Sampling psychological content with LLMs

Recent research has documented output homogeneity in large language model (LLM) generated content, with models converging on similar response distributions despite architectural and training differences (e.g., Guo et al., 2023; 2024; Wenger & Kenett, 2025). We build on this research to explore how LLMs sample psychological content, which has important implications for clinical, organizational, psychometric and other applications like automatic item or scale generation, persona simulations, and human-LLM interactions in therapeutic or selection contexts. Our broad research question is: do LLMs generate homogeneous personality content, and does this homogeneity vary systematically across model characteristics and psychological content?