Tuesday, July 9, 2024

On Generative AI Art: A Prompter is not an Artist (yet)

I have heard a number of people claiming that they are creative authors of these works of art because they are refining their prompts until they get what they want. However, this fails the litmus test: Isn't refining a prompt still more similar to what a manager or client would traditionally do? They'd tell the artist what to make, the artist makes it, then they tell it what edits they want, etc. until the piece is done to their liking. Therefore a prompter's role is still more like a client who hired an artist, the artist in this case being the AI.

Additionally, I propose a way to objectively measure creative contribution, without any technical/knowledge gatekeeping nor requiring it be done a certain way; a mathematical definition of creative authorship:

The fewer variations of an output that can match your spec, the more you contributed to the creative process.

Consider the 3 scenarios: A classical piece with all notes fully written out; a jazz piece which allows significant improvisation from the soloists; an AI-generated piece where someone prompted “Epic orchestral soundtrack, sci-fi battle, heroic with French horn melody, high-quality”.

1. The composer already restricted the subspace of possible sounds to a small slice, so they’re recognized as contributing to the majority of the creative process. There is still room for interpretation for dynamics, expression etc., and the musicians are recognized for putting their personal touch when performing these pieces, reducing the subspace from that small slice to that unique recording.

2. There is much more leeway for interpretation via improvisation. In any particular recording, the composer and the soloists are all rightfully recognized as having contributed significantly to the work, because the soloists are composing important parts of the piece on the fly.

3. It is such a vaguely specified instruction that there’s an uncountable number of wildly different pieces that could’ve fit the requirements. As such, your role as the prompter was more like a client who hired the AI to do all the creative work for you, rather than an actual artist.

This concept can be applied universally to both music and visual arts. It also leaves the door open in the future for a prompter to legitimately be an artist; they'd just have to be incredibly specific with their instructions and the model would need to be very good at following those instructions accurately.