This study examines how the expertise assigned to a generative large language model influences early-stage design ideation. Twenty students majoring in mechanically related fields generated ideas for bicycles that make commuting to school more comfortable. Each student wrote an initial idea of 50–150 characters, received feedback from ChatGPT-4o instructed as an expert in Mechanical Engineering, Product Design, Environmental Studies, or with no expertise specification, and then revised or added ideas. All ideas were rated on seven criteria: expertise level, diversity, specificity, logical consistency, novelty, usefulness, and feasibility. Expertise increased under every AI condition. Diversity rose in both Mechanical Engineering and Environmental Studies settings, indicating that viewpoint expansion is not limited to distant domains. Novelty decreased when the AI was set to Mechanical Engineering or Environmental Studies, suggesting that distant expertise does not automatically boost originality. Usefulness improved except in the Mechanical Engineering setting, while feasibility improved in every setting except Product Design. These findings imply that designers can tune different quality dimensions of their ideas by selecting proximate, distant, or non-specified expertise for the AI. Future work will analyze AI feedback content and measure user uptake to clarify the mechanisms behind these effects.
Shimada et al. (Wed,) studied this question.