Human-AI Co-Creation in Product Ideation: the Dual View of Quality and Diversity

Generative AI models, such as GPT and DALL-E, have emerged as some of the most impressive technology innovations with profound impact on many businesses. Their abilities to generate content with not only high fidelity but also signs of creativity open new possibilities of human-AI collaboration in creative tasks. In this paper, we conduct two online experiments to understand the relative strengths of humans and GPT-4 in creative product ideation, and compare the merits of different co-creation modes. Importantly, we take a dual view to evaluate product ideas, measuring both their quality and diversity using human evaluations and deep-learning-based assessments. We find clear complementarity between humans and GPT: human-generated ideas exhibit high diversity but low quality, whereas GPT-generated ideas have high quality but low diversity. Such complementarity also persists when humans or GPT engage in idea revisions. Overall, a co-creation mode where humans propose the initial ideas, followed by a GPT-assisted revision process, can achieve greater balance between idea quality and diversity (than relying on humans or GPT alone). Additionally, we explore the underlying mechanisms behind the diversity limitations of GPT-generated ideas, as well as the quality-diversity trade-off in co-creation and how to mitigate it through prompt engineering. Our research has direct implications for individuals and firms in the creative industries: humans and AI bring complementary values to the creative process, and the performance of co-creation depends jointly on the creativity of humans and their ability to effectively use AI to improve their creations.