Prompting Diverse Ideas: Increasing AI Idea Variance

Core Framing

Positions idea generation as exploration of a rugged solution landscape where diversity drives success. Unlike routine tasks where consistency matters, innovation requires variance, the best idea emerges from a broad pool, not from optimizing a single direction.

Key Findings

• GPT-4 baseline produces less diverse ideas than aggregated human groups (0.377 vs 0.243 cosine similarity)
• Chain-of-Thought prompting nearly matches human diversity (0.255)
• CoT expands the “idea pond” from ~3,700 to ~4,700 unique ideas
• Different prompting strategies produce different ideas (low between-pool overlap)
• Persona prompts (“Steve Jobs”) outperform elaborate methodologies (HBR brainstorming techniques)
• Idea exhaustion occurs after ~750-800 ideas regardless of strategy

Methodology

• Domain: consumer products for college students under $50
• Measured: cosine similarity (Google Universal Sentence Encoder), unique idea count, exhaustion rate
• Compared 35 prompting strategies across 10 sessions each
• Threshold of 0.8 cosine similarity for “identical” ideas

Extracted Content

→ 05-atom—ai-diversity-deficit → 05-atom—idea-space-as-landscape → 05-atom—cot-diversity-effect → 05-atom—idea-exhaustion-dynamics → 03-atom—cosine-similarity-threshold → 05-molecule—prompt-engineering-for-diversity