Aligning Large Language Models with Diverse Political Viewpoints

TL;DR

This paper presents a method to align large language models with diverse political viewpoints using a large dataset of Swiss parliamentary comments, improving their ability to generate balanced political summaries.

Contribution

The authors introduce a novel alignment procedure using a large dataset of political comments to reduce bias and generate balanced political overviews in LLMs.

Findings

Models aligned with Swiss political data produce more accurate viewpoints.

Aligned models generate more balanced and comprehensive political summaries.

The approach outperforms commercial models like ChatGPT in political viewpoint accuracy.

Abstract

Large language models such as ChatGPT exhibit striking political biases. If users query them about political information, they often take a normative stance. To overcome this, we align LLMs with diverse political viewpoints from 100,000 comments written by candidates running for national parliament in Switzerland. Models aligned with this data can generate more accurate political viewpoints from Swiss parties, compared to commercial models such as ChatGPT. We also propose a procedure to generate balanced overviews summarizing multiple viewpoints using such models. The replication package contains all code and data.