Spherical analysis of communities and AI counternarratives against online abusive content

Online offensive speech represents a persistent social problem, causing discrimination and exclusion against marginalized communities. In Italy, Roma communities are frequently targeted on social media, highlighting the need for interventions that go beyond content removal. Counternarratives have emerged as an effective strategy to address this issue, offering alternative stories and perspectives that confront harmful discourses while amplifying the voices of those directly affected. By positioning themselves in relation to dominant “master narratives,” counternarratives act as mechanisms of counterforce, shaping meaning through intentional and contextually grounded discourses. In this study, we focus on offensive content targeting Italian Roma communities and the generation of counternarratives as a response. We constructed an annotated dataset of 450 social media comments written in Italian containing anti-Roma toxic speech. Three annotators from Italian Roma communities actively generated counternarratives, producing a total of 150 responses each. We compared these human-generated texts with those produced by different Large Language Models. From a methodological perspective, we use a statistical framework for the analysis of textual data based on hyperspherical geometry. Counternarratives are embedded into a high dimensional semantic space and normalized onto the unit hypersphere. The resulting ensemble of semantic orientations is modelled using a finite mixture of von Mises–Fisher distributions. This allows the semantic profile of both human and machine-generated responses to be characterized in terms of dominant orientations and withincluster dispersion. In particular, differences in the estimated concentration parameters offer a formal criterion to evaluate whether AI-generated texts reproduce the intrinsic heterogeneity of community-based discourse or instead exhibit higher concentration, indicative of more homogeneous and potentially formulaic semantic patterns when compared to the lexical and conceptual diversity characterizing human counternarratives.