Artificial Intelligence Chatbot to Combat Trolling on Social Media Platforms

• In Event: Research Design

Abstract

As social media grows in popularity, so does the incidence of trolling. The method of countering hate speech with proactive intervention can be used to address online harassment and increase inclusivity of an online platform while respecting freedom of discussion in various contexts, including across partisan lines. For our project, we propose an AI-driven chatbot that can detect instances of online harassment and intervene to spare the target. The chatbot will be built using a state-of-the-art generative language model and trained on varied dialogues to (1) detect a harassing conversation or comment, and (2) generate an appropriate response to address the hateful speech in different contexts. To evaluate the chatbot’s effectiveness, we will design and run experiments that involve deploying the chatbot on several social media platforms. As part of this experimental design, we plan on training the chatbot to engage in different forms of intervention, such as appealing to positive social norms or attempting to shame the harasser, in order to better understand the types of behaviors that most successfully reduce hate speech. Since a primary concern of ours is to respect the freedom of expression on many online platforms, we further plan to evaluate whether our chatbot demonstrates partisan bias in the comments it classifies as harassment.

Authors:
Maya Srikanth(1), Nicholas Adams-Cohen(2), Betty Wang(3), Angie Liu(1), Anima Anandkumar(1), and R. Michael Alvarez(3),

1Department of Computing and Mathematical Sciences, California Institute of Technology
2Immigration Policy Lab, Stanford University
3Division of Humanities and Social Sciences, Caltech Institute of Technology

Authors

• Maya Srikanth, California Institute of Technology

• Nicholas Joseph Adams-Cohen, Stanford University

• Betty Wang, California Institute of Technology

• Anqi Liu, Caltech

• Anima Anandkumar, Caltech

• R. Michael Alvarez, California Institute of Technology