Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

A team from Columbia University led by Yuting Deng, Melanie Brucks, and Olivier Toubia has published groundbreaking research examining why Large Language Models (LLMs) generate less diverse ideas than humans, raising concerns about AI homogenizing innovation at a societal level. Their paper, 'Examining and Addressing Barriers to Diversity in LLM-Generated Ideas,' identifies two key mechanisms: at the individual level, LLMs exhibit fixation where early outputs constrain subsequent ideation, and at the collective level, they aggregate knowledge into a unified distribution rather than exhibiting the knowledge partitioning inherent to human populations where each person occupies distinct regions of the knowledge space.

The researchers conducted four studies demonstrating that targeted prompting interventions can address each mechanism. Chain-of-Thought (CoT) prompting reduces fixation by encouraging structured reasoning (effective only in LLMs, not humans), while using ordinary personas (versus 'creative entrepreneurs' like Steve Jobs) improves knowledge partitioning by serving as diverse sampling cues. When combined, these approaches produce the highest idea diversity, actually outperforming human ideation. This research offers both a theoretical framework for understanding LLM idea diversity and practical strategies for human-AI collaboration that leverage AI's efficiency without compromising the essential diversity needed for healthy innovation ecosystems.