# Language Models Trained on Media Diets Can Predict Public Opinion

**Authors:** Eric Chu, Jacob Andreas, Stephen Ansolabehere, Deb Roy  
**arXiv:** [2303.16779](https://arxiv.org/abs/2303.16779)  
**Date:** 2023-03-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.16779v1/

## Key points

- The study proposes a new approach to measuring public opinion by using language models adapted to online news, TV broadcast, or radio show content that can emulate the opinions of subpopulations that have consumed a set of media.
- The researchers used U.S. nationally representative surveys on COVID-19 and consumer confidence as ground truth for opinions expressed to validate this method.
- Results indicate that this approach is predictive of human judgments found in survey response distributions and robust to phrasing and channels of media exposure.
- This method is more accurate at modeling people who follow media more closely and aligned with literature on which types of opinions are affected by media consumption.
- Probing language models provides a powerful new method for investigating media effects, supplementing polls, and forecasting public opinion.
- Neural language models can predict human responses with surprising fidelity, suggesting a need for further research in this area.
- The authors of the study are Eric Chu, Jacob Andreas, Stephen Ansolabehere, and Deb Roy.

## Summary

The study titled "Language Models Trained on Media Diets Can Predict Public Opinion" introduces a novel approach to measuring public opinion by probing media diet models. The authors propose using language models adapted to online news, TV broadcast, or radio show content that can emulate the opinions of subpopulations that have consumed a set of media. To validate this method, the researchers used U.S. nationally representative surveys on COVID-19 and consumer confidence as ground truth for opinions expressed. The results indicate that this approach is predictive of human judgments found in survey response distributions and robust to phrasing and channels of media exposure. Additionally, it is more accurate at modeling people who follow media more closely and aligned with literature on which types of opinions are affected by media consumption. The study suggests that probing language models provides a powerful new method for investigating media effects, supplementing polls, and forecasting public opinion. Furthermore, it highlights the surprising fidelity with which neural language models can predict human responses and suggests a need for further research in this area. The authors of the study are Eric Chu, Jacob Andreas, Stephen Ansolabehere, and Deb Roy. Their findings have significant implications for understanding how public opinion is shaped by media consumption and how it can be measured more accurately in the future.