# AI-Augmented Surveys: Leveraging Large Language Models and Surveys for Opinion Prediction

**Authors:** Junsol Kim, Byungkyu Lee  
**arXiv:** [2305.09620](https://arxiv.org/abs/2305.09620)  
**Date:** 2023-05-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.09620v2/

## Key points

- Authors Junsol Kim and Byungkyu Lee explore the intersection of large language models (LLMs) and social surveys to enhance research practices in the social sciences.
- They introduce a novel methodological framework that combines LLMs with survey data to predict individual responses accurately.
- The framework involves personalizing LLMs by analyzing survey questions, inferring latent beliefs, and considering temporal contexts across survey periods.
- Kim and Lee demonstrate the effectiveness of fine-tuning LLMs, specifically the Alpaca-7b model, in predicting individual responses to missing survey questions using data from the General Social Survey spanning from 1972 to 2021.
- The refined model's capabilities enable researchers to fill in missing trends with high confidence and identify pivotal moments when public attitudes shift.
- The authors address practical constraints, concerns regarding socio-demographic representation, and ethical considerations related to individual autonomy and privacy when utilizing LLMs for opinion prediction.
- This study showcases how LLMs and surveys can complement each other's strengths in advancing social science research practices.

## Summary

In their paper titled "AI-Augmented Surveys: Leveraging Large Language Models and Surveys for Opinion Prediction," authors Junsol Kim and Byungkyu Lee explore the intersection of large language models (LLMs) and social surveys to enhance research practices in the social sciences. They introduce a novel methodological framework that combines LLMs with survey data to accurately predict individual responses to survey questions that were previously unasked. This framework involves personalizing LLMs by analyzing the textual meaning of survey questions, inferring individuals' latent beliefs from response patterns, and considering temporal contexts across various survey periods. Through their study utilizing data from the General Social Survey spanning from 1972 to 2021, Kim and Lee demonstrate the effectiveness of fine-tuning LLMs, specifically the Alpaca-7b model, in predicting individual responses to both partially missing and entirely missing survey questions. The refined model's remarkable prediction capabilities enable researchers to fill in missing trends with high confidence and identify pivotal moments when public attitudes shift, such as the increasing support for same-sex marriage. Furthermore, the authors address practical constraints, concerns regarding socio-demographic representation, and ethical considerations related to individual autonomy and privacy when utilizing LLMs for opinion prediction. They emphasize that this study showcases how LLMs and surveys can complement each other's strengths: while LLMs expand survey potential by providing human-like responses, surveys contribute to improving the alignment of LLM predictions with real-world data. Overall, this research highlights the promising synergy between advanced language models and traditional survey methodologies in advancing social science research practices.