Source-linked AI summary
LLM-based Detection of Manipulative Political Narratives
Sinclair Schneider, Florian Steuber, Gabi Dreo Rodosek
TL;DR
The paper asks how politically manipulative strategic narratives can be identified and structured in large-scale social media while distinguishing them from legitimate critique. It proposes an LLM-driven pipeline that filters posts with few-shot reasoning, clusters intent-focused embeddings using density-based methods, and extracts narratives from the resulting groups. Applied to over 1.2 million posts, the approach identified 41 distinct narrative clusters, while its filtering remains limited by false positives and ambiguity between coordinated manipulation and personal radical views.
Problem
The paper addresses the challenge of separating manipulative storylines and malicious reframing from legitimate political critique in large-scale social media data.
Method
The pipeline uses few-shot reasoning for FIMI filtering, intent-driven embeddings, density-based clustering, and specialized narrative extraction.
Results
41 distinct narrative clusters were identified around German political figures, while the filtering classifier achieved 0.92 recall and 0.66 precision.
Takeaways & Limitations
Combining stringent prompt-based filtering with HDBSCAN enables narrative clustering that can exclude noise and does not require a predefined list of narrative categories.
Takeaways & Limitations
The method has difficulty separating coordinated manipulative narratives from populist, radical, personal political views and depends on manually engineered prompts.
Abstract
from arXiv · showhide
We present a new computational framework for detecting and structuring manipulative political narratives. A task that became more important due to the shift of political discussions to social media. One of the primary challenges thereby is differentiating between manipulative political narratives and legitimate critiques. Some posts may also reframe actual events within a manipulative context. To achieve good clustering results, we filter manipulative posts beforehand using a detailed few-shot prompt that combines documented campaign narratives with legitimate criticisms to differentiate them. This prompt enables a reasoning model to assign labels, retaining only manipulative narrative posts for further processing. The remaining posts are subsequently embedded and dimensionality-reduced using UMAP, before HDBSCAN is applied to uncover narrative groups. A key advantage of this unsupervised approach is its independence from a predefined list of target categories, enabling it to uncover new narrative clusters. Finally, a reasoning model is employed to uncover the narrative behind each cluster. This approach, applied to over 1.2 million social media posts, effectively identified 41 distinct manipulative narrative clusters by integrating prompt-based filtering with unsupervised clustering.
1 Introduction
The paper addresses how to identify manipulative strategic narratives in large-scale social media while separating them from legitimate political critique. It proposes an LLM-driven pipeline combining prompt-based filtering, intent-focused embeddings, and narrative-oriented clustering.
- Motivation: Manipulative campaigns increasingly use social media to inject strategic content into domestic political discourse.Modern campaigns combine established rhetorical motifs with techniques such as cloned news outlets, narrative laundering, forged evidence, and staged videos.
- Research problem: The central challenge is distinguishing manipulative storylines and malicious reframing from legitimate but controversial political critique.The approach targets manipulative intent and rhetorical motifs rather than verifying whether every claim is strictly true.
- Approach: The proposed pipeline uses FIMI characteristics and few-shot examples to guide a reasoning model in separating manipulative content from legitimate critique.This prompt-based reasoning approach focuses on rhetorical nuances and isolates strategic manipulative content.
- Approach: Posts are embedded according to manipulative intent so related storylines become close in the embedding space.This shifts the adapted BERTopic pipeline from topic similarity toward intent-driven narrative grouping.
- Approach: A specialized extraction prompt captures complete FIMI storylines by including the core claim, targeted adversary, and manipulative angle.The method replaces simplistic topical keywords with strategic narrative descriptions.
2 Related Work and Fundamentals
Prior work largely builds datasets or applies topic modeling to predefined collections of manipulative content. The paper situates its approach around strategic narratives, FIMI, and the distinction among disinformation, misinformation, and malinformation.
- Related work: Existing research primarily constructs datasets or applies topic modeling to established corpora of political disinformation and malinformation.Related datasets include general collections and human-annotated posts focused on specific topics such as elections.
- Related work: DiNaM is the closest comparison, but it mines narratives from fact-check articles rather than unfiltered social media posts.The paper instead addresses the need to separate manipulative narratives from legitimate critique and unrelated content in social media.
- Fundamentals: Disinformation is false information deliberately created to cause harm, whereas misinformation is false information not created with that intention.These categories are distinguished by both falsity and harmful intent.
- Fundamentals: Mal-information is reality-based information used to inflict harm on a person, organization, or country.Its defining feature is harmful use of information grounded in reality.
- Strategic Narratives: A strategic narrative is a story that constructs meaning about political events to shape domestic and international actors’ behavior.The paper emphasizes that narratives extend beyond isolated topics or keywords.
The overall interaction
FIMI campaigns disseminate manipulative content through social platforms and other channels to shape audience behavior. Their operational motifs include cloned outlets and other mechanisms that turn political ambiguity into malicious storylines.
- Campaign-to-behavior interaction: A FIMI campaign connects deployed manipulative content with intended audience behavior change.The paper presents this interaction as the progression from campaign activity to behavioral effects.
- Campaign-to-behavior interaction: Campaigns can use false claims to reinforce negative perceptions and increase opposition to future political support.The Ukraine trafficking example links a fabricated claim to perceptions of corruption and opposition to support for Ukraine.
- Operational mechanisms: FIMI delivery mechanisms range from cloned websites to fabricated whistleblowers and rely on recurring rhetorical motifs.These mechanisms transform political ambiguity into malicious storylines.
- Operational mechanisms: The paper uses real-world FIMI campaigns and their motifs to establish criteria for few-shot detection queries.Table 1 summarizes the campaigns and operational motifs used for this purpose.
- Evidence base: The methodology draws on credible gray literature from public institutions, security agencies, fact-checkers, research organizations, and reliable news outlets.These sources are used to analyze recent strategic narratives and evolving social-media manipulation tactics.
3 Dataset
The dataset contains social media posts collected from X, Reddit, and Telegram around German political figures and parties. Figure 2 summarizes the flow from raw data to narrative labels.
- Dataset composition: 1,255,895 short posts were collected from X, Reddit, and Telegram.The corpus is an unfiltered dataset with an 80% German and 20% English split.
- Platform sources: X contributed 829,191 tweets, making it the largest platform portion of the dataset.Posts were collected by searching for German Bundestag politicians before the February 2025 federal elections.
- Platform sources: Reddit contributed 362,753 posts gathered through searches for German parties, politicians, and popular political channels.The sampled Reddit sources were more prominent among left-leaning content than in the X subset.
- Platform sources: Telegram contributed 63,951 messages from 219 groups using a more selective group-based sampling strategy.The Telegram sample mainly covered right-wing conspiracy groups.
- Processing flow: Figure 2 overviews the data flow from raw data to narrative labels.The paper states that the individual processing steps are described in Section 4.
4 Methodology
The methodology filters posts with few-shot, reasoning-based prompts before embedding and density-based clustering. It then uses specialized narrative extraction to label clusters rather than relying on conventional topic keywords.
- Prompt-based Filtering: Prompt-based filtering isolates manipulative content from legitimate critique and unrelated posts using FIMI characteristics and few-shot campaign examples.Human expertise and machine optimization iteratively refine the structured prompt.
- Prompt-based Filtering: The filtering prompt is processed with the Qwen3.5-122B-A10B-FP8 reasoning model through vLLM.The model uses a mixture-of-experts design with 122 billion total parameters and 10 billion activated parameters.
- Embedding and Reduction: Filtered posts are embedded with Qwen3-Embedding-8B using an instruction prompt targeting strategic narrative, manipulative intent, and disinformation motive.The resulting 4096-dimensional vectors are L2-normalized after generation.
- Embedding and Reduction: UMAP reduces embeddings to two dimensions for visualization and five dimensions for unsupervised clustering.Five dimensions follow the BERTopic standard and limit computational demands for large clusters.
- Clustering: HDBSCAN determines cluster structure without requiring a predefined number of narrative groups.The method tests minimum cluster sizes from 100 to 1000 and adjusts min_samples to balance clusters against noise.
- Narrative Labeling: Cluster narratives are extracted by supplying c-TF-IDF keywords and cluster documents to a larger reasoning model.This replaces simplistic topical keywords with complete strategic storylines containing the core claim, adversary, and manipulative angle.
5 Evaluation
Evaluation combines a balanced manual audit, reasoning-coherence assessment, and HDBSCAN hyperparameter testing. The filtering model favors recall and is stricter than human raters, while clustering identifies 41 distinct groups at the selected threshold.
- Validation: A two-stage manual audit evaluated 200 balanced posts, including 100 model-flagged manipulative posts and 100 non-manipulative posts.The first stage used blind, randomized human assessment, while the second examined reasoning coherence.
- Validation: 0.77 F1 was achieved by prompt-based filtering, with 0.92 recall and 0.66 precision.The model accepts more false positives to avoid discarding potential narrative fragments before clustering.
- Validation: The filtering model’s high recall is supported by HDBSCAN, which can classify falsely flagged posts as noise when they lack cohesive semantic density.This interaction mitigates the downstream impact of false positives.
- Boundary Analysis: 95.5% of model explanations received human agreement in the secondary reasoning-coherence evaluation.The high agreement contrasts with the lower F1 score because coherent reasoning can still accompany overly strict classifications.
- Boundary Analysis: The central boundary case is distinguishing coordinated manipulative narrative fragments from genuine private but highly populist opinions.A remigration slogan illustrates how the model may classify a political expression as manipulative despite the absence of evidence for coordinated external activity.
- Cluster Optimization: A minimum HDBSCAN cluster size of 400 posts was selected as the configuration’s sweet spot for limiting noise while maintaining semantic separation.At this threshold, HDBSCAN identified 41 distinct clusters.
6 Results
The pipeline identified 41 distinct FIMI narrative clusters around German political figures, which were manually grouped into four thematic pillars. The extracted narratives prominently feature betrayal and conspiracy motifs across migration, Ukraine, climate politics, and political salvation.
- 41 distinct narrative clusters surrounding German political figures were identified, with the five largest ranked by associated social-media post volume.The clusters were manually grouped into four main thematic pillars.
- Pillar 1: The “Great Replacement”: Migration narratives portray the government as deliberately enabling replacement, insecurity, and the erosion of native culture.Examples include state betrayal, civilizational replacement, and systemic betrayal involving migrants and political elites.
- Pillar 2: The Proxy War: Ukraine-related narratives frame Germany as non-sovereign and accuse elites of sacrificing national prosperity and peace for a proxy war.These narratives include occupation, deliberate betrayal, and energy-security claims linked to severed Russian gas ties.
- Pillar 3: Climate Politics: Climate-politics narratives depict climate policy as a fabricated authoritarian project that intentionally harms Germany’s economy and energy security.The Green Party and Robert Habeck are portrayed as ideological or incompetent agents of deliberate sabotage.
- Pillar 4: Political Salvation: Political-salvation narratives portray established parties and media as a corrupt deep state while positioning the AfD and external allies as saviors.The associated motifs include fraudulent democracy, internal destruction of sovereignty, and an external-savior role for Elon Musk and global allies.
7 Discussion
The discussion argues that intent-driven clustering can adapt topic modeling to produce narrative storylines rather than isolated topics. It also highlights complementary benefits from stringent LLM filtering and density-based noise removal, while noting ambiguity and reliance on manual prompt engineering.
- Intent-driven strategic narrative clustering maps posts by manipulative intent so clusters represent storylines rather than single-word topics.Similar manipulative intents are grouped together in the embedding space.
- Prompt-based filtering achieved recall 0.92 and precision 0.66 because it flags posts whenever a provided FIMI strategy might apply.HDBSCAN subsequently excludes false positives as noise when they do not form coordinated large clusters.
- Large reasoning models can use trained knowledge beyond prompt vocabulary, such as recognizing rtde.media as a proxy for Russia Today.The domain was not included in the filtering prompt, yet the model used that connection when flagging the post.
- The approach remains limited by the blurry boundary between coordinated manipulation and radical or populist personal political views.Its filtering prompt also depends on iterative human-AI collaboration and manual choices about campaign characteristics and few-shot examples.
8 Outlook
The outlook proposes reducing reliance on heuristic prompt design, tracing narrative origins across platforms, and extending FIMI analysis to multimodal content and reliable-source assessment.
- Programmatic prompt optimization could replace heuristic prompt engineering by searching for prompts and few-shot examples against an annotated FIMI corpus.The proposed corpus would contain both FIMI and non-FIMI posts to support systematic optimization for precision and recall.
- Cross-platform analysis of web and social-media clustering outputs could investigate narrative origins, campaign dynamics, and vulnerability to manipulative content.The outlook also connects this analysis with identifying sources and individuals particularly vulnerable to manipulation.
- A broader FIMI framework could incorporate visual language models and reliable-source repositories to process memes, manipulated images, videos, and contextual authenticity cues.This is presented as a possible defense against future disinformation waves.
9 Data and Code Availability
The paper provides the source code for its FIMI detection pipeline to support scientific transparency and reproducibility.
- The FIMI detection pipeline’s source code is available on GitHub.
10 Ethical Considerations and Statement of Objectivity
The paper presents its strategic narratives as unedited LLM-pipeline outputs for transparency and reproducibility, while disclaiming endorsement of the political content. The authors also report no relevant competing interests and acknowledge institutional support.
- The strategic narratives are unedited LLM-pipeline outputs provided for transparency and reproducibility.
- The authors reject the political opinions, conspiracy theories, and manipulative claims represented in the dataset.
- The authors acknowledge the System Sciences Chair for Communication Systems and Network Security under Prof. Dr. Gabi Dreo Rodosek’s direction.
- The authors declare no competing interests relevant to the article’s content.