Classification bias of LLMs in detecting incivility towards female and male politicians in German social media discourse

This study explores potential biases in current language models for detecting different forms of incivility in online discussions. Our work builds on prior research demonstrating that numerous classification approaches disproportionately focus on overt forms of uncivil language, while more subtle forms of incivility remain underrecognized. Such classification bias may lead to the unfair treatment of social groups that are more frequently targeted by implicit forms of incivility in online debates, such as stereotyping and discrimination, including, for instance, female politicians. On a dataset of 24,681 user comments from YouTube and X, we evaluate state-of-the-art language models including BERT, GPT-4, and Meta Llama 3, for their ability to reliably identify different subtypes of incivility directed at female and male politicians, combining comparative performance evaluation and in-depth error analysis. Our results suggest that stereotyping and discriminatory comments are less reliably classified across all models and learning strategies than, for example, vulgar language and insults, indicating a continued risk of bias toward overt forms of incivility in modern language models. As a result, incivility directed at social groups that are more frequently targeted with stereotyping and discrimination may remain underdetected. With this study, we aim to offer valuable implications for the use of current language models in both incivility research and online moderation, supporting the development of more transparent and fairer artificial intelligence grounded in democratic principles.