Pitfalls of Conversational LLMs on News Debiasing

TL;DR

This study evaluates the effectiveness of conversational Large Language Models in news debiasing, revealing their limitations in maintaining author style, avoiding misinformation, and matching expert evaluation quality.

Contribution

It introduces a tailored evaluation checklist for news debiasing and systematically assesses popular conversational LLMs' performance in this task.

Findings

LLMs are imperfect in debiasing news articles.

Some models introduce unnecessary changes and misinformation.

Models underperform compared to domain experts in evaluation.

Abstract

This paper addresses debiasing in news editing and evaluates the effectiveness of conversational Large Language Models in this task. We designed an evaluation checklist tailored to news editors' perspectives, obtained generated texts from three popular conversational models using a subset of a publicly available dataset in media bias, and evaluated the texts according to the designed checklist. Furthermore, we examined the models as evaluator for checking the quality of debiased model outputs. Our findings indicate that none of the LLMs are perfect in debiasing. Notably, some models, including ChatGPT, introduced unnecessary changes that may impact the author's style and create misinformation. Lastly, we show that the models do not perform as proficiently as domain experts in evaluating the quality of debiased outputs.