Counteract Homogenization with Diversity
Resist statistically-common ideas; deploy diverse specialist personas and, where feasible, multiple models with different training data. A single LLM voice narrows the collective idea space even as it may improve individual idea quality.
- Anderson et al. (2024, PNAS Nexus): LLMs matched individual human originality scores but AUT variability was 0.459 (LLM) vs. 0.699 (human) — structurally convergent at population level
- Meincke/Nave/Terwiesch (Nature Human Behaviour, June 2025): across 5 experiments, 37 of 45 comparisons showed ChatGPT decreases population-level idea diversity; 94% of AI-assisted ideas shared overlapping concepts
- MultiAgent Research Ideator (SIGDIAL 2025, 7,000 ideas empirically): diverse agent personas are the single highest-leverage variable for idea quality, outperforming parallelism or iteration depth
- Always ask "what would be the most surprising version of this?" to reach the low-probability tail of the idea distribution rather than the statistical center
- Prefer genuine multi-model diversity (different providers, different training data) over multi-prompt diversity within one model where feasible