Identifying the sources of ideological bias in GPT models through linguistic variation in output

TL;DR

This paper investigates ideological biases in GPT models by analyzing linguistic variations across different languages, revealing biases linked to training data and filtering policies, and emphasizing the need for curated datasets.

Contribution

It introduces a novel method to detect ideological bias in GPT models through linguistic analysis across languages with contrasting political attitudes.

Findings

GPT responses are more conservative in Polish and more liberal in Swedish.

Bias differences persist from GPT-3.5 to GPT-4 despite filtering policies.

Training data quality is crucial to reduce ideological bias.

Abstract

Extant work shows that generative AI models such as GPT-3.5 and 4 perpetuate social stereotypes and biases. One concerning but less explored source of bias is ideology. Do GPT models take ideological stances on politically sensitive topics? In this article, we provide an original approach to identifying ideological bias in generative models, showing that bias can stem from both the training data and the filtering algorithm. We leverage linguistic variation in countries with contrasting political attitudes to evaluate bias in average GPT responses to sensitive political topics in those languages. First, we find that GPT output is more conservative in languages that map well onto conservative societies (i.e., Polish), and more liberal in languages used uniquely in liberal societies (i.e., Swedish). This result provides strong evidence of training data bias in GPT models. Second, differences across languages observed in GPT-3.5 persist in GPT-4, even though GPT-4 is significantly more liberal due to OpenAI's filtering policy. Our main takeaway is that generative model training must focus on high-quality, curated datasets to reduce bias, even if it entails a compromise in training data size. Filtering responses after training only introduces new biases and does not remove the underlying training biases.