment opens the question to other ears.

Exercise 21 — Three New Studies

This early example brings together three short AI-composed musical phrases, heard one after another as a continuous piece. The video, titled Exercise 21 — Three New Studies, was made directly with MuseScore Studio and ends with the caption “Made with MUSESCORE STUDIO.” It documents a stage in the teaching process that led to the listening experiment.

If you have not yet tried the listening game, you may prefer to do so before hearing this identified AI example.

From teaching experiment to listening experiment

The game brings together ten piano miniatures and asks a simple question: who composed each one—a human or an AI?

All were composed in standard notation, intended for a human pianist. All ten were prepared in Steinberg Dorico 5 and rendered using the same SONiVOX Eighty Eight Ensemble virtual piano, rather than recorded in live performance. MusicXML served as the editable score-interchange format during the AI composition workflow, including transfers between notation programs.

You can listen, pause, seek and replay, give your judgment and confidence, and then discover the authorship after finishing your blind session. No formal musical training is needed. I am interested in the responses of attentive listeners as well as musicians.

Further listening and reading

• Musical Alphabet Evolution and Contemporary Composition Techniques: The Theory of Zvi Harry Nadel—my article on the musical ideas behind the project.
• Six Short Pieces for Piano Solo (1987), performed by Hagai Yodan—a later development of my own approach, following the 1983–1984 miniatures used in the experiment.
• The original Facebook post and video, where I described this stage of the teaching experiment.