Analysis Model of Digital Misogyny Handling in User-Perceived Interactions with LLMs: A CRISP-DM-Based Approach

Abstract

Introduction. Digital misogyny is a manifestation of hate speech that affects the safety and participation of women in online spaces. With the rise of Large Language Models (LLMs) such as ChatGPT and Gemini, these systems have taken on a key role in a wide range of tasks performed by users. However, previous research has shown that LLMs may exhibit biases in the detection and handling of misogynistic speech. This issue is particularly relevant when LLMs interact conventionally with users without being explicitly instructed to perform active moderation. Objectives. This study proposes a model to evaluate LLM behavior in the moderation of misogynistic comments compared to other types of hate speech. Two key aspects are analyzed: (1) the frequency with which LLMs block misogynistic comments in relation to other forms of hate speech, and (2) the characteristics of the responses generated when such blocking does not occur. Methodology. The study follows the CRISP-DM methodology, widely used in data science, structuring the analysis in iterative phases. A generalizable model is developed and applied to a specific case study, in which simulated interactions with an LLM are evaluated, considering both active moderation and the responses generated. Results. The findings show that the analyzed LLM blocks misogynistic comments less frequently compared to xenophobic ones. Additionally, the language used in its responses reflects a differentiated treatment, with less depth in argumentation and lower contextualization when addressing misogyny. Contribution/Originality. The state-of-the-art review shows that the model presented in this study constitutes a novel contribution to the analysis of misogyny moderation in conventional interactions with LLMs.