”Tomorrow morning I will finally be gone”: Assessing and Analyzing Risk Indicators in Violent Lone Actors' Pre-Attack Communications

Sammanfattning

Introduction: This thesis investigates risk indicators in pre-attack communications by different types of violent lone actors, and the extent to which computational methods can identify them. Research Question: Which risk indicators are present in pre-attack communications produced by violent lone actors, how do these indicators vary across different categories of actors, and to what extent can computational methods support automated identification of these indicators? Method: A database of more than 100 writings from 82 violent lone actors was created and analyzed through a mixed-methods approach combining manual annotation by two independent annotators as well as automated analysis using the computational threat assessment tool Dechefr. Actors were categorized along three dimensions: grievance, context and motive, and suicidality, which together make an actor profile. Indicator prevalence was compared across actors using descriptive statistics and chisquare tests. Results: Indicators such as Grievance, and Leakage were present in over 90% of communications of violent lone actors, whereas Linguistic Alignment, Warrior Mentality, and Influences from Previous Offenders varied significantly across categories and were concentrated among ideologically organized actors. Actors expressing explicit suicidality showed the highest prevalence of escalation-linked indicators such as Preparation (95%), Leakage (100%), and Last Resort (85%), despite some of these actors having lower overall indicator counts. Dechefr demonstrated high precision (mean 85% for all indicators) but limited recall (mean 78%). Discussion: These findings suggest that risk indicators serve two partially distinct functions: universal markers of violent intent shared across actor profiles, and markers of suggested actor types. A lower indicator count should not be interpreted as lower risk. Dechefr can support initial screening but cannot substitute for human judgment, particularly for contextual and subcultural language. Further research could explore the longitudinal development of communicative patterns, expand the dataset to underrepresented actor types, and compare different automated methods against each other.