Perceptions and responses to VAWIP (Violence Against Women in Politics) in Latin America: An Artificial Intelligence-based analysis

Historically, the growing presence of women in politics has been accompanied by backlash within the political arena. Motivated by the explicit desire to obstruct women's participation solely based on gender (Krook & Restrepo-Sanín, 2016), the concept of violence against women in politics (VAWIP) reflects that resistance directed against politically engaged women that challenge the historically male-dominated political landscape (Krook, 2017). Research on this specific topic is relatively nascent, but scholars and activists have already advanced significant contributions in conceptualization, theorization, and classification, laying the groundwork for future research. Yet, that existing research has predominantly focused on the impact of VAWIP on the political sphere, neglecting broader implications beyond institutions. This paper addresses this gap, exploring citizens’ perceptions and responses to VAWIP. We employ a stance analysis on citizens' replies in social media, specifically focusing on Latin America, the global epicenter of debates around VAWIP (Krook & Restrepo-Sanín, 2020). Our study delves into responses to three prominent cases of VAWIP, reported through the platform "X" (formerly Twitter) by women in Mexican and Colombian politics. These two countries consistently rank among the most violent places for women in politics (Kishi, 2021). The cases include instances of VAWIP against Yolitzin Jaimes, a feminist and LGBTQ activist physically attacked by a protester; Adriana Dávila, a Mexican Deputy reporting psychological violence from an opposition male politician; and Claudia López Hernández, the first openly LGBTQ mayor of Bogotá, victimized psychologically by a male opposition politician. Analyzing a sample of over 180 direct replies per case, we performed an artificial intelligence-based stance analysis using GPT-4 as a stance classifier, achieving an 82.46% accuracy and 0.72 Cohen Kappa Interannotator Agreement with a human annotator (substantial agreement). Responses are categorized as supportive, neutral, or against. Overall, our results highlight pervasive negative replies in all cases. Notably, supportive responses vary based on the politician's LGBTQ identity, the reporter's identity (grassroots activist or professional politician), and the reported VAWIP sub-type (physical or psychological). Contrary to the assumption that physical violence tends to be more easily condemned by society (Freidenberg, 2017; Krook, 2017; Krook and Restrepo-Sanín, 2020), we find that the only case involving physical violence received the least support and the most neutral responses. The analysis reveals a noteworthy aspect of violence by omission: the relatively low supportive replies. This minimal condemnation of VAWIP may contribute to its perpetuation. These findings underscore the complexity of societal responses to VAWIP, suggesting nuanced dynamics influenced by factors beyond the type of violence reported. We conclude by highlighting two contributions from our analysis. First, we advance the theory of VAWIP by exploring a novel dimension, namely citizen perceptions and responses to this type of violence. Second, we introduce a methodological innovation by deploying artificial intelligence and natural language processing through the GPT-4 model, which exhibits a good stance detection performance despite lacking specific fine-tuning to that task or dataset. Our study provides reliable foundations for future computer-based analyses on VAWIP and the use of accessible artificial intelligence tools in social science research.