Exploring Human-AI Co-Creativity under Human Control: Framing, Reframing, Brainstorming, and Future Challenges

Generative AI has the potential to support human creativity. In our work, we investigate how one or more humans can collaborate with an AI agent to co-create their contributions, while maintaining human control over process and outcomes. In earlier work, we had developed a conversational UI to large language models (LLMs) for software engineering tasks. Ross and colleagues showed that a well-tuned UI could make a back-end LLM behave in a humble, polite, and highly supportive way. We re-used this architecture to explore creativity and co-creativity opportunities, through careful prompt-engineering. After surveying human-human co-creativity strategies, we first experimented with the well-known strategy of framing a problem with a productive representation. Next, we explored the more powerful concept of reframing a problem after the initial frame had been found to be flawed or insufficient in some way. The conversational UI allowed the human to control how the conversation developed, and which aspects of the conversation would be preserved in an analogy-based design. In our third experiment, we moved from specialist methods to the more generally-adopted processes of brainstorming. A human was able to guide the UI+LLM in exercises based on divergent-thinking, convergent-thinking, summarization, and structured organization/re-organization of outcomes. While these initial experiments were successful, we were only able to implement a dialog between one human and one AI. Our next projects will use a specialized environment in which multiple humans can interact with the UI+LLM configuration, with preservation of each human’s identity, thus adding aspects of Mutual Theory of Mind to the co-creative exercises. After that, we hope to revisit multi-agent symbiotic cognitive computing architectures for a richer configuration of multiple humans and multiple AI agents. Throughout this work, we have focused on principles of IBM’s Augmented Human Intelligence, in which AI is used to support and extend the work of humans – not to replace humans. Following a recent debate of Shneiderman and Muller, we label all AI conversational turns with an “AI” or “APP” marker – i.e., we explicitly avoid any so-called Turing test confusions about who or what is speaking or acting. We maintain human control of both process and outcomes. As we showed in a recent CHIWORK paper, these are design choices. It is possible to create interactive AI solutions that channel and control the work of humans. Recent work by many researchers have documented the potential and actual harms of such systems. We make a different choice: We design for AI applications that support, educate, and enable human abilities and human agency.