The Content Moderator's Dilemma: Removal of Toxic Content and Distortions to Online Discourse

There is an ongoing debate about how to moderate toxic speech on social media and how content moderation affects online discourse. We propose and validate a methodology for measuring the content-moderation-induced distortions in online discourse using text embeddings from computational linguistics. We test our measure on a representative dataset of 5 million US political Tweets and find that removing toxic Tweets distorts online content. This finding is consistent across different embedding models, toxicity metrics, and samples. Importantly, we demonstrate that content-moderation-induced distortions are not caused by the toxic language. Instead, we show that, as a side effect, content moderation shifts the mean and variance of the embedding space, distorting the topic composition of online content. Finally, we propose an alternative approach to content moderation that uses generative Large Language Models to rephrase toxic Tweets to preserve their salvageable content rather than removing them entirely. We demonstrate that this rephrasing strategy reduces toxicity while minimizing distortions in online content.