Using Large Language Models for Idea Generation in Innovation

This paper investigates whether Large Language Models (LLMs), specifically GPT-4, can effectively generate high-quality product ideas compared to human students. The authors frame innovation as a search process where the goal is to identify the best possible ideas rather than maximizing the average quality of all generated concepts.

The researchers compared three pools of product ideas for items priced under $50 targeted at college students: one pool generated by university students, one by GPT-4 using zero-shot prompting, and one by GPT-4 using few-shot prompting. They evaluated the quality of these ideas using market research techniques to predict purchase intent. Additionally, they used text mining to measure pairwise similarity and human raters to assess the novelty of the ideas.

The study reveals that AI-generated ideas outperform human-generated ideas in terms of average purchase intent, with few-shot prompting providing a slight boost over zero-shot. While AI-generated ideas are perceived as less novel and exhibit higher similarity (indicating a less diverse range of solutions), they demonstrate a significant advantage at the high end of the quality distribution. Specifically, AI-generated ideas are seven times more likely to rank in the top 10% of all ideas, suggesting that AI is a powerful tool for finding exceptional innovation opportunities.

In innovation, the overall success of a project is often determined by the quality of the best ideas rather than the average quality of the entire pool. By demonstrating that AI can consistently produce top-tier ideas, this research suggests that organizations can leverage LLMs to significantly enhance their innovation pipelines, even if the AI's output is less diverse than that of human teams.