Social bots sour activist sentiment without eroding engagement

TL;DR

The paper addresses how social bots influence online political activism by analyzing Twitter discourse around Extinction Rebellion protests in 2019. It uses a mixed-method pipeline combining Botometer and self-trained models for bot detection, topic modeling, Granger causality to identify information cascades, and Difference-in-Difference models to estimate the causal impact of bot interactions on individual users over a 30-day window. Key findings show that bots exert a stronger influence on human behavior than humans on bots during heated periods; bot interactions consistently reduce subsequent sentiment toward protests, while the impact on tweeting activity depends on bot type (astroturf bots increase activity; other bots decrease it); sentiment changes depend on users’ initial stance, with neutral users most affected, though bots do not alter activists’ engagement. The study highlights that small per-encounter effects accumulate due to the large volume of bot communication, calls for greater platform transparency and data access, and emphasizes policy considerations around automated accounts in online activism.

Abstract

Social media platforms have witnessed a substantial increase in social bot activity, significantly affecting online discourse. Our study explores the dynamic nature of bot engagement related to Extinction Rebellion climate change protests from 18 November 2019 to 10 December 2019. We find that bots exert a greater influence on human behavior than vice versa during heated online periods. To assess the causal impact of human-bot communication, we compared communication histories between human users who directly interacted with bots and matched human users who did not. Our findings demonstrate a consistent negative impact of bot interactions on subsequent human sentiment, with exposed users displaying significantly more negative sentiment than their counterparts. Furthermore, the nature of bot interaction influences human tweeting activity and the sentiment towards protests. Political astroturfing bots increase activity, whereas other bots decrease it. Sentiment changes towards protests depend on the user's original support level, indicating targeted manipulation. However, bot interactions do not change activists' engagement towards protests. Despite the seemingly minor impact of individual bot encounters, the cumulative effect is profound due to the large volume of bot communication. Our findings underscore the importance of unrestricted access to social media data for studying the prevalence and influence of social bots, as with new technological advancements distinguishing between bots and humans becomes nearly impossible.

Figures (4)

• Figure 1: Information flow of humans and bots. The directions and amount of tweets retweeted between bots and humans.
• Figure 2: Panel A: Number of tweets posted by bots (blue line) and humans (red line) in a cascade mutually-driven by bots and humans within the "Anti-XR protests" topic. Panel B: Ratio of bot and human generated tweets throughout the time period of the cascade. The number of tweets is the rolling average aggregated on 5-minute intervals.
• Figure 3: Two examples of Twitter timelines. Top timeline shows daily tweet counts and daily mean sentiment for a human account that directly interacted with a bot on day 0; bottom timeline shows a matched account without a direct bot interaction. Bar heights are proportional to tweet counts, color indicates sentiment from $-1$ to $1$.
• Figure 4: Predicted change resulting from a bot interaction. Panel A, C: Predicted change in Amount (number of tweets). Panel B,D: Predicted change in Sentiment. Panel A-B separate predictions (by color) are shown for astroturf bots (light blue), and other bots (dark blue), while Panel C-D separate predictions by users support level towards XR - supporters (green), neutral users (blue), anti-XR (purple)