Introduction

We launched Bot Sentinel to help diminish the effectiveness of trollbot accounts that infest Twitter. We believe Twitter users should be able to engage in healthy online discourse without foreign countries and organized groups manipulating the conversation.

We designed Bot Sentinel to be easy to use and also to be as informative as possible. We publicly display the Twitter accounts the system is tracking to give visitors of this website a better understanding of how trollbot accounts spread disinformation, and to be as transparent as possible. Our system updates the accounts listed on this website daily.

Bot Sentinel is a non-partisan platform; we track both left and right trollbot accounts. The system uses machine learning and artificial intelligence to classify Twitter accounts who exhibit trollbot behavior, and we then add accounts that exceed a certain threshold to a publicly available database that anyone can browse.

We trained Bot Sentinel to identify specific types of trollbot accounts using thousands of accounts and millions of tweets for our machine learning model. The system can correctly identify trollbot accounts with an accuracy of 95%. Unlike other machine learning tools designed to detect “bots,” we are focusing on specific activities deemed inappropriate by Twitter rules. We analyze hundreds of Tweets per each Twitter account to determine if an account exhibit irregular tweet activity, engaging in harassment, or troll-like behavior

Machine Learning Model

Researchers rarely agree on what someone considers a troll or what constitutes harmful bot activity, so we took a different approach when training our machine learning model. Instead of creating a model based on our interpretation of a troll or bot, we used Twitter rules as a guide when selecting Twitter accounts to train our model. We searched for accounts that were repeatedly violating Twitter rules and we trained our model to identify accounts similar to the accounts we identified as “trollbots.” Note: Ideology, political affiliation, religious beliefs, geographic location, or frequency of tweets are not factors when determining the classification of a Twitter account.

Trollbot Score and Rating

We use the term trollbot because of their troll-like behavior with the repetitive bot-like nature of their trolling. We rate accounts based on a score from 0% to 100%, the higher the score the more likely the account is a trollbot. We analyze several hundred tweets per account, and the more someone engages in behavior that is troll-like, the higher their trollbot rating is. We feel since trollbot accounts are likely violating Twitter rules, most Twitter users would want to report and avoid these accounts because they offer little value to meaningful public discourse. 

Foreign Influence 

Most trollbots are not part of a large conspiracy attempting to influence American policies and/or elections. However, there are trollbots who engage in deceptive tactics and there is a correlation between troll-like behavior and trollbots who are part of an influence campaign. A trollbot that is actively trying to cause division and discord will behave in a manner consistent with someone who receives a high trollbot score.