Abstract
This thesis explores artist–Generative AI (GenAI) co-performance to examine how agency and authorship are reconfigured within Human-AI Interaction. Framed through Agential Realism and Speech Act Theory, the research moves beyond the binary distinction of AI-as tool/AI-as-partner. Instead, it analyses GenAI as a performative system whose outputs lack intent yet exert perlocutionary force, fundamentally restructuring human creative trajectories and decision-making processes. Methodologically, the inquiry employs a mixed-method ethnographic approach that integrates Design Anthropology, Entangled Ethnographies, and Diffractive Methodology. This study draws on my interventionist positionality as both researcher and curator to conduct longitudinal observations across three case studies involving six internationally renowned artists. The analysis shows how these practitioners deliberately repurpose and re-stage generative systems, articulating an Aesthetics of Errors that treats algorithmic hallucinations and glitches as diagnostic features, rendering visible the opaque datasets, infrastructures, and labour conditions underlying the model. The research advances three contributions to the HCI community. First, it proposes AI Art Ethnography as a field of study that positions researchers as problem framers. They move across studio practice, exhibition contexts, and academic inquiry. Their work unfolds between the artists, as problem makers, and the AI developers, as problem solvers, to investigate transformative practices and epistemic troubles arising from artists-GenAI co-performance. Second, it develops three performative metaphors that articulate the interaction: the Theatre of Projections frames the stage of the collaboration; the Choreographer ⇄ Performer continuum maps the shifting distribution of agency; the Smark attitude describes the human state of mind that oscillates between suspension of disbelief and critical awareness of the system’s inner working. Third, it conceptualises Perlocutionary Agency and Entangled Authorship to provide a vocabulary for describing co-performance without attributing intentionality to machines, thus maintaining human accountability. Overall, it outlines implications for responsible GenAI development grounded in accountability, ownership, and transparency. By situating AI Art as a critical site for HCI inquiry, the dissertation argues how responsible AI design must be shaped through practices that render prompting, datasets, labour, and curatorial decisions legible, fostering a more transparent and reflexive form of calibrated reliance in human–AI collaboration.