The authors used an open Lithuanian organisational-psychology survey with 263 human respondents and 68 items.
Neither decision models nor new LLMs beat statistical baselines as synthetic survey respondents
A psychometric audit of 61 language models on a 68-item Lithuanian survey shows the best generative model is level with a Gaussian copula fit on other respondents and newer model releases are not reliably better.
Academic
Mantas Lukauskas · Viktorija Šarkauskaitė
Hostinger · Kaunas University of Technology · AI Insight Lab · Independent Researcher
Research Digest··3 min read
Lukauskas and Šarkauskaitė tested 14 decision models and 47 generative LLMs as substitutes for human survey respondents.
§