Creativity and AI

Abstract

We investigate whether AI systems outperform humans in creative tasks that vary in their degree of “openness.” To this end, we generated creative responses using three versions of ChatGPT and recruited 738 participants to blindly evaluate six creative answers randomly drawn from three pools—comprising 160 responses each—generated by both humans and AI. This process yielded 4,428 individual evaluations. Our results show that, regardless of the GPT version employed, humangenerated responses achieve significantly and substantially higher average scores than machinegenerated responses in open tasks. Conversely, AI-generated responses outperform human ones in closed tasks. Furthermore, we estimate that human imagination accounts for between 22% and 45% of the creative score in open tasks.

Keywords: creativity, Artificial Intelligence, experiment, text analysis, structural estimation

JEL Classification: C91, O33, O39