💬 NLP

Language Predicts Identity Fusion Across Cultures and Reveals Divergent Pathways to Violence

This paper introduces a new method using cognitive linguistics and LLMs to measure identity fusion through language, demonstrating that it can accurately predict fusion scores and distinguish between two different psychological pathways to extremist violence.

Original authors: Devin R. Wright, Justin E. Lane, F. LeRon Shults

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Devin R. Wright, Justin E. Lane, F. LeRon Shults

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "Psychological DNA" of Extremism: A Simple Breakdown

Imagine you are looking at a massive, swirling ocean of social media posts, comments, and messages. Most of it is just noise—people talking about lunch, weather, or movies. But hidden deep within that ocean are a few "rogue waves"—individuals who are moving toward extreme violence.

For a long time, security experts have tried to spot these waves by looking at what people are saying (e.g., "Are they talking about politics? Are they using hate speech?"). The problem is, that’s like trying to predict a storm by looking at the color of the water. It’s not accurate enough, and there is simply too much water to check by hand.

This paper introduces a new way to find those rogue waves. Instead of looking at the topic, they look at the psychological "fingerprint" left behind in the way people use language.

1. The Core Concept: Identity Fusion

To understand this paper, you first need to understand Identity Fusion.

Think of your identity like a collection of different hats you wear: the "Professional" hat, the "Parent" hat, the "Sports Fan" hat. Usually, these hats are separate. You can be a dedicated fan of a team without losing yourself in it.

Identity Fusion is what happens when those hats melt together. It’s when the "Sports Fan" hat fuses so strongly to your actual head that you can no longer tell where you end and the team begins. If someone insults the team, it doesn't feel like a critique of a hobby; it feels like a physical attack on your very soul. When this happens to political or religious groups, it becomes a powerful engine for violence.

2. The New Tool: CLIFS (The "Language Microscope")

The researchers created a tool called CLIFS. Think of CLIFS as a high-tech microscope that doesn't look at the words themselves, but at the metaphors hiding underneath them.

Most AI looks for "bad words." CLIFS looks for "fusion patterns." It uses advanced AI (Large Language Models) to see if a person’s language suggests they are "fusing" with a group.

For example, it looks for:

The Family Connection: Does the person talk about their political group as if they are brothers and sisters? (This is called "Fictive Kinship").
The Mirror Effect: Does the person describe the group using the same words they use to describe themselves? (If they say "I am a soldier" and "The Party is a soldier," the two are fusing).

3. The Big Discovery: Two Different Paths to Darkness

The most exciting part of the study is that they analyzed "manifestos" (the writings left behind by violent extremists) and discovered that there isn't just one way to become dangerous. They found two distinct "pathways":

Pathway A: The Ideologue (The "Soldier")

The Vibe: These people are driven by a grand cause or a strict belief system.
The Language: They use "family" language. They see the group as their kin. They are deeply woven into the fabric of an ideology.
Analogy: Think of a soldier who lives and breathes for the flag. The flag isn't just a piece of cloth; it is their family, their identity, and their life.

Pathway B: The Victim (The "Mirror")

The Vibe: These people aren't necessarily driven by a grand political theory. Instead, they feel personally wronged—bullied, rejected, or treated unfairly.
The Language: They don't necessarily talk about the group as "family." Instead, they project themselves onto the group. They see the group as a reflection of their own pain.
Analogy: Think of someone who has been pushed into a corner and decides to join a crowd, not because they love the crowd's rules, but because they see the crowd as a giant version of their own anger.

Why does this matter?

Currently, social media companies try to stop extremism by "censoring" topics or deleting "bad words." This often fails and can actually make people angrier (the "Streisand Effect").

This paper suggests a smarter way: Don't look at what they are talking about; look at how they are feeling. By using CLIFS, we can identify the psychological signature of someone who is becoming dangerously "fused" with a group, regardless of whether they are talking about religion, politics, or something else entirely. It’s the difference between trying to stop a fire by looking for smoke, and using a thermal camera to find the heat before the flames even appear.

Technical Summary: Language Predicts Identity Fusion Across Cultures and Reveals Divergent Pathways to Violence

1. Problem Statement

The rapid rise in political violence and online radicalization has rendered traditional content moderation (e.g., keyword filtering and hate speech classification) insufficient. Current AI systems often fail because they focus on topics (what is being said) rather than the psychological mechanisms (why it is being said) that drive extremism.

A critical psychological predictor of extreme violence is Identity Fusion—a state where an individual’s personal identity becomes so intertwined with a group or ideology that the boundaries between "self" and "target" blur. Existing computational methods to measure fusion, such as the Violence Risk Index (VRI), are criticized for being non-scalable, requiring heavy manual curation, and lacking generalizability across different cultural or linguistic contexts.

2. Methodology: The CLIFS Approach

The authors introduce the Cognitive Linguistic Identity Fusion Score (CLIFS), a scalable, cognition-informed framework that estimates fusion levels using written text alone. Unlike lexicon-based methods, CLIFS integrates four sophisticated components:

Cognitive Linguistic Patterns: Utilizing established research on how identity is expressed through language.
LLM Opaque Embeddings: Using large language models (e.g., all-mpnet-base-v2) to capture high-dimensional semantic signals.
Fine-tuned Classifiers: Using a ModernBERT classifier for coarse-grained fusion-level probabilities.
Implicit Metaphor Detection: A central innovation using Masked Language Models (MLM) to derive four key metrics:
- Fictive-Kinship ( $K_f$ ): How much the individual views the group as "family."
- Fusion Proximity ( $f(I,T)$ ): The bidirectional overlap between self ( $I$ ) and target ( $T$ ).
- Directional Identity Scores ( $S_{I \to T}$ and $S_{T \to I}$ ): Measuring the perceived overlap from self-to-group and group-to-self.

Experimental Design:
The researchers validated CLIFS using three distinct datasets:

MTurk Dataset: Religious conversion/reaffirmation descriptions from the UK and Singapore.
Singapore Field Data: Semi-structured interviews from local church participants.
Extremist Manifesto Corpus: A collection of violent manifestos, categorized into "Ideologues" (driven by ideology) and "Victims" (driven by perceived personal grievance/victimhood).

3. Key Contributions

Scalable Psychological Profiling: Provides a method to detect the "psychological signature" of extremism without needing to monitor specific political topics.
Cross-Cultural Validation: Demonstrates that the linguistic markers of identity fusion are robust across different countries (UK vs. Singapore) and communication domains (interviews vs. manifestos).
Methodological Critique: The paper provides a rigorous debunking of the VRI-Fusion method, proving that prior "state-of-the-art" models were often reliant on post-hoc manual curation rather than true predictive power.
Theoretical Refinement: Expands Comprehensive Identity Fusion Theory (CIFT) by showing that fusion can occur through non-kinship pathways (e.g., asymmetric self-projection).

4. Results

Superior Predictive Accuracy: In all cross-cultural tests, CLIFS significantly outperformed baseline methods (UAI, nUAI, and VRI-Fusion). CLIFS achieved the highest Spearman correlations ( $r_s$ ) with ground-truth psychometric scores and the lowest Mean Absolute Error (MAE).
Divergent Pathways to Violence: The study identified two distinct linguistic profiles for extremists:
- Ideologues: Exhibit higher Fictive-Kinship ( $K_f$ ) and Self-to-Target ( $S_{I \to T}$ ) scores. They frame the group in familial terms and align their identity with the group's semantics.
- Victims: Exhibit higher Fusion Proximity ( $f(I,T)$ ) and Target-to-Self ( $S_{T \to I}$ ) scores. They do not necessarily view the group as "family" but instead project their own identity onto the group, viewing the group as an extension of their personal grievances.
Robustness: CLIFS maintained performance even when text length was restricted, proving it captures deep semantic structures rather than mere word frequency.

5. Significance

This research shifts the paradigm of extremism detection from content-based moderation (which risks censorship and polarization) to mechanism-based detection. By identifying the underlying psychological state of identity fusion, social media platforms and security institutions can more accurately triage massive datasets to find individuals approaching a radicalization threshold. This approach offers a way to address the "needle in the haystack" problem of online radicalization through a scientifically grounded, scalable, and privacy-preserving computational lens.