Abstract
While Artificial Intelligence (AI) research aims to emulate human capabilities through computing technologies, human-centred AI takes into account human values beyond algorithmic epistemic values such as accuracy or computational cost. This thesis particularly focuses on one of these values: creativity. It has long been established as a quality of interaction that has huge downstream societal implications, yet remains elusive and difficult to evaluate. The emerging capabilities of generative AI models further complicate the matter. It is now possible for AI to directly generate semantic or visual idea in a range of creative tasks. While this seems promising, it raises questions about the ownership of creativity and introduces new challenges stemming from the vagueness and uncertainty inherent in conversational, natural language based interactions. On top of that, the increasingly anthropomorphized AI, both in public discourse and research literature, might implicitly influence interaction or collaboration patterns, whose impact on both creativity processes and outcome remains unknown.The thesis first aims to understand human-AI collaboration patterns and dynamics during a co-creativity process. An empirical study was conducted to investigate human-AI co-creativity in prewriting tasks. Drawing upon creativity theories and CSCW perspectives, the study unravels the GPT-3's anthropomorphized, cognitive role in the co-creativity process, its contributions to the final creative outcome, and the user experience of the collaboration workflow and strategies. The findings reveal a three-stage, iterative co-creativity process that includes Ideation, Illumination, and Implementation. This collaborative process champions the human in a dominant role, in addition to mixed and shifting levels of initiative that exist between humans and Large Language Models (LLMs).
The thesis then reports on two creativity support tools. Polymind explores a novel interaction paradigm to support human-AI co-creativity. The system simulates human collaboration scenarios and uses multiple LLM agents to facilitate creative visual diagramming. It operates in a mixed initiative manner that proactively provides generated suggestions in parallel to users' own diagramming process.
Metamorpheus probes the potential of metaphorical visual storytelling as a co-creative way of journaling emotional dream experiences to prompt meaningful self-reflection. The phenomenological study of Metamorpheus shows that the system promotes affective mindfulness, and creates a sense of connectedness --- including the engagement between participants and researchers, and motivations to share dream experiences. These findings highlight the potential of creative storytelling as a way of journaling personal information to promote empathy and self-understanding.
The thesis concludes with a summary of key insights of all the work reported herein, and discusses future opportunities for research in human-AI co-creativity.
| Date of Award | 22 Jul 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Yuhan LUO (Supervisor) & Zhicong LU (External Co-Supervisor) |