Large Language Models as Information Sources: Distinctive Characteristics and Types of Low-Quality Information

TL;DR

This paper characterizes the types and distinctive features of low-quality information generated by large language models, highlighting their differences from traditional sources and implications for information quality and harm mitigation.

Contribution

It introduces a typology of LLM-generated low-quality information and identifies key characteristics that distinguish these outputs from human-generated content.

Findings

Low-quality information includes misprioritization and exaggeration.

LLMs exhibit unique affordances that differ from previous technologies.

The study provides a framework for understanding and addressing LLM-related informational harms.

Abstract

Recent advances in large language models (LLMs) have brought public and scholarly attention to their potential in generating low-quality information. While widely acknowledged as a risk, low-quality information remains a vaguely defined concept, and little is known about how it manifests in LLM outputs or how these outputs differ from those of traditional information sources. In this study, we focus on two key questions: What types of low-quality information are produced by LLMs, and what makes them distinct than human-generated counterparts? We conducted focus groups with public health professionals and individuals with lived experience in three critical health contexts (vaccines, opioid use disorder, and intimate partner violence) where high-quality information is essential and misinformation, bias, and insensitivity are prevalent concerns. We identified a typology of LLM-generated low-quality information and a set of distinctive LLM characteristics compared to traditional information sources. Our findings show that low-quality information extends beyond factual inaccuracies into types such as misprioritization and exaggeration, and that LLM affordances fundamentally differs from previous technologies. This work offers typologies on LLM distinctive characteristics and low-quality information types as a starting point for future efforts to understand LLM-generated low-quality information and mitigate related informational harms. We call for conceptual and methodological discussions of information quality to move beyond truthfulness, in order to address the affordances of emerging technologies and the evolving dynamics of information behaviors.