Plausible but Not Valid: A Psychometric Audit of LLMs as Synthetic Survey Respondents

A research study critically examines the use of Large Language Models (LLMs) as synthetic survey respondents, moving beyond mere plausibility to evaluate their psychometric validity. The study posits that LLMs must accurately replicate the joint distribution, latent structure, reliability, mediation pathways, and demographic effects observed in real human survey data. Using a Lithuanian organizational-psychology dataset, a comprehensive lineup of LLMs was tested under various conditions to assess their ability to preserve these psychometric properties.

Why it matters

The increasing reliance on LLMs for data generation, including synthetic survey responses, necessitates rigorous validation to ensure the reliability and integrity of insights derived. If LLMs fail to accurately replicate the psychometric characteristics of human data, strategic decisions based on such synthetic data could be flawed, leading to misallocated resources or ineffective initiatives.

Key insights

• Existing evaluations of LLMs as synthetic survey respondents often focus on individual-level answer plausibility rather than psychometric validity.
• The appropriate evaluative question for LLMs as survey respondents is whether they preserve key psychometric properties of real human data.
• These properties include joint distribution, latent structure, reliability, mediation pathways, and demographic effects.
• A study utilized a Lithuanian organizational-psychology dataset (n=263 employees, 68 items across 12 subscales) for evaluation.
• A diverse lineup of 37 LLMs (OpenAI, Anthropic, Google, and open-weight models) was tested.
• LLM performance was assessed under varying persona-disclosure levels, presentation formats, reasoning-effort ablations, and counterfactual demographic scenarios.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.14606