tuta.io
Toxicity on social media is often assessed using aggregate trends, yet different forms of engagement may shape these patterns in distinct ways. We analyse approximately 100 million Twitter posts collected during the 2018 Brazilian presidential election and the following year to examine how engagement type, automation, and user activity relate to online toxicity. Using a fine-tuned toxicity model, we find that 12% of posts are classified as toxic at a 0.5 threshold (6% at 0.8). Although aggregate toxicity appears relatively stable over time, this masks diverging engagement dynamics. Retweets, which account for about 60% of posts, are consistently less toxic and exhibit a slight downward trend. In contrast, replies are twice as likely to be toxic and increase significantly over time. Replies to toxic parent posts are themselves more likely to be toxic, indicating conversational propagation of hostility. Contrary to the common assumption that automated accounts primarily amplify content through retweeting, we find that higher automation levels are associated with fewer retweets and more replies and original tweets. These findings demonstrate that rising hostility is concentrated in conversational exchanges and partially obscured by large volumes of low-toxicity amplification, challenging simplified assumptions about automation and online toxicity.
Proceedings of the 18th ACM Web Science Conference 2026 (WebSci ’26), pp. 211–220.
[Download preprint here]
[Link to publication]
Cite this paper
@inproceedings{karimi2026toxicity,
author={Karimi, Soroush and Oliveira, Marcos and Pacheco, Diogo},
title={The Hidden Rise of Toxicity: How Retweets Obscure Increasing Hostility in Brazilian Politics},
booktitle={Proceedings of the 18th ACM Web Science Conference 2026},
series={WebSci '26},
publisher={ACM},
pages={211--220},
year={2026},
month=May,
DOI={10.1145/3795766.3799769},
url={https://doi.org/10.1145/3795766.3799769}
}