This study examines the persistence of hate speech on X (formerly Twitter) in Spanish digital media, using a mixed-methods design. The research analyzed 9894 hate messages from 2.1 million comments across five Spanish digital news outlets (20 Minutos, ABC, La Vanguardia, El Mundo, and El País), collected between January 2021 and May 2022, with verification conducted in 2024. The study used machine learning algorithms, ANOVA, and Random Forest modeling to investigate the factors influencing message persistence versus deletion. The findings show that 88% of hate speech messages remained active after three years, challenging the traditional “pyramid of hate” models. Message intensity was not statistically significant in predicting deletion (P=0.512), whereas categorical factors emerged as the primary determinants. Random Forest analysis identified hate type (mean-decreasing Gini coefficient (MDG) =54.97) and media source (MDG =48.69) as the strongest predictors, with engagement levels correlating with message survival (2.07 for active versus 0 for deleted messages). This study shows that content moderation operates through categorical pattern recognition rather than intensity-based responses. This research proposes an “inverted pyramid” model in which algorithmic virality supersedes severity in moderation decisions, thereby supporting autopoietic systems theory. Hate-speech persistence creates operational latency, where undeleted messages function as nodes that are cyclically reactivated during social crises, thereby normalizing dehumanizing narratives. These findings suggest that current moderation strategies do not adequately address the persistence of hate speech and highlight the need for contextual, human-AI collaborative approaches over automated systems.
Said-Hung et al. (2026) studied this question.