💬 NLP

Generative Large Language Models in Automated Fact-Checking: A Survey

This survey systematically reviews 199 research papers to present a comprehensive taxonomy and analysis of how generative large language models are integrated into automated fact-checking workflows, covering their roles in verification, data generation, evaluation, and multilingual applications while identifying current limitations and future research directions.

Original authors: Ivan Vykopal, Matúš Pikuliak, Simon Ostermann, Marián Šimko

Published 2026-06-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ivan Vykopal, Matúš Pikuliak, Simon Ostermann, Marián Šimko

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 Big Picture: The Information Flood and the New Librarian

Imagine the internet is a massive, chaotic ocean. Every day, millions of waves of information crash onto the shore. Most are true, but some are dangerous "fake waves" (misinformation) that can wash away trust, confuse people, and cause real-world harm.

For a long time, we relied on human "lifeguards" (fact-checkers) to swim out, grab the fake waves, and check if they were real. But the ocean is getting too big, and the waves are coming too fast. The lifeguards are drowning.

Enter Generative Large Language Models (LLMs). Think of these as super-smart, tireless robot assistants that can read, write, and reason. This paper is a massive report card on how we are currently using these robots to help the lifeguards. The authors looked at 199 different research papers to see exactly what these robots are doing, where they are failing, and how we can make them better.

The Three Main Jobs of the Robot Assistant

The paper organizes the robots' work into three main categories, like a factory assembly line:

1. The Data Factory (Synthetic Data Generation)

The Problem: To train a robot to spot fake news, you need thousands of examples of real and fake news. But finding and labeling these examples by hand is slow and expensive. It's like trying to teach a child to recognize apples by only showing them three apples.
The Robot's Job: The robots are now being used to make up their own practice tests. They can take a real news story and rewrite it, translate it, or even invent a fake version that looks real.

The Catch: If the robot makes up too many "perfect" fake stories, it might get too good at spotting robot-made fakes and fail to spot human-made fakes. It's like training a security guard only on photos of fake ID cards; they might miss the real forgeries.

2. The Assembly Line (Prediction Tasks)

This is the core work: actually checking the claims. The paper breaks this down into four stations:

Station A: The Claim Detector. The robot scans a messy social media post and says, "Hey, this sentence is a claim we can check!" It cleans up the sentence, removing slang and confusion, so it's ready for inspection.
Station B: The Archive Search. The robot asks, "Has anyone checked this before?" It searches a database of past fact-checks to see if this story is an old lie that has already been debunked.
Station C: The Evidence Hunter. If it's a new claim, the robot goes out and finds proof. It searches the internet for articles, scientific papers, or official reports that support or deny the claim.
Station D: The Verdict. Finally, the robot looks at the evidence and decides: True, False, or "Not Enough Info." Crucially, it can now write an explanation for why it made that decision, acting like a teacher explaining the answer to a student.

3. The Quality Control Inspector (Evaluation)

The Problem: How do we know if the robot is doing a good job? In the past, we used simple math to compare the robot's answer to the "correct" answer. But if the robot writes a long, fancy explanation, simple math can't tell if the reasoning is actually logical or just sounds good.
The Robot's Job: Now, we are using one robot to grade another robot. We ask a second robot to read the first robot's work and say, "Is this explanation honest? Did it use the right evidence?"

The Catch: If the grading robot is biased or confused, it might give a bad grade to a good job, or vice versa. It's like asking a student to grade their own homework; they might be too kind to themselves.

The Special Case: The Multilingual Challenge

The paper points out a huge imbalance. Most of the robots are trained on English. They are like native English speakers who are very smart but struggle to understand a conversation in Hindi, Swahili, or Arabic.

The Current Fix: Many researchers translate the foreign claim into English, let the robot check it, and then translate the answer back.
The Risk: This is like playing a game of "Telephone." By the time the message gets back to the original language, the meaning might have changed, or a subtle cultural joke might be lost, leading to a wrong verdict. The paper argues we need robots that can think and check facts directly in those local languages, not just rely on translation.

The Robot's Weaknesses (Hallucinations)

The paper highlights a major flaw: Hallucinations.
Imagine a robot that is so confident it invents facts. It might say, "This claim is false because a famous scientist said so," and then invent a fake name for that scientist.

The Danger: The robot gets the final answer right (the claim is false), but for the wrong reason (fake evidence). This is dangerous because it looks trustworthy but is actually lying. The paper notes that in many cases, robots invent details about 80% of the time when they try to explain their answers.

The Future: Teamwork and New Rules

The paper suggests that the future isn't about one robot doing everything alone. Instead, we are moving toward Agentic Systems—a team of specialized robots working together.

One robot breaks the claim into small pieces.
Another hunts for evidence.
A third checks the math.
A fourth acts as a "debate coach," arguing against the first robot to find errors.

The Bottom Line:
Generative AI is a powerful new tool for fact-checking, capable of doing the heavy lifting that humans can't keep up with. However, it's not a magic wand. It still makes things up, struggles with languages other than English, and needs careful supervision. The goal of this research is to build a system where these robots are reliable, transparent, and inclusive, helping us navigate the ocean of information without getting swept away by the fake waves.

Technical Summary: Generative Large Language Models in Automated Fact-Checking: A Survey

Problem Statement

The rapid proliferation of false and misleading information on digital platforms has overwhelmed manual fact-checking capacities, creating an urgent need for scalable, reliable automation. While traditional automated fact-checking pipelines have focused on accuracy-driven prediction using neural networks and transformers, recent advances in generative Large Language Models (LLMs) have expanded the scope of these systems. However, the integration of generative LLMs introduces new complexities, including hallucinations, limited faithfulness to evidence, and difficulties in reliable evaluation. Existing surveys either predate the generative LLM era, focus primarily on factuality and hallucinations in isolation, or address only specific components of the pipeline without a unified view. There is a lack of comprehensive research synthesizing how generative LLMs are utilized across the entire fact-checking lifecycle—from data creation to prediction and evaluation—particularly in multilingual and low-resource settings.

Methodology

This survey employs a systematic literature review following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology.

Data Collection: The authors searched major repositories (ACL Anthology, ACM Digital Library, CLEF CheckThat! shared tasks) and conducted citation searching.
Selection Criteria: From an initial pool of 6,848 records, 199 research papers were selected for inclusion.
Inclusion/Exclusion: Papers were included if they employed generative LLMs for synthetic data generation, prediction, or LLM-based evaluation within the textual fact-checking pipeline. Studies focusing solely on encoder-only models, non-textual inputs (multimedia), or tasks outside the fact-checking pipeline (e.g., hate speech detection) were excluded.
Analysis Framework: The collected literature was analyzed through a manual inspection process guided by six key dimensions: (1) Tasks and Challenges, (2) Datasets and Limitations, (3) Methodological Paradigms, (4) Evaluation, (5) Multilingual Fact-Checking, and (6) Challenges and Future Directions.

Key Contributions

The survey introduces a unifying taxonomy that categorizes the roles of generative LLMs in automated fact-checking into three primary use cases:

1. Synthetic Data Generation

Recognizing the scarcity and imbalance of annotated fact-checking datasets, the survey details how LLMs are used to:

Augment existing datasets: Through paraphrasing, translation, and style transfer to improve robustness against reworded misinformation.
Create synthetic corpora: Generating entirely new claims, narratives, or fact-check pairs from scratch to address data scarcity in low-resource languages and specialized domains.
Limitations: The authors note risks of distributional mismatch, where synthetic data may be too clean or stylistically homogeneous compared to real-world misinformation, potentially leading to overfitting.

2. Prediction (Downstream Tasks)

The survey analyzes how LLMs perform core fact-checking tasks, distinguishing between structured output (constrained labels, scores, spans) and unstructured output (free-form text).

Core Tasks:
- Claim Detection: Identifying check-worthy statements, normalizing noisy social media posts, and decomposing complex claims into atomic sub-claims.
- Previously Fact-Checked Claim Retrieval: Matching input claims to existing verifications to avoid redundant work.
- Evidence Retrieval: Generating search queries, selecting relevant passages, and ranking evidence.
- Fact Verification & Fake News Detection: Predicting veracity (binary, multi-class, or regression) and generating explanations or full fact-checking articles.
Methodological Paradigms: The review covers prompting strategies (zero-shot, few-shot, Chain-of-Thought, self-consistency), fine-tuning (single-task vs. multi-task), and retrieval-augmented generation (RAG).
AI Pipelines and Agentic Systems: The survey highlights a shift from single-step prompting to multi-stage pipelines (decomposition $\to$ retrieval $\to$ reasoning $\to$ verification) and agentic systems where multiple LLM agents collaborate, debate, or critique to improve reliability.

3. Evaluation

The survey examines the emerging paradigm of "LLM-as-a-judge," where LLMs evaluate the outputs of fact-checking systems.

Rubric-Based Evaluation: Assessing quality, coherence, and persuasiveness against structured criteria.
Factuality-Based Evaluation: Decomposing generated text into atomic claims to verify their truthfulness against external evidence (e.g., FactScore).
Limitations: The authors emphasize that LLM evaluators can suffer from bias, inconsistency, and hallucination, necessitating careful calibration and human oversight.

Multilingual and Cross-Lingual Fact-Checking

A significant portion of the analysis focuses on the disparity in language coverage. While 148 of the reviewed papers focused solely on English, the survey identifies strategies for multilingual settings, including:

Prompting Strategies: English-only prompting vs. target-language prompting.
Fine-Tuning: Multilingual joint fine-tuning vs. language-specific fine-tuning.
Cross-Lingual Transfer: Leveraging high-resource languages to support low-resource ones via translation or shared representations.
Findings: Despite progress, significant performance gaps remain for low-resource and non-Latin script languages, and translation-based pipelines risk semantic drift.

Results and Observations

Task Evolution: The field is shifting from simple classification toward generative formulations that support reasoning, explanation, and article generation.
Data Scarcity: Synthetic data generation has become a critical strategy for addressing the lack of annotated data, particularly in multilingual contexts, though it introduces challenges regarding data quality and bias.
Reliability vs. Flexibility: While agentic systems and multi-stage pipelines offer improved interpretability and robustness through iterative refinement, they introduce complexity and potential error propagation.
Evaluation Gaps: Standard automatic metrics (e.g., BLEU, ROUGE) are insufficient for generative outputs. LLM-based evaluation is gaining traction but requires rigorous validation against human annotations to ensure reliability.
Language Bias: Research remains heavily skewed toward English, with limited systematic coverage of low-resource languages, despite the global nature of misinformation.

Significance

The paper positions itself as a comprehensive synthesis that consolidates fragmented research efforts into a coherent framework. Its significance lies in:

Providing a Unified Taxonomy: It offers the first systematic categorization of generative LLM roles across the entire fact-checking pipeline (data, prediction, evaluation), distinguishing it from prior surveys that focused on traditional pipelines or isolated hallucination issues.
Highlighting Methodological Patterns: It identifies emerging trends such as the move toward agentic systems, the use of synthetic data for low-resource languages, and the adoption of LLMs as evaluators.
Identifying Critical Gaps: The survey explicitly maps limitations, including the "faithfulness" problem (hallucinated reasoning), the scarcity of multilingual benchmarks, and the instability of LLM outputs under prompt variations.
Guiding Future Research: By consolidating current knowledge, the survey aims to support researchers in developing more reliable, interpretable, and inclusive fact-checking systems, outlining necessary directions such as improving evidence grounding, developing robust multilingual benchmarks, and creating evaluation frameworks that account for reasoning quality rather than just output similarity.

The authors conclude that while generative LLMs offer transformative potential for scaling fact-checking, their responsible deployment requires addressing fundamental challenges related to hallucination, bias, and the need for human-in-the-loop oversight.