Conditioned Stochasticity: Maintaining Compositional Agency with Text-to-Audio Generative AI

Source of Publication

Proceedings of the 31st International Symposium on Electronic/Emerging Art (ISEA2026)

Conference Year

2026

Abstract

Commercial generative AI music platforms are displacing musical agency by reducing artists to prompt engineers. This contrasts sharply with earlier generative and algorithmic music practices where artists maintained compositional control. This paper proposes conditioned stochasticity as a framework for maintaining meaningful artistic agency while composing with GenAI systems. This grounds the text-to-audio generative process in experimental music’s long engagement with chance, indeterminacy, and algorithmic procedures. Through an original composition using a latent diffusion model fine-tuned on the first author’s own compositions and recordings, this paper demonstrates how artists can subvert GenAI outputs through deliberate intervention. Rather than accepting the role of prompt engineer within commercial platforms, this approach gives the artist more creative control at different stages: curating training data, directing generation, and composing with outputs. Instead of framing the use of GenAI as a novel approach, this paper situates it as an extension of contemporary experimental music practices that are technologically mediated. The goal here is to offer pathways for artistic agency that neither reject GenAI nor surrender authorship to it.

Document Type

Conference Proceeding

Publication Date

11-6-2026

This document is currently not available here.

Share

COinS