Fig. 1

The indirect influence of bots on social information networks. (a) Representation of opinion dynamics network mediated by a content recommender system (grey box on top). Bot and Human agents (circles) consume and share content. A bot agent can influence human opinions via direct interaction with human agents (e.g., retweets, at-mentions, likes, and comments) or indirectly via affecting the internal representation of the content recommendation algorithm. (b) Schematic representation of the effect of bot presence on the internal representation learned by a simple recommender system trained to predict a user’s engagement with various types of content. Including the bot behavior in the training set skews the model to think that engagement with extreme content is more likely than it would be without the bot presence. (c) Agents in the simulation were modeled to include a true private opinion and an expressed public opinion. Agents were presented with one of their neighbors' public opinions on every round based on the recommender’s predicted content engagement. Then the agents decided whether to engage with this content or not according to their engagement function (Eq. 3). Opinion change took place only if the agent decided to engage with the recommended piece of content