Counterspeech for Mitigating the Influence of Media Bias: Comparing Human and LLM-Generated Responses (arXiv)

In addition to escalating societal divisiveness, biased news is frequently exacerbated by hurtful reader comments that propagate false narratives. This study demonstrates how these remarks hurt targeted populations and reinforce bias. With the use of a manually annotated dataset that connects bias, offensive comments, and counterspeech, it presents the first counterspeech generating method specifically designed for news items. More than 70% of hostile comments, according to analysis, promote biased content, highlighting the necessity of counterspeech. Model-generated counterspeech lacks uniqueness and diversity, yet it is more courteous than human responses. The generated counterspeech is more diverse and relevant thanks to improvements made through few-shot learning and news context integration.

https://arxiv.org/abs/2508.15855