Generalization of the Diversity-Reduction Effect Beyond GPT-4 and Text Tasks

Determine whether the observed reduction in idea diversity caused by human–generative AI collaboration generalizes to generative AI models other than GPT‑4 and to creative tasks beyond text-based modalities, by conducting comparative experiments across models and modalities using consistent diversity metrics and designs.

Background

The meta-analysis finds that human–GenAI collaboration significantly reduces idea diversity (pooled Hedges’ g ≈ −0.86) based on six observations from four studies, though with considerable heterogeneity. These findings suggest a homogenization risk when collaborating with LLMs.

However, the evidence base for diversity effects is limited in size and scope, largely centered on text tasks and GPT‑4. The authors explicitly note that it remains unclear whether the negative diversity effect extends to other models or modalities, motivating targeted generalization studies.