Adapting the Dangerous Speech Paradigm to Identify Incitement from Polarized and Hate Speech

Online platforms urgently need better ways of managing inciting speech that will not unreasonably limit lawful political speech. To address this, we use discourse on the minimally-moderated platform Gab surrounding the Charlottesville Unite The Right Rally in 2017. We first adapt the Dangerous Speech Paradigm to a three-tiered framework with messages classified as either allowable , harmful , or inciting . The bisection of the previous ”dangerous” category into ”harmful” and ”inciting” enables more targeted classification. Using a customized GPT-4o-mini large language model (LLM) framework that utilizes Chain-of-Thought prompting (CoT) and in-context examples, we classify over 3.5 million posts under the three-tiered framework. To validate the compatibility of our framework with diffusion methods which may be used in future work, we then build a directed repost graph to model the spread of harmful and inciting speech through the use of a DeGroot diffusion model. We find support for LLM classifiers to assist human content moderators in identifying incitement according to our framework. With this assistance, intervention methods can be more effectively targeted, better protecting the public while preserving freedom of speech.