Indirect Causal Influence of a Single Bot on Opinion Dynamics Through a Simple Recommendation Algorithm

Abstract

The ability of social and political bots to influence public opinion is often difficult to estimate. Recent studies found that hyper-partisan accounts often directly interact with already highly polarised users on Twitter and are unlikely to influence the general population’s average opinion. In this study, we suggest that social bots, trolls and zealots may influence people’s views not only via direct interactions (e.g. retweets, at-mentions and likes) but also via indirect causal pathways mediated by platforms’ content recommendation systems. Using a simple agent-based opinion-dynamics simulation, we isolate the effect of a single bot – representing only 1% of the population – on the average opinion of Bayesian agents when we remove all direct connections between the bot and human agents. We compare this experimental condition with an identical baseline condition where such a bot is absent. We used the same random seed in both simulations so that all other conditions remained identical. Results show that, even in the absence of direct connections, the mere presence of the bot is sufficient to shift the average population opinion. Furthermore, we observe that the presence of the bot significantly affects the opinion of almost all agents in the population. Overall, these findings offer a proof of concept that bots and hyperpartisan accounts can influence average population opinions not only by directly interacting with human accounts but also by shifting platforms’ recommendation engines’ internal representations.

Supplementary Materials

Different Engagement Functions

In the main text, we presented results assuming that agents are more likely to engage when the distance between their own opinion and the other agent’s opinion is high. Here we study the sensitivity of our results to other engagement functions. Figure S1 shows the same results as Fig. 2 in the main text. Instead of assuming that agents are more likely to engage when content is dissimilar, we assume that agents are more likely to engage when the observed content is similar. In Fig. S2, we study a bimodal engagement function where agents are more likely to engage with very similar or very dissimilar content and less likely to engage with content that is neither too similar nor too dissimilar.

Fig. S1.

Homophilous engagement function. Agents are more likely to engage with content in their feed that is closer to their own opinions. (a) population’s mean opinion; (b) population’s mean engagement; (c) the number of people influenced by the bot. (d) agents’ mean opinion shift. The analysis shows that the results reported in the main text might be sensitive to the specific engagement function used by the agents to chose which content items they engage with.

Fig. S2.

Bimodal engagement function. Agents’ engagement with content follows a binomial distribution with bimodal probability for content that is close to the target agent’s private opinion as well as content that is distant from the agent’s opinion. Content that falls between these two extremes is less likely to generate engagement. (a–d) population’s mean opinion, population’s mean engagement, number of people influenced by the bot and mean opinion change, as a function of time. Notice that contrary to the main text results (Fig. S2) here no difference emerges between conditions. This is likely due to the bimodal engagement function. Agents using this engagement function were more likely to engage with content that was similar to their own opinion. According to the Bayesian update rule (Eq. 4), this preference for similar content generates escalation dynamics that lead to agents reinforcing their own opinion [23] and thus being deaf to the bot’s different opinion.