Unmasking Digital Falsehoods: A Comparative Analysis of LLM-Based Misinformation Detection Strategies
Tianyi Huang ∗ \ast, Peiyang Yu, Jingyuan Yi, Xiaochuan Xu · 2025
Read Unmasking Digital Falsehoods: A Comparative Analysis of LLM-Based Misinformation Detection Strategies by Tianyi Huang ∗ \ast, Peiyang Yu, Jingyuan Yi, Xiaochuan Xu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00724.txt.
Introduction
Tianyi Huang and co-authors compare ways to use large language models as lie detectors for online falsehoods in a 2025 paper. False claims spread fast on social media and do real harm. Language models read and reason well, so many teams have tried them as fact checkers.
The review sorts methods into three kinds: text-only, mixed media, and agent teams that check step by step. It weighs tuned models, no-example use, and strict checking routines across health, politics, and finance topics. Each style has weak spots, from made-up facts to attack tricks and high compute cost. The authors back hybrid designs that pair firm check steps with models that keep learning. Future work may track falsehoods live.
Catalog record
| CREATOR | Tianyi Huang ∗ \ast, Peiyang Yu, Jingyuan Yi, Xiaochuan Xu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-00724
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Unmasking Digital Falsehoods: A Comparative Analysis of LLM-Based Misinformation Detection Strategies Thanks: ∗ Corresponding author: [email protected] Tianyi Huang ∗ \ast, Peiyang Yu, Jingyuan Yi, Xiaochuan Xu Abstract The proliferation of misinformation on social media has raised significant societal concerns, necessitating robust detection mechanisms. Large Language Models such as GPT-4 and LLaMA2 have been envisioned as possible tools for detecting misinformation based on their advanced natural language understanding and reasoning capabilities. This paper conducts a comparison of LLM-based approaches to detecting misinformation between text-based, multimodal, and agentic approaches. We evaluate the effectiveness of fine-tuned models, zero-shot learning, and systematic fact-checking mechanisms in detecting misinformation across different topic domains like public health, politics, and finance. We also discuss scalability, generalizability, and explainability of the models and recognize key challenges such as hallucination, adversarial attacks on misinformation, and computational resources. Our findings point towards the importance of hybrid approaches that pair structured verification protocols with adaptive learning techniques to enhance detection accuracy and explainability. The paper closes by suggesting potential avenues of future work, including real-time tracking of misinformation, federated learning, and cross-platform detection models. I Introduction I-A Background and Motivation 1) The Rise of Misinformation on Social Media The Social media’s proliferation has transformed information dissemination by enabling instant access but also accelerated fake news, rumors, and conspiracy theories [ 1 ] . Large language models (LLMs) exacerbate this by mass-producing convincing misinformation [ 2 ] . Misinformation’s societal impacts include eroded public trust, disrupted democratic processes (e.g., 2016 U.S. election polarization), and public health crises like COVID-19 vaccine hesitancy [ 3 ] . It also threatens economic stability and scientific progress by fostering distrust in markets and research [ 4 ] . Multimodal misinformation (images, videos, audio) further complicates detection, necessitating advanced cross-modal analysis tools [ 5 ] . 2) Role of LLMs in Misinformation Detection Large language models (LLMs) like GPT-4 and LLaMA show promise in combating misinformation through advanced language understanding and reasoning. They can fact-check claims by cross-referencing factual databases and detecting inconsistencies in text or multimodal content. Ensemble NLP models may further improve detection by refining linguistic coherence and analytical rigor [ 6 ] . Various techniques have been tried for applying LLMs in identifying misinformation, including: • Prompt Prompt-based Detection: Zero-shot, few-shot, or chain-of-thought prompting to evaluate claim validity. • Agentic Verification: Systems like FactAgent mimic human fact-checking by breaking verification into subtasks [ 7 ] . • Multimodal Analysis: Combining LLMs with vision-language models (LVLMs) to detect inconsistencies in text-image pairs. LLMs face limitations like hallucination risks and reliance on high-quality training data [ 8 ] . Future work must prioritize model transparency, explainability, and adversarial robustness. I-B Research Objectives 1) Analyzing the Effectiveness of LLMs in Misinformation Detection The research in this paper is to access the usability of LLMs to detect misinformation in different domains of study, e.g., politics, public health, and finance. 2) Comparing Text-Based vs. Multimodal Approaches The study tries to compare text-based and multimodal approaches in identifying misinformation based on effectiveness. While text-based models deal with linguistic signals and fact-checking variety alone, multimodal models process textual, visual, and auditory inputs for higher accuracy [ 9 ] . 3) Evaluating Performance Across Datasets and Benchmarks To give a general idea of the efficacy of LLM, the research will compare different models with benchmarked datasets that are standardized, i.e., FakeNewsNet, SciNews, and MM-COVID. Recent studies on COVID-19 misinformation emphasize the need for datasets that unify multiple modalities, as combining text, images, and user behavior data improves misinformation detection accuracy [ 10 ] . I-C Contributions of This Review 1) Comprehensive Overview of Existing Detection Approaches The paper offers a unified overview of state-of-the-art detection algorithms of misinformation, ranging from text to multimodal and hybrid. 1) Performance-Based Comparison of Leading Models Comparative analysis shall be conducted amongst leading LLMs for detection of misinformation as per key performance parameters such as accuracy, precision, recall, and explainability. 2) Identification of Key Challenges and Future Directions Identifying such major challenges for detecting misinformation as model prediction bias, evolving misinformation methods, and scalability, the review would recommend future research directions on how LLM-based detection mechanisms can be optimized and made adaptive. II Taxonomy of Misinformation and Detection Challenges II-A Types of Misinformation TABLE I: Category Definition Characteristics Fake News Deliberately fabricated or misleading information presented as legitimate news to deceive audiences. Exploits sensationalism and emotional appeal to drive engagement and influence public opinion. Rumors Unverified information that spreads rapidly. Fueled by uncertainty and speculation, it often lacks credible sources. Conspiracy Theories Propose hidden, malevolent forces behind significant events. Lack credible evidence but resonate with those seeking explanations. Disinformation Intentionally spread falsehoods aimed at misleading or manipulating audiences. Often politically or ideologically motivated, strategically crafted to deceive. Misinformation Unintentionally shared incorrect information without malicious intent. Can still contribute to misinformation spread and undermine trust in reliable sources. Clickbait Sensationalized headlines designed to attract engagement. Prioritizing clicks and views over factual accuracy, may mislead audiences. Propaganda Information crafted to influence public perception and promote agendas. Used for political, ideological, or persuasive purposes, often one-sided or biased. II-B Characteristics of Misinformation 1) Spread Patterns (Viral Nature, Echo Chambers) Misinformation spreads rapidly due to social media algorithms that boost engaging material, which leads to viral spreading. 2) Psychological Influence and User Biases Cognitive biases such as confirmation bias and illusory truth effect are responsible for the perpetuation of the acceptance of misinformation [ 11 ] . Users verify information that reinforces their pre-existing beliefs and will find it easier to reject corrections. Emotion plays a key role in misinformation spread. Studies show LLaMA3 excels at identifying emotions in short text, which could aid in detecting emotionally charged misinformation [ 12 , 13 ] . 3) Multimodal Misinformation (Text, Image, Video) Modern misinformation typically uses a combination of text, images, and video for greater credibility and effectiveness. The multimodal format makes detection more difficult because sophisticated cross-modal analysis methods are required [ 14 ] . Context is key in geospatial misinformation, as location verification studies show surrounding tweets improve model accuracy [ 15 ] . II-C Challenges in Misinformation Detection 1) Contextual Ambiguity and Intent Detection Satire, opinion, and deliberate misinformation are hard to identify due to contextual ambiguity. Differences in culture, linguistic subtleties, and diverse perceptions prevent machines from properly inferring intent. 2) Scalability and Generalizability of Detection Methods Misinformation keeps on evolving, and therefore detection also has to be adaptive and scalable. Generalization across different languages, domains, and platforms is a key challenge because models trained on specific data may not fare well in different environments. Layer-wise task encoding in LLMs affects misinformation detection, as complex inferential reasoning may require deeper layers for accurate assessment [ 16 ] . Multi-event extraction models like DEEIA improve contextual understanding and efficiency, which could enhance misinformation detection across related events [ 17 ] . 3) Resource Limitations in Real-Time Detection Real-time misinformation detection requires significant computational resources and infrastructure, which requires high processing power to handle massive volumes of content in different modalities. II-D Comparison of Detection Challenges Across Approaches 1) Challenges in Text-Based Approaches Text-based detection mechanisms struggle with sarcasm, irony, and hidden communication, which can hide the real intent of messages.