Roberto Mochetti
Biography Recordings Works Documentation

Recursive Misinterpretation - a musical telephone game (2025)

About the Piece

This piece is a set of instructions on how to interact with AI music generators. It works like a musical telephone game, and the final result is a recursive dialogue between human and machine, unfolding over multiple movements.

The work allows for a firsthand exploration of the concept of post-AI authorship. It also serves as a test of the chosen AI model, as the discrepancy between how the human and the machine interpret musical descriptions through language are essential for the piece to work.

Instructions

This is how to generate your own version of Recursive Misinterpretation:

  1. Create a simple initial prompt. Try to start with the simplest musical ideas you can think of. This prompt will guide your first movement, which you should compose yourself.
  2. Use your original prompt to generate audio through an AI model of your choice. This becomes the second movement.
  3. Listen to the AI’s output. Based on your interpretation of it, write a new prompt that captures its content, mood, or structure. Just like the initial prompt, no technical or musical language needs to be used.
  4. Compose the next movement based on your new prompt. Then, feed that same prompt into the AI to generate the following movement.
  5. Repeate steps 3 and 4, alternating between your compositions and the AI's, each time using a newly generated prompt. Stop when you find that the AI’s output became too boring, predictable, or broken. You can also keep going until you feel the piece is complete. The final movement must be AI-generated.

Variations of the piece are possible. For example, some models support audio input as prompts. You may choose to build your version of Recursive Misinterpretation entirely through audio-to-audio recursion.

Recursive Misinterpretation I, by Roberto Mochetti and Stable Audio

This is the first version of the piece. Since there are no rules regarding the creation of the prompts that describe the AI music, I decided to not think about it too much and describe what I was listening to like I was explaining to a friend. I later realized that my prompts are a little on the technical side (most of my friends are musicians, after all) but that is not necessary and prompts can be as subjective as the human creator desires.

I used Stable Audio’s AudioSparx 2.0 for this because it was the only model available that gave me interesting initial results (which is also subjective, and any AI can be used for the creation of this piece). My interpretation of the prompts was completely executed through Plug Data.

A video of all the movements in sequence can be found here, but you can also listen to the movements and read the respective prompts below:

  1. "pure and unfiltered white noise”
    1. My creation:
      Download audio
    2. Stable Audio:
      Download audio
      Notice that noise was the most prominent sound, but it was clearly filtered. In the background you could hear what it seems to me like strings playing 20th century music counterpoint with fast notes in legato, but instrument and actual pitches were hard to define as the filtered noise was too strong.
  2. "barely filtered white note, obscuring the convoluted atonal counterpoint played by a chamber group in the background, like a radio that is playing mostly static but lets a faint idea of the music come through”
    1. My creation:
      Download audio
    2. Stable Audio:
      Download audio
      I think this creation sounds even less like the prompt. Still, I found it very interesting even though I was disappointed to see it approach tonality so early in the process. I followed up with the process and used a description of this album as my next prompt.
  3. "ethereal synth sound with faint filtered white note. only the pitch C for about 10 seconds, and then a pitch Eb is added. After about 50 seconds, a diminished chord is solved into a brief fifth interval (pitches F and C)”
    1. My creation:
      Download audio
    2. Stable Audio:
      Download audio
      This got progressively uninteresting to me and I already got about 6 minutes of music, so I decided to stop right here.

Variations and Next Steps

These are suggested variations on the piece that I invite anyone to try, and want to try in the future as well: