Effects of Different Human-In-The-Loop Approaches on Human-AI Co-Design: A Comparison Between Human-Learning HITL Approach and Machine-Learning HITL Approach

Abstract The speed at which the quantity of data is increasing stymies the deployment of some AI tools for design as the basis for training the tools is not stable. Also, selecting which data can be used as training data is a challenge. To address these issues, the concept of human-in-the-loop (HITL) approaches have been proposed. HITL includes human knowledge in the model-building processes to mitigate the high volume requirement of training data. Although how to apply AI in a design process and the effect of Generative AI tools (such as ChatGPT and Midjourney) in design processes has been explored, limited attention has been given to how different HITL approaches affect design performance. This study thus aimed to explore how different HITL approaches (Human-learning HITL approach and Machine-learning HITL approach) affect solutions of human-AI co-design. Fifteen participants were recruited to undertake two design tasks under the guidelines of Human-learning HITL and Machine-learning HITL approach respectively. The solutions from participants were assessed using four criteria (novelty, aesthetics, functionality, and feasibility). The results of the study indicated that the solution generated from Human-learning HITL approach is better compared with that of the Machine-learning HITL approach, especially in terms of aesthetics and functionality of the solution. The study supports the importance of expressing human insights through prompts on improving human-AI co-design solutions compared with providing more training data to AI.