AI-Induced Idea Diversity Collapse: Ex Ante Evaluation Framework

A new paper on arXiv (2605.06540) introduces a framework for evaluating how creative AI systems may reduce idea diversity across populations, even when individual outputs improve. The authors propose a human-relative benchmarking method that estimates crowding risk from model-only generations and matched unaided human baselines, without requiring human-AI interaction data. Ideas are modeled as congestible resources, yielding an excess-crowding coefficient Δ and a human-relative diversity ratio ρ, where ρ ≥ 1 indicates no excess crowding. The framework connects Δ to an adoption game with exposure-dependent redundancy costs, providing an ex ante protocol to assess AI-induced diversity collapse.