Large language model accurately predicts online chat derailments

Online chat rooms and social networking platforms frequently experience harmful behavior as discussions drift from their intended topics toward personal conflict. Traditional predictive models typically depend on platform-specific data, limiting their applicability and increasing implementation costs.

In a new study, researchers at the University of Tsukuba applied a zero-shot prediction method to LLMs to detect conversational derailments. The performance of various untrained LLMs was compared to that of a deep learning model trained on curated datasets. The results showed that untrained LLMs achieved comparable, and in some cases superior, accuracy.

These findings, published in the journal IEEE Access, suggest that platform operators can implement effective moderation tools at reduced cost by leveraging general-purpose LLMs, supporting healthier online communities across diverse platforms.