Human Preferences in Large Language Model Latent Space: A Technical Analysis on the Reliability of Synthetic Data in Voting Outcome Prediction

2025-02-22

AI Summary

This study investigates whether synthetic data generated by large language models (LLMs) can reliably substitute for human survey data in vote prediction. We propose a latent-space probing method to disentangle political orientation and systematically evaluate 14 mainstream LLMs across three dimensions: demographic subgroup opinion distributions, personality–partisanship mappings, and prompt robustness. Results reveal that LLMs substantially underestimate inter-individual opinion variance observed in real populations, exhibit poor discriminability in political stance attribution, and produce outputs highly sensitive to prompt phrasing—collectively undermining predictive stability and construct validity relative to empirical survey data. To our knowledge, this is the first cross-model empirical study to identify and characterize systematic opinion distortion mechanisms inherent in LLMs. Our findings establish critical methodological boundaries for public opinion modeling, social science experimentation, and AI-augmented survey research.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Voting
• Natural Language Processing: (Large) Language Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Generative AI (GenAI) is increasingly used in survey contexts to simulate human preferences. While many research endeavors evaluate the quality of synthetic GenAI data by comparing model-generated responses to gold-standard survey results, fundamental questions about the validity and reliability of using LLMs as substitutes for human respondents remain. Our study provides a technical analysis of how demographic attributes and prompt variations influence latent opinion mappings in large language models (LLMs) and evaluates their suitability for survey-based predictions. Using 14 different models, we find that LLM-generated data fails to replicate the variance observed in real-world human responses, particularly across demographic subgroups. In the political space, persona-to-party mappings exhibit limited differentiation, resulting in synthetic data that lacks the nuanced distribution of opinions found in survey data. Moreover, we show that prompt sensitivity can significantly alter outputs for some models, further undermining the stability and predictiveness of LLM-based simulations. As a key contribution, we adapt a probe-based methodology that reveals how LLMs encode political affiliations in their latent space, exposing the systematic distortions introduced by these models. Our findings highlight critical limitations in AI-generated survey data, urging caution in its use for public opinion research, social science experimentation, and computational behavioral modeling.

Problem

• Assessing LLM reliability for human preference simulation.
• Analyzing demographic and prompt effects on LLM opinions.
• Evaluating AI-generated survey data's predictive accuracy.

Innovation

• Probe-based methodology analysis
• Latent space encoding exploration
• Prompt sensitivity assessment