The final speaker in this session at the AoIR 2025 conference is the brilliant Fabio Giglietto, presenting a study of pro-Bolsonaro narratives on Facebook in Brazil. The key question here is whether online hyperpartisan groups are as stable as they are thought to be; is that true, and how does such stability fare in times of intense political crisis?
Brazil is an obvious case for the study of such questions. The project tracked some 59 pro-Bolsonaro accounts between 2021 and 2023, a timeframe including Bolsonaro’s election loss against Lula and his subsequent coup attempt. The dataset contains some 12 million posts.
The approach here is to process these posts and calculate several affective engagement metrics, whose volatility over time was then also assessed; actors were also classified. The key indices used here were the Emotional Polarisation Index (EPI), which measures the balance between supportive (love) and hostile (angry) emotional reactions to posts; and the Engagement Balance Index (EBI), which measures the balance between shares and comments on a given post.
From key volatile periods over these years, some 700k posts were identified; a subset of 2,300 posts from these timeframes were coded manually, and six categories of political actors were identified (Bolsonaro and allies, Lula and allies, the Supreme Court, etc.). This coding scheme was then translated into an LLM prompt and refined, eventually reaching strong alignment between LLM and human coding. The LLM was then also able to pick up fairly obscure references to such political actors.
The analysis of EPI and EBI patterns over time in response to such posts then shows some surprising changes: the coup attempt in 2022 in particular leads to a considerable change in patterns. Strong positive reactions to pro-Bolsonaro posts decline considerably towards the end of the period, in fact, and shares of such posts also decline in favour of more critical commenting. Conversely, positive reactions to posts about Lula in these pro-Bolsonaro spaces increase, and shares of such posts also increase.
Partisan communities are therefore far from static; the ‘echo chamber’ metaphor fails to capture their evolution. Major political events serve as inflection points for their development, and shifts in activity can signal community stress or restructuring. Further, LLM classification of such posts clearly can provide critical assistance to the analysis of large-scale social media data.
