Large Language Models (LLMs) as Agents for Augmented Democracy

A guide to the publication

The overview

This research paper tests six adapted language models using policy choices from 267 volunteers in an online project during Brazil’s 2022 election. The authors report that the models predicted choices better than a simple rule based on political orientation, and that adding model predictions improved estimates of the volunteers’ combined preferences in some tests; accuracy varied across groups, and reversing the order of options sometimes changed predictions. The study does not show how people across Brazil would vote or whether AI helpers would work safely in real decisions, which the authors say these models are not ready to make.