Entangled Mediation: An Empirical Study of Intent Drift in Human–AI Co-Creation with a Text-to-Image Model
TL;DR
An empirical study examining how creative intent evolves during interaction with Dreamina, a text-to-image generative model, indicates that creative intent functions as a processual attribute shaped through iterative generation–evaluation–modification cycles, rather than a stable precondition.
Abstract
This poster presents an empirical study examining how creative intent evolves during interaction with Dreamina, a text-to-image generative model. Moving beyond the assumption that creative intent is fixed prior to action, we propose an entangled mediation framework—drawing on post-phenomenological mediation theory and Frauenberger’s Entangled HCI—to analyze the recursive interplay between human decision-making and model outputs. Through a mixed-methods study with 10 experienced creators using think-aloud protocols, process logging, and retrospective interviews, we identify three interaction patterns: local correction, gradual emergence, and intent reorganization. Our findings indicate that creative intent functions as a processual attribute shaped through iterative generation–evaluation–modification cycles, rather than a stable precondition. We further observe that creators’ perceptions of authorship are tied to their locus of decision-making rather than the proportion of AI-generated content. These findings contribute to understanding human–AI co-creativity as a situated, relational process and offer preliminary implications for designing generative tools that support flexible interaction strategies.