Linguistic Markers of Population Replacement Conspiracy Theories in YouTube Immigration Discourse

Erik Bran Marino, Davide Bassi, Renata VieiraCLiC-it 2025

Presented at the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025), this research explores the subtle linguistic boundaries between mainstream political dissent and fringe conspiracy theories. The study focuses specifically on Population Replacement Conspiracy Theories (PRCT)—frequently termed the "Great Replacement"—within the high-volume, often volatile ecosystem of YouTube comments.

The Challenge of Fringe Discourse Detection

Online moderation remains a complex challenge. Conspiracy theories like PRCT are notoriously difficult to detect because they are often "lexically sparse," avoiding obvious slurs or banned keywords. Instead, they frequently hide within legitimate political debates regarding immigration policy.

Existing automated tools often struggle to distinguish between a user expressing a standard policy-based objection (a "CONTRA" stance) and one promoting a narrative of an orchestrated "elite plot" to replace native populations. This study aims to bridge that gap by identifying the specific psycholinguistic "fingerprint" of conspiracy discourse and testing the efficacy of modern open-weight Large Language Models (LLMs) in identifying these patterns.

Methodology: A Hybrid Pipeline

The researchers employed a multi-stage hybrid pipeline to analyze a corpus of 71,137 unique YouTube comments scraped from 15 immigration-related videos:

  1. Stance Classification: GPT-4o categorized comments into PRO, NEUTRAL, and CONTRA (anti-immigration) stances.
  2. Conspiracy Detection: DeepSeek-v3, an open-weight LLM, was utilized in a few-shot learning configuration to identify PRCT elements. DeepSeek-v3 achieved a remarkable 94.5% accuracy on a human-annotated gold standard dataset.
  3. Psycholinguistic Profiling: The team used LIWC (Linguistic Inquiry and Word Count) to analyze text across dimensions such as parts of speech, emotional categories, and cognitive processes.
  4. Statistical Validation: Differences were validated using Kruskal-Wallis tests and Cohen’s d to ensure findings represented substantial trends rather than statistical noise.

The "Intensification" Hypothesis

One of the study's most striking findings is the 89.7% linguistic overlap between conspiracy theories and non-conspiratorial anti-immigration discourse. This suggests that PRCT is not a separate dialect or a distinct "language" of its own. Rather, it is a radicalized version of mainstream rhetoric that "intensifies" specific themes already present in the broader discourse.

Psycholinguistic Profiling: Analytic but not Insightful

The LIWC analysis revealed a unique cognitive profile for PRCT proponents. While one might expect conspiracy rhetoric to be disorganized, the data suggests the opposite.

LIWC analysis comparing PRCT, PRO, NEUTRAL, and CONTRA stances across key linguistic dimensions including 'Analytic' and 'Insight'.

As illustrated in the figure above, PRCT comments scored the highest in "Analytic Thinking" but the lowest in "Insight" language. This indicates that conspiracy theorists are not merely rambling; they utilize structured, logical-sounding frameworks to validate pre-existing beliefs. However, they show a marked lack of cognitive flexibility or "aha moments" typically associated with insight-related vocabulary.

The Four Pillars of PRCT

The researchers identified four specific dimensions where PRCT significantly deviates from standard anti-immigration talk:

  • Religion (+115.6%): Framing immigration as a spiritual or civilizational threat (e.g., "Islamization").
  • Conflict (+95.2%): Heavy use of warfare metaphors, such as "invaders" or "defending borders."
  • Power (+41.4%): Attributing demographic changes to orchestrated manipulation by "globalist" elites.
  • Negative Tone (-22% positive sentiment): A marked absence of positive sentiment and a focus on apocalyptic urgency.

Conclusion and Future Directions

The study demonstrates that open-weight models like DeepSeek-v3 are now viable for complex, niche moderation tasks previously reserved for closed-source giants. By identifying the specific linguistic markers of radicalization—such as the shift toward warfare metaphors and elite-focused power dynamics—platforms can better distinguish between healthy political disagreement and the spread of harmful conspiracy narratives.

Future work will look toward longitudinal studies to track how these linguistic markers evolve over time as individual users move from mainstream dissent toward radicalized conspiratorial worldviews.

Summary updated

BibTeX
@inproceedings{marino2025linguistic,
  author = {Erik Bran Marino and Davide Bassi and Renata Vieira},
  title = {{Linguistic Markers of Population Replacement Conspiracy Theories in YouTube Immigration Discourse}},
  booktitle = {CLiC-it},
  year = {2025},
}
Figure