Brainstorming with a Generative Language Model: Effect of Exposure to AI Ideas on Brainstorming Performance and Cognitive Load

Abstract The adoption of generative AI (artificial intelligence) has sparked increased interest in how individual humans can generate ideas and solve problems collaboratively with AI, aiming to achieve superior performance compared to working alone. Based on group effects from all-human brainstorming groups, the paper investigates a human–AI brainstorming setting using OpenAI’s GPT-3.5 as an embedded generative large language model (GLM). In a between–subjects experiment (n = 75) comparing solitary humans and human–GLM dyads, results show that humans do not perform better individually when supported by a GLM. However, collectively, the human–AI dyad achieves superior (or complementary) performance on common brainstorming performance metrics (fluency, flexibility, novelty, and value). Findings are discussed in relation to effort allocation and potential loafing behavior when working with GLMs. The paper advances our understanding human–AI group dynamics, the transferability of group mechanisms (cognitive stimulation, free riding, cognitive inertia) from all-human to human–AI groups, and the discourse on smart loafing with AI group members.