Population Replacement Conspiracy Theories
- Population Replacement Conspiracy Theories (PRCT) are narratives alleging that native populations are deliberately replaced by migrants through orchestrated plans and provocative metaphors.
- The dataset comprises 1,617 headlines from Spanish, Italian, and Portuguese media, annotated with binary PRCT labels along with stance and rhetorical biases.
- Classification experiments using models like mBERT, Llama, and Mistral highlight varied precision, recall, and the challenges of automated detection amid subjective linguistic cues.
Population Replacement Conspiracy Theories (PRCT) designate a category of conspiratorial narratives characterized by the claim that "native populations are being deliberately replaced by migrants." These theories operate by invoking a framing in which migratory processes are not simply demographic or economic, but instead the effect of a coordinated plan—frequently referenced by terms such as “population replacement,” “ethnic substitution,” or metaphorical language (“poisoning our blood”). PRCTs represent a salient component of contemporary hyperpartisan discourse, exhibiting strong correlations with political polarization, institutional distrust, and potential for motivating extremist violence. Their computational detection is integral to the broader objective of misinformation mitigation and is now studied across multiple linguistic and socio-political contexts, notably in PartisanLens, a multilingual benchmark covering Spanish, Italian, and Portuguese (Maggini et al., 7 Jan 2026).
1. Formal Definition and Taxonomy
Within the PartisanLens framework, PRCT is operationalized as a binary label assigned to headlines or statements if they contain any explicit or implicit reference to a malicious, strategic process intended to “replace” an ingroup population with an outgroup (typically migrants). The exact decision rule states: PRCT=True is triggered by the presence of coordinated demographic plot frames or analogous conspiracy metaphors. Examples include both direct phrasing (“Invaders plan ethnic takeover”) and more oblique allusions (“poisoning our blood”).
The annotation taxonomy remains flat (True/False), without subdivision by alleged perpetrator or mechanisms, but is consistently co-annotated with:
- Hyperpartisan style (binary: one-sided/extreme)
- Stance (pro / neutral / against)
- Three span-level rhetorical biases (loaded language, appeal to fear, name-calling)
This delineation ensures consistent cross-lingual operationalization and enables joint analysis with related phenomena such as rhetorical polarization and stance.
2. Dataset Structure and Label Distributions
PartisanLens presents 1,617 news headlines sourced from Spanish (525), Italian (565), and Portuguese (527) media outlets. Each item is manually annotated with PRCT status plus the aforementioned co-occurring labels.
A breakdown of PRCT versus non-PRCT headlines by language is shown below:
Across all languages, hyperpartisan narratives are present in 336 (SPA), 337 (ITA), and 367 (PT) headlines. Stance, rhetorical bias, and other label distributions are comprehensively tabulated in the dataset, with loaded-language and appeal-to-fear marking substantial fractions of hyperpartisan content, especially in Spanish and Italian. These figures reveal sharp imbalances in PRCT prevalence by language, a pattern mirrored across stance orientation and rhetorical device distributions.
3. Annotation Protocol and Reliability
Annotation was conducted by nine native speakers (three per language, mixed gender, Master’s+ educational background) in a masked, multi-stage procedure: pilot, refinement, and full annotation rounds, totaling roughly 66 hours. Outlets were masked to avoid bias.
The headline annotation protocol includes:
- Hyperpartisan: one-sided or highly emotional framing
- PRCT: presence of demographic replacement conspiratorial narrative per the above definition
- Stance: pro/neutral/against in reference to immigration policy
- Span-level rhetorical marking: loaded language, appeal to fear, name-calling
Inter-annotator agreement was quantified using Fleiss' κ:
| Language | Hyperpartisan | PRCT | Stance |
|---|---|---|---|
| SPA | 0.876 | 0.880 | 0.837 |
| ITA | 0.706 | 0.774 | 0.744 |
| PT | 0.770 | 0.721 | 0.737 |
By conventional standards, these κ values denote substantial to almost perfect agreement. The highest consistency is observed in Spanish, especially for PRCT detection, indicating that conspiratorial replacement narratives manifest with salient and cross-annotator-recognizable linguistic features.
4. Baseline Model Benchmarking and Classification Performance
Classification experiments systematically evaluate multilingual transformer architectures, employing both zero/few-shot prompting and supervised fine-tuning:
- mBERT (fine-tuned, 110M parameters)
- Llama 3.1–8B (zero-/few-shot; fine-tuned; rationale-augmented)
- Mistral–NeMo–Base–2407 (12B parameters, identical schemes)
- Llama 3.3–70B (all four strategies)
Model outputs conform to a standardized JSON format including hyperpartisan, PRCT, and stance keys. Classification quality is measured by macro-averaged precision (P), recall (R), and F1, ameliorating class imbalance effects.
PRCT detection results (excerpted):
| Model | Precision | Recall | F1 |
|---|---|---|---|
| mBERT (FT) | .8589 | .8148 | .8345 |
| Llama 8B (FS) | .7274 | .8041 | .7540 |
| Llama 70B (ZS) | .7814 | .8537 | .8090 |
| Llama 70B (FS) | .8281 | .8175 | .8226 |
| Nemo (FS) | .9325 | .7237 | .7814 |
Key insights:
- PRCT detection outperforms hyperpartisan identification, attributed to a more distinct conspiratorial lexicon.
- Nemo achieves peak precision but at the cost of recall, indicating a conservative bias.
- Few-shot prompting with Llama 70B rivals supervised mBERT, while rationale-augmented fine-tuning yields minimal gains (explanations encode decision style rather than conspiracy features).
5. LLM Annotation and Human Alignment
Evaluations of LLMs as automatic annotators involve mean pairwise Cohen’s κ between human–human (H–H), Llama–human (L–H), and gpt-5-mini–human (G–H) pairs:
| Label | H–H | L–H | G–H |
|---|---|---|---|
| SPA-PRCT | .893 | .483 | .549 |
| ITA-PRCT | .743 | .632 | .541 |
| PT-PRCT | 1.000 | .124 | .662 |
LLMs lag behind human agreement, markedly in Spanish and Portuguese. Leading sources of error include:
- Misinterpretation of sarcasm or irony
- Inability to resolve polysemy (e.g., “Imigrantes” as both place and demographic category)
- Cultural rigidity in LLM understanding (e.g., direct mapping of “invadere” to migration contexts)
Persona-based prompting partially mitigates these deficits (see below).
6. Persona-Based Conditioning for Subjectivity Modeling
Each human annotator's demographic and ideology were profiled via Political Compass and PRISM surveys, then rendered into persona prompts for zero-shot LLM classification. Agreement statistics (PRCT κ):
| Persona | Llama 70B | gpt-5-mini |
|---|---|---|
| SPA_1 | 0.5531 | 0.6252 |
| ITA_1 | 0.7409 | 0.7274 |
| PT_1 | 0.3151 | 0.6622 |
Persona conditioning increases PRCT annotation agreement across all languages, particularly for Portuguese with gpt-5-mini (κ up to .662). Stance labeling also improves under persona prompts, though hyperpartisan trends are less consistent. This approach supports a nuanced simulation of individual annotator variability and subjectivity in conspiracy detection.
7. Methodological Implications and Future Directions
Empirical findings demonstrate that PRCT detection is the most tractable among hyperpartisan, stance, and conspiracy-related classification problems, owing to the salience of conspiratorial linguistic motifs. Hyperpartisan style occupies an intermediate tier of difficulty: its distributed lexical markers are frequently interwoven with stance and emotional rhetoric. Stance detection is the most challenging, requiring subtle pragmatic and socio-cultural knowledge.
Contemporary LLMs, including those at 70B parameter scale, remain inferior to human annotators for subtle or contextually-dependent phenomena (e.g., sarcasm, polysemy). Few-shot prompting and rationale-based techniques yield measurable gains, but task-dependent patterns emerge: whereas rationales encode decision structures, they do not directly enhance conspiracy content recognition.
Persona-based model conditioning is an effective vector for modeling subjective dimensions—critical for contexts where misinformation is evaluated through culturally- and ideologically-contingent lenses. A plausible implication is that future architectures should combine stronger context modeling, greater linguistic diversity, and cross-lingual transfer learning to further close the gap between automated and human annotation.
Comprehensive resources, including code, data, and annotation protocols, are archived at https://github.com/MichJoM/PartisanLens, supporting transparent evaluation and extension of PRCT detection in multilingual and cross-cultural settings (Maggini et al., 7 Jan 2026).
