Rhetorical features from social media predict real-world violent outcomes of manifestos

The delicate balance between free speech and harmful speech could benefit from an unbiased means to identify rhetorical techniques that characterize harmful discourse irrespective of its factual or ideological content. It remains unclear, however, whether any such techniques are empirically related to harm. We show that historically documented violence linked to political manifestos can be predicted on the basis of their use of certain rhetorical techniques. Alienating ‘us-vs-them’ rhetoric was a particularly dominant predictor of both violent outcomes of manifestos and of perceived manipulative intent in online comments. Such rhetoric may be identified at scale using machine-learning models trained on accessible data, permitting the proactive identification of potentially harmful speech.