A Review of Automatic Fake News Detection: From Traditional Methods to Large Language Models

Abstract

In the current digital era, the spread of fake news presents serious difficulties. This study offers a thorough analysis of recent developments in false news automatic detection techniques, from traditional methods to the most recent developed models like large language models. The review identifies four perspectives on automatic detection of fake news that are oriented towards knowledge, style, propagation, and source of the misinformation. This paper describes how automatic detection methods use data science techniques such as deep learning, large language models, and traditional machine learning. In addition to discussing the shortcomings of existing approaches, such as the absence of datasets, this paper emphasizes the multidimensional function of large language models in creating and identifying fake news while underlining the necessity for textual, visual, and audio common analysis, multidisciplinary collaboration, and greater model transparency.

1. Introduction

The massive flow of information from multiple media platforms exposes people to a constant barrage of data in the current information-filled society. The rapid sharing of unconfirmed and frequently inaccurate content, commonly referred to as “fake news”, is facilitated by the accessibility of information and the absence of strong verification procedures [1]. The ease with which such content can be shared and commented on, frequently without verification, increases its reach and influences local and global public opinion and behavior [2].

When fake news affects important domains like public health or political events, the repercussions are very dire. For example, more than half of the voters in the 2016 US presidential election were predicted to be exposed to false information, demonstrating the power of fake news to influence public opinion [3]. Outside the political realm, misinformation has also weakened public health initiatives, particularly vaccination efforts, highlighting the critical necessity for effective methods of detection and counteraction [4].

False or misleading content can take many forms, including misinformation, disinformation, satire, or conspiracy theories. It is more difficult to tell the difference between deception and truth because of these various forms. Fake news is typically defined by researchers as purposefully and verifiably inaccurate material intended to deceive [5]. In addition to text, examples include deepfakes, which make it harder to tell the difference between truth and deceit by using sophisticated media manipulation to depict popular individuals in made-up situations [6].

Social networks are important conduits for the fast and sometimes uncontrolled spread of false information. These platforms frequently achieve a high level of credibility between their users, who then unknowingly disseminate misleading information as a result of ignorance, confirmation bias, or a preference for social media sources over traditional ones [7]. According to different studies, a few bogus postings can quickly garner a lot of attention. For example, during natural disasters, edited photos or spectacular claims may receive thousands of shares and reposts [8,9].

Malicious actors and unsuspecting individuals are the two main user categories involved in the dissemination of fake news, according to certain study groups. In order to influence public opinion, malicious actors purposefully create and disseminate false information, such as automated bots and pseudonymous accounts (sock puppets) [10]. However, unsuspecting people may spread this information without conducting any research, unintentionally expanding its audience. Consequently, these diverse patterns of distribution and motivations must be taken into account by the detection techniques now in use [11].

Despite extensive research, the volume, heterogeneity, and complexity of fake news continue to challenge state-of-the-art detection methods [12]. By systematizing contemporary automatic detection techniques, this study reviews how existing approaches address these challenges and identifies new solution pathways enabled by large language models, particularly through text augmentation and summarization. Automated techniques use deep learning, machine learning, and newly developed large language models (LLMs) to analyze bogus news more thoroughly. In particular, the emergence of LLMs presents fresh possibilities for identifying subtle cues in written and multimodal content. This allows for strategies that combine text, vision, and audio analysis to hold promise for thriving within the complexity of misinformation. However, to improve the dependability and accessibility of these detection methods, more developments in transparency, model robustness, and interdisciplinary cooperation are required.

A substantial body of work surveys automatic fake news detection [11,13,14,15,16]; however, our review shows that existing studies rarely provide a comprehensive synthesis of model usage patterns, particularly the rapid uptake of large language models (LLMs) and other transformer-based architectures. Prior surveys typically focus on a single family of models or on narrow facets, such as propagation dynamics or credibility assessment, without examining the evolving roles of the latest automatic detection methods. In this article, we catalog the full spectrum of automatic approaches and elucidate their internal mechanisms, highlighting problematic trends and current limitations. We conduct an in-depth analysis of recent experimental results and assess the dual role of LLMs within the fake news ecosystem. We conclude with Table 1, which summarizes the datasets currently used for fake news detection, and Table 2, which overviews the latest research. Addressing the strengths and weaknesses of prior work, our systematic survey (i) organizes fake news detection research into four perspectives, (ii) classifies and analyzes recent studies by architectural family, and (iii) discusses their limitations.

2. Materials and Methods

This review and similar research uses the term fake news to denote verifiably false or fabricated claims presented in a news-like format, intended or liable to deceive readers, across text, images, audio, or video.

It is distinct from neighboring categories often encountered during screening: misinformation refers to false claims shared without demonstrated intent to deceive; disinformation refers to false claims shared with intent to deceive, including coordinated campaigns; malinformation describes genuine material deployed in a misleading context that substantially alters meaning; satire or parody comprises humorous fabrications that do not purport to report facts and are excluded unless shown to be consumed as factual; conspiracy narratives and rumors may be considered only when they include concrete, checkable assertions. In practical terms, an item is treated as fake when it contradicts high-confidence evidence available at the time of the claim, when it fabricates entities, events, quotations, or data, when authentic material is placed in a false or materially distorting context, or when synthetic media depicts events or attributes that did not occur while being presented as authentic reporting.

Signals from propagation patterns and source credibility may corroborate these determinations but are not sufficient on their own to establish falsity. Stylistic cues such as sensational tone, excessive punctuation, or typography function as features for modeling rather than determinants of ground truth. Ambiguous or presently uncheckable assertions are labeled unverified and excluded from counts of fake items.

This operationalization aligns the selection of datasets and the interpretation of model outputs with a fact- and context-first standard while still acknowledging the supporting role of source and propagation signals.

For the current review, a structured literature review of automatic fake news detection methods was conducted. The review synthesizes model families and usage patterns across four analytical perspectives: knowledge/fact-checking, content/style, source reliability, and propagation, followed by a taxonomy of computational approaches (from traditional ML to deep learning, transformers, and large language models). The initial search was performed in Google Scholar, chosen because it indexes the main scientific portals relevant to this field, including MDPI, IEEE Xplore, Elsevier, and Springer Link, as well as smaller or lesser-known repositories that may have useful insights.

The search period was 2022–2025, ensuring coverage of the latest surge of transformer-based and LLM-based approaches while retaining sentinel pre-2022 works where necessary. Boolean strings combined task, method, and modality terms (e.g., “fake news” OR misinformation OR rumor AND detection AND (BERT OR transformer OR LLM OR CNN)).

The search yielded 432 initial records. Studies were included if they addressed automatic fake or misinformation detection, presented a concrete automatic method, reported empirical evaluation on a recognized dataset, and were published in English.

The exclusions consisted of purely theoretical works, studies without an automatic detection component, duplicates, or tasks unrelated to news credibility like spam or phishing.

As depicted in Figure 1, the screening proceeded with the following two phases. The title/abstract stage consisted of 432 initial records, from which 281 were excluded for irrelevance. In the full-text stage, of the 151 assessed, 107 were excluded for different reasons. The final corpus comprised 44 primary studies included for synthesis.

To ensure representativeness while avoiding redundancy, for each method type (e.g., SVM-based, CNN-based, transformer-based), the latest and most relevant 2–3 examples were retained. Newer studies were prioritized, reflecting the review’s emphasis on state-of-the-art progress.

For LLM-based methods, we deliberately included papers covering different application strategies (prompting, fine-tuning, retrieval-augmented approaches). We also ensured that multilingual contributions were represented, covering detection beyond English (e.g., Romanian, Chinese, and cross-lingual settings).

3. Perspectives on the Automatic Detection of Fake Content

Researchers must gain a thorough grasp of fake news’s several forms before attempting to identify it. Therefore, an important first step is to investigate the causes and motivations behind the spread of fake news, including malevolent actors, monetary gain, and algorithm-driven social media. Additionally, it helps increase awareness of the negative effects of fake news on society, such as the erosion of media credibility, the disruption of political debate, or even the inciting of violence [1,17]. Researchers can only start looking into ways to detect disinformation after they have a thorough understanding of how it operates.

When examining fake news, researchers have taken different tasks into account. While some study the methods used to create false news [13,18], others look into the degree of clarity in news content [19,20]. Conversely, some research evaluates source dependability [15,21], while others concentrate directly on the ways via which fabricated content is spread [14,22].

The detection of false news is currently approached from four main viewpoints, as illustrated in Figure 2: knowledge-based approaches, content-based approaches, source evaluation approaches, and content propagation approaches. The latter two viewpoints look at fake news after it has been spread, while the first two analyze it while it is being created.

3.1. Knowledge Approaches

Past research has frequently examined the stylistic components of written content in order to detect writing patterns that point to deception. This is conducted after the primary data has been verified through fact-checking. Knowledge-based approaches, which are also known as fact-checking approaches, entail the collection of data from a variety of sources, often in an open format, in order to establish datasets or similar knowledge repositories. The methodologies used for this process are computationally oriented (automatic), human expert-oriented, or crowd-sourced [16]. Fake news datasets have shown issues such as redundancy, low validity, imbalance, or incompleteness. For these reasons, the data that has been obtained must be further processed by matching, time-marking, checking for uniformity within the dataset, completeness assessment, and reliability check in order to render the datasets usable. Time-stamping evaluates the temporal validity of facts, while data matching establishes connections between records of related or similar facts. The inclusion of all pertinent information is verified through exhaustiveness assessment; uniformity analysis confirms the factual consistency; and the reliability assessment validates the data veracity. The successful implementation of fact-checking approaches is facilitated by the appropriate refinement of the data through the execution of these five tests [11].

Figure 2. Perspectives on the automatic detection of fake content. Four primary analytical stances are depicted in the diagram according to stage: the “Content Fabrication stage,” which includes stylistic and fact-checking analyses, and the “Dissemination Stage”, which includes assessments of source credibility and propagation dynamics.

Figure 2. Perspectives on the automatic detection of fake content. Four primary analytical stances are depicted in the diagram according to stage: the “Content Fabrication stage,” which includes stylistic and fact-checking analyses, and the “Dissemination Stage”, which includes assessments of source credibility and propagation dynamics.

3.2. Content Style Analysis Approaches

Researchers employ these methodologies to assess whether the actual content is intended to deceive the reader. This primarily involves the identification and preservation of the written style of the journal/media content by classifying it based on quantifiable characteristics or features. A variety of methods are employed to categorize the news [23]. The data that is stylistically analyzed can be both visual or textual. To give some examples, the textual pattern of a false written media can include visionary rhetoric, a varied content or sources included to appear credible, emotional charged language, and a relaxed tone, while, in contrast, fake images frequently exhibit a high level of lucidity and coherence, despite the fact that they lack diversity [24].

3.3. Source Reliability Measuring

These methodologies evaluate and investigate the credibility of sources by evaluating the credibility of organizations and individuals that generate and distribute journalistic content. The information and broader social landscape that encumbers the news can be the focus of this evaluation [25]. Analyzing the responsibilities of news creators and publishers is an effective method for determining the credibility of a source. Malicious users on social platforms have the ability to create and circulate news stories that are widely disseminated, while ordinary users may share fake news without further checking for its accuracy [26].

3.4. Content Propagation Studying

This method is employed by researchers to examine the manner in which information is disseminated across various media platforms, with the expectation that news that originates from anomalies will be inaccurate. Propagation-based research investigates the methods by which users are involved in the dissemination of disinformation [27]. Graphs or cascades of news may serve as the input. Graphs give a visual way to understand more about how information moves and spreads, while the propagation is explicitly displayed for the content input. The false news source is the origin of the news cascade, a tree branch layout that is made up of a variety of nodes that were created by users to further disseminate the media [28]. The graph models the structure of a network through which fake news is propagated [29]. The structure may be homogeneous, featuring only one type of node/edge; heterogeneous, incorporating various types of nodes and connections; or hierarchical, where nodes are arranged in layered levels with the root node acting as the origin point of the fake media.

4. Computational Approaches to the Automatic Detection of Fake News

Despite the existence of a variety of techniques for distinguishing between true and deceptive news, none of them is capable of completely distinguishing between the two due to their respective limitations. These challenges involve managing large datasets, dealing with rapidly changing and diverse information, and the insufficient study of multimodal data. The generalizability of fake news detection models is significantly limited by the absence of comprehensive datasets tailored to accommodate a wide range of contexts and languages. As a result, researchers have classified false news detection methods into distinct approaches.

The methods for detecting false news can be broadly classified into pre-deep learning methods, deep learning methods, and large language models (LLMs), as illustrated in Figure 3. This paper will provide a detailed breakdown of these methods.

4.1. Pre-Deep Learning Techniques

These methods, both supervised and unsupervised, are also referred to as traditional machine learning methods and are widely used to address the issue of detecting false news. Research teams still employ, with encouraging results, fundamental machine learning algorithms for the purpose of identifying fake news as of today, frequently making modifications to enhance their efficacy in this regard.

4.1.1. Support Vector Machines (SVMs)

These models simply classify fake news according to the article’s media source, stylistic approach, and secondary data. The core process involves entering the text, extracting the relevant data, and then classifying it as either real or fake news. SVMs can handle big datasets fast, are accurate, and can make predictions in real time. Using a COVID-19 dataset with 1,375,592 tweets, a 2024 study assessed the detection of fake news on social media using a number of machine learning methods (Naïve Bayes, Logistic Regression, SVM, Decision Tree, Random Forest, K-Nearest Neighbor) and deep learning models (CNN, LSTM). Tokenization, TF-IDF feature extraction, and normalization were all part of the data pretreatment process. The algorithms that produced the lowest accuracy (65%) were LSTM architecture, while the SVM attained the highest score (98%), and Logistic Regression (LR) had a 2.8% lower metric (95.2%) [30].

Another noteworthy method used the ISOT Fake News dataset from Kaggle and an SVM classifier. Dataset preprocessing and data division into 80%/20%, training/testing sets were part of the experimental setup. For classification, the process used a linear SVM kernel. The findings showed a 99% overall accuracy rate, properly classifying 4264 true and 4542 fraudulent news articles with only 20 false positives and 26 false negatives [31].

By highlighting the shortcomings of conventional Naïve Bayes, particularly the presumed conditional independence between features, which frequently is not reflected in real-world textual media, a hybrid approach combining Support Vector Machine architecture and K-Nearest Neighbor (KNN) improved upon the understanding of nuanced fake news classification. The experimental process included text preprocessing, reduced Singular Value Decomposition (SVD) feature extraction, and training of the classifier using a BuzzFeed News dataset. The improved Naïve Bayes model outperformed the Random Forest (80%), normal Naïve Bayes (69%), and Passive Aggressive (87%) classifiers with an accuracy of 99% [32].

4.1.2. Naïve Bayes (NB)

The foundation of this approach is the Bayes theorem, which determines the likelihood that a hypothesis is correct in light of available data. In the field of detecting fake news, NB can analyze big datasets and find patterns that point to the existence of false information. It can, for example, identify words and phrases that are frequently associated with false news items and assess the writing style of an article by contrasting it with other sources. It can also determine an article’s likelihood of being real or false by analyzing its general tone.

Using an improved hybrid Naïve Bayes model in conjunction with shortened Singular Value Decomposition (SVD) for fake news identification, this new model variation achieved 99% accuracy on a dataset compiled from the BuzzFeed media platform. Text preprocessing, attribute selection using the SVD dimensionality reduction, and classifier training were all part of the experiments. The accuracy of the novel variation was much higher than that of typical classifiers, such as Random Forest (RF) classifiers (80%), base Naïve Bayes models (69%), or the Passive Aggressive algorithm (87%), when conducted in a supervised learning environment [33].

A related study used the Naïve Bayes Multinomial approach to identify fake or unsourced news in online media. Text was vectorized using CountVectorizer as part of the experimental process to create numeric vectors. Using a split of 80% training and 20% testing data, the model’s accuracy was 94.73%. The results of the confusion matrix, notated in true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), were TN/4555, FP/178, FN/295, and TP/3952, indicating a high reliability of the model in real-world situations [34].

4.1.3. Logistic Regression (LR)

LR is a statistical approach that is effective in identifying false news by assessing various characteristics of news articles. It examines a variety of factors, including source, title, news body, and any other attributes that may suggest that the news is misleading. The likelihood of a true or false media is estimated by the algorithm through this analysis.

One of the primary advantages of LR is the model’s ability to generate precise results, irrespective of the dataset’s dimensions. This is particularly advantageous for extensive compilations of news articles.

A research team achieved a 97.92% accuracy (Acc) score in the prediction of false job postings by employing an optimized LR algorithm in comparison to traditional Linear Regression. The experimental setup comprised data preprocessing, attribute selection, and classifier training, all of which were conducted using a dataset of job postings [35].

The accuracy scores of 95.09% and 95.62% were achieved by a procedure using Support Vector Machines, Logistic Regression classifiers, and Long-Short-Term Memory (LSTM) models through Stacking and Delegation. The experimental procedure entailed the evaluation of the detection efficiency by analyzing preprocessing methods like count vectorization or Term Frequency–Inverse Document Frequency (TF-IDF). The results indicated that ensemble methods, in particular probability-based stacking, obtained good scores for the area under the curve (AUC) metric (0.9394, 0.9509). Similarly, the use of delegation strategies demonstrated satisfactory performance, with repeated delegation achieving AUC scores as high as 0.9477 and 0.9280. According to the present review, the presented method has achieved the second highest precision score in fake news classification, surpassed only by SVMs [36].

4.1.4. Decision Tree (DT)

DTs are established algorithms able to promptly and accurately process large datasets, making them a great choice for the classification of fake news. They classify the source, style, primary and contextual data, and identify connections between all these data points. Consequently, DTs can be of assistance in identifying the individuals or groups that are responsible for the dissemination of false news.

In this 2022 experiment, DT methods typically outperform SVMs, as evidenced by the accuracy and precision intervals between 90% and 97% obtained over the course of approximately 10 iterations. This proposed model outperformed an SVM that obtained an accuracy score of 91.5% [37]. Nevertheless, other studies conducted in 2024 [38] assert that the scores are higher. For instance, a study that utilized a dataset from the 2016 US election processed from the Kaggle media outlet compared Logistic Regression, Decision Tree, and Random Forest models in a Python-based supervised learning environment, involving dataset preprocessing, feature selection, training, and validation, and demonstrated that the Decision Tree yielded superior results. The DT model outperformed the Random Forest classifier (99.23% Acc) and Logistic Regression (98.80% Acc), achieving the highest accuracy score of 99.64%.

A somewhat distinct study suggested a method for detecting false news that is based on Decision Trees and incorporates a combined optimization algorithm entitled African Vultures Optimization-Aquila Optimization (IBAVO-AO) combined with a DT classifier (XGBoost). Global Vectors for Word Representation (GloVe) embeddings and Relief feature selection were employed to preprocess more than 44,000 ISOT news articles in experimental contexts. The hybrid algorithm improves the accuracy of classification by optimizing the selection of pertinent features. The experimental results iterated that the IBAVO-AO method outperformed the existing methods, attaining an accuracy of over 92.5% [39].

4.1.5. Random Forest Classifiers (RF)

RF is a classifier that relies on the majority vote for its predictions and constructs a multitude of DTs during its training process. This algorithm examines the news content, stylistic cues, writing tone, and the media source for fake news classification. It evaluates all of these characteristics in conjunction to ascertain the probability that an article is fraudulent. The RF algorithm is also capable of identifying biases in the data, exhibits a reduced sensitivity to overfitting, and performs well on new data.

In 2023, a modified Random Forest classifier obtained a 99.32% accuracy score using a 2016 Kaggle dataset [40].

Another research team that used RF in their experiments proposed a three-stage method. The first, the data preprocessing stage, included stop-word removal, stemming, and tokenization. The second stage, for feature selection, was optimized using the Honey Badger (HB) algorithm. The third stage used a lightweight convolutional random forest (LCRF) classifier. The dataset used by the team consisted of news articles concerning COVID-19. The LCRF-HB model obtained 98.7% accuracy, a 98.3% precision score, 97.6% recall, and 95.4% specificity [41].

4.1.6. Bayesian Modeling (BM)

BM employs the probability theory to work with data and has proven to be a highly effective method for fake news identification based on its content.

For instance, if a piece of news contains inaccurate data and includes specific terms and phrases frequently used or associated with deception, Bayesian Modeling is capable of identifying these patterns in order to develop a classification model for stories of this nature. Bayesian methods are distinguished from other modeling techniques, such as deep learning algorithms, by the utilization of “Bayesian Modeling priors”, which allows research teams to incorporate prior knowledge about the likelihood of certain news being true or false.

An Indonesian team experimented with a rapid fake news detection system using a BM variant called Complements Naïve Bayes (CNB). It was purposed for COVID-19-related media content that was disseminated over social networks. The researchers implemented four models: CNB, Multinomial Naïve Bayes (MNB), DTs, and Gradient Boosted Decision Trees (GDBT). The experimental setup consisted of a dataset containing a total of 10.700 tweets, preprocessed by tokenization, using the Term Frequency–Inverse Document Frequency (TF-IDF) and Discrete Fourier Transform (DFT) methods, class balancing by the Synthetic Minority Over-sampling Technique (SMOTE), and Grid Search with Cross-Validation for hyperparameter optimization. The CNB and MNB models demonstrated the highest performance, with approximately 92% scores for Acc, Prec, recall, and F1. The best scores were achieved by the CNB model, which also had the shortest runtime (0.55 s) [42].

4.2. Deep Learning Models (DL)

Models from this category are used in fake news detection and similar research due to their ability to automatically acquire layered feature representations from raw data. Fake news classification has shown significant potential in research on neural networks, particularly those that are specifically designed for natural language processing (NLP) [43]. As shown in Figure 2, fake news creation, dissemination, its challenges, and other various perspectives have been the subject of research for a considerable time. Nevertheless, the lack of a generally acknowledged definition of what “fake news” is poses substantial obstacles for the research conducted in this field, especially for dataset uniformity.

This conceptual problem causes a lack of datasets that satisfy all the criteria. A possible solution may be found in the use of deep learning techniques, particularly those that are based on neural networks, as these have been especially effective in the management of mixes and unstructured data. The capacity for acquiring hierarchical representations allows the model to recognize intricate patterns inside written content, thus simplifying the process of distinguishing between genuine and fake news.

In the detection of false news, DL models such as Recurrent Neural Networks (RNNs) or Convolutional Neural Networks (CNNs) have been effectively employed by capturing contextual nuances and linguistic features. These models provide significant adaptability, enabling the fine-tuning of performance across datasets with different structures or label systems. This adaptability is especially useful because of the persistent absence of dataset standardization. Deep learning models systematically achieve superior performances on different datasets by employing techniques such as domain adaptation and transfer learning, which effectively mitigate some challenges presented by dataset inconsistency [44].

4.2.1. Recurrent Neural Networks (RNNs)

This subclass is particularly good for managing sequential data, rendering them the optimal choice for the analysis of text, audio, and vision content for the detection of false news. Recurrent Neural Networks are able to identify patterns and relationships that indicate deception, such as specific word combinations and phrasing, by retaining a memory of previous inputs. Furthermore, they are capable of identifying discrepancies between a variety of sources.

In 2024, a method was proposed to improve the accuracy of false news detection by combining GloVe word embeddings with an RNN version called a Bidirectional Long Short-Term Memory (BiLSTM) neural network. The experimental design entailed text preprocessing with GloVe to extract word relations, followed by data classification using deep learning models like RNN, CNN, and DNN. The combination between RNN and GloVe preprocessing obtained the best accuracy (98.974%) when using the Curpos fake news dataset, surpassing other models in the detection of false news [45].

4.2.2. Long-Short Term Memory Models (LSTMs)

LSTM networks are a particular type of RNN intended to resolve the vanishing gradient issue and learn long-term dependencies. They are especially efficient in recognizing patterns within extended sequences, which recommends them for analyzing longer articles. Furthermore, LSTMs are able to analyze language usage and writing patterns.

The most recent research from 2024 remains somewhat underwhelming for fake news classification. In this recent study, an LSTM model was assessed in comparison to an SVM Classifier, LR, and Multinomial Naïve Bayes methods using a Kaggle-sourced fake news dataset. The LSTM obtained an accuracy score of 94% [46].

Nevertheless, a hybrid RNN-LSTM network proposed in 2023 demonstrates that this approach still has potential. The researchers proposed a system for detecting false news that employs a combined feature extraction technique (Term Frequency–Inverse Document Frequency, N-grams, Part-of-Speech Tagging, etc.) in conjunction with a combined Recurrent Neural Network and a Long Short-Term Memory one (RNN-LSTM). The experimental environment consisted of a standard CPU configuration with a Java-based implementation. The team used the widely used LIAR fake news dataset, which was preprocessed and divided into a 70% quota for the training data and 30% for testing. The RNN-LSTM network was employed to perform classification, and it was compared to models such as Support Vector Machines, Artificial Neural Networks, Random Forests, Naïve Bayes classifiers, Decision Trees, and Deep Neural Networks. The RNN-LSTM version with a Rectified Linear Unit (ReLU) activation attained the highest accuracy score (99.10%), particularly under a 15-fold cross-validation split [47].

4.2.3. Bidirectional Long Short-Term Memory (Bi-LSTM)

These networks have the ability to concurrently capture context from both past and future states by analyzing data in both forward and backward directions. This bidirectional analysis improves the network’s ability to comprehend contextual information and recognize patterns associated with fake news.

A recent experiment has employed SA-BiLSTM, a Self-Attention Bidirectional Long Short-Term Memory architecture that was trained and tested using the ISOT Fake News dataset and contains more than 44,000 labeled news articles. The method demonstrates superior outcomes in comparison to conventional LSTM models. Created in 2024, the experimental environment comprised a CPU–GPU configuration that included Python, Keras, and TensorFlow. The procedure was optimized with ADAM and SGD and consisted of data processing, word embedding and a 70/30 training/testing quota. The obtained results (99.98% Acc, 99.96% F1, and 99.98% AUC) showed that the SA-BiLSTM model outperformed similar models like Convolutional Neural Networks, Gated Recurrent Unit (GRU), LSTM, and similar baselines [48].

4.2.4. Convolutional Neural Networks (CNNs)

This class of network was initially developed for image analysis but has been repurposed for text processing in various tasks. For fake news detection, CNNs are able to recognize local features and patterns within texts (such as N-grams and phrases) that indicate deception. These models have the capacity to effectively manage large datasets, which renders them especially advantageous for the analysis of the extensive information associated with false news.

A model that utilizes word embeddings to identify COVID-19-related false news has been introduced in recent contributions to the field of fake news detection using CNN. The optimal CNN architecture was identified through hyperparameter optimization using grid search in the experimental setup. The model’s performance was assessed in comparison to a variety of machine learning algorithms. The proposed CNN model outperformed other methods in the detection of COVID-19 false news, achieving the following mean scores: 96.19% accuracy, 95% F1, and 0.985 AUC [49].

A similar, hybrid CNN–Bi-LSTM–SelfAttention model achieved an accuracy of 98.71% over a COVID-19 dataset, outperforming its base models. The hybrid method demonstrated superior metrics like accuracy, recall, F1, and loss by a 1% minimum over the other tested models. The experimental settings used the TensorFlow and Keras libraries in Python and implemented data preprocessing, feature extraction using a CNN, sequence learning using a Bi-LSTM, and highlighting relevant features through an Attention Mechanism (AM) [50].

4.2.5. Graph-Based Models

Custom models for NLP like the Propagation Graph Neural Networks (PGNNs) [51] concentrate on the dissemination of false news within social networks, utilizing graph structures to model and identify anomalies within the media dissemination. This group of models has also been effectively implemented in multimodal datasets (vision and textual).

For instance, a 2024 study suggested the development of a multimodal false news detection system that integrates Text Graph Convolutional Networks (TextGCNs) and Vision Transformers (ViTs), utilizing both textual and visual features from the Fakeddit dataset. TensorFlow, Scikit-learn, and HuggingFace tools were implemented in the experimental environment on a system that featured an Intel i7 processor, 16 GB of RAM, and an RTX 3060 GPU. Textual input was transformed into graphs using Term Frequency–Inverse Document Frequency (TF-IDF) and Pointwise Mutual Information (PMI), while images were encoded using a pretrained Vision Transformer (ViT). Random Kitchen Sink (RKS) mapping was employed to perform feature fusion, and an Artificial Neural Network (ANN), an RF, and a Support Vector Machine were employed to classify the data. The results showed that the ANN outperformed the similar classifiers by attaining an accuracy of 94.17% for a binary classification, 90.14% accuracy on the three-class experiment, a 75.91% accuracy score for the six-class experiment, and 0.98 for precision and recall. This study illustrates the superior performance for multimodal systems in comparison to unimodal approaches and emphasizes the need for future research in the areas of class imbalance and the application of explainable AI techniques [52].

4.2.6. Transformer-Based Models

These represent a subset of the deep learning model architecture specifically designed to manage sequential data, with a particular emphasis on natural language processing (NLP) tasks. The Transformer architecture’s use of the attention mechanism to process or generate data has been widely used in research. This subclass was introduced by the foundational study “Attention is All You Need” [53]. A frequently employed family of models for the classification of false news is constructed on the Bidirectional Encoder Representations from Transformers (BERT) architecture [54]. In contrast to a binary true/fake approach, recent advancements using BERT variants involve more sophisticated exploration of nuanced multiclass datasets, as this architecture consistently reports [55] high metrics on established fake news datasets.

As an illustration, 2024 research investigated the detection of multiclass fake news on the CheckThat-2022 dataset by employing transformer-based models, mBERT, SBERT, and XLM-RoBERTa, to categorize news as true/false/partially false/other categories. Using an Azure framework, the experiments were conducted to address class imbalance by training with original and Chat Generative Pre-trained Transformer (ChatGPT)-augmented data. The mBERT model demonstrated the highest level of performance (accuracy: 34%; F1-score: 0.23) among the results. The macro F1-score of 0.26 was achieved when ChatGPT-generated data was used to enhance classification, particularly for underrepresented classes [56].

An additional recent study suggested the implementation of a multiclass fake news detection framework on the LIAR dataset. This framework consists of an ensemble of models, including BERT, Robustly Optimized BERT Approach (RoBERTa), Bi-LSTM, and Light Gradient Boosting Machine (LightGBM). Contextual features were extracted using BERT and RoBERTa in the experimental configuration; sequential dependencies were captured using Bi-LSTM; text and numerical features were integrated using LightGBM; and the training and evaluation used k-fold cross-validation [57]. The proposed ensemble model outperformed individual models, achieving the following metrics: 41% Acc and 0.42 F1 as the highest performance.

Another novel perspective is implemented on three datasets: Twitter15 (containing 1490 claims), Twitter16 (containing 818 claims), both labeled multilabeled in four classes: true (T), fake (F), nonfake (NF), unverified (U), and PHEME, labeled using 9 tags and containing a total of 2402 claims. This approach incorporates a Graph-based GCN and BERT, which are combined with Co-Attention as a GBCA model. The proposed model exhibits an enhancement over the individual models for analyzing unstructured multilabel data and considers user interactions [58].

A different research path that employs transformer architecture involves the application of a combination of BERT and a Visual Geometry Group Network variant (VGG-19) to multimodal datasets (text and image). By incorporating transformer and DL techniques on text and vision data, the study proposed a dual phase framework for the detection of fake news. The ISOT dataset was employed in textual experiments, which employed RF, SVM, and LR classifiers. The Random Forest classifier achieved 99% across all metrics (accuracy, precision, recall, and F1). The multimodal approach used the MediaEval 2016 dataset for evaluation. The approach incorporated BERT for text classification and ResNet for vision feature extraction plus an attention tool. The proposed model demonstrated robust detection across diverse data types, surpassing existing baselines such as SpotFake, with an F1-score of 0.892 and 94.4% accuracy [59]. These results demonstrate a 3.1% improvement in accuracy.

In a comparable investigation, pretrained BERT and VGG-19 models were locally fine-tuned. BERT was completely retrained to better adapt its language representations, and VGG-19 underwent similar modifications, which included the development of a redesigned classifier and the inclusion of a global average pooling layer. This method obtained a high accuracy rate of 92% in experiments conducted on the Fakeddit binary dataset [60].

The Ensemble Learning Deep model for Fake News (ELD-FN), used for multimodal fake content detection with the Fakeddit dataset, which contains more than 1 million samples, exhibits additional encouraging results in nuanced (multiclass) classification. The experimental setup comprises NLP preprocessing (tokenization, lemmatization, and sentiment analysis), feature extraction using Visual BERT for textual data and vision, and an ensemble of bagged and boosted CNN and LSTM models. The model validation consisted of a 10-fold cross-validation technique. ELD-FN outperformed the baseline models (FakeNED, MultiFND) with 88.83% accuracy, 93.54% precision, 90.29% recall, and 91.89% F1. The results also emphasize the beneficial effects of sentiment analysis, preprocessing, and oversampling [61].

Additionally, as a recent development, the transformer models have been optimized for multilingual applications. In 2024, a newly constructed dataset, TR_FaRe_News, was employed to present a comprehensive framework for the detection of false news in Turkish. This dataset comprises 18.695 manually labeled tweets. The experimental setup involved the preprocessing of data using the Zemberek NLP tool and classification using traditional machine learning classifiers (LR, NB, RF, Voting Classifier, SVM) and deep learning classifiers (BERTurk, DistillBERTurk). In addition to TR_FaRe_News, other datasets were used for the experiments (Twitter15, Twitter16, ISOT, GPT-2, LIAR, etc.). The base classifiers were outperformed by the BERTurk+CNN model, which obtained a 94% accuracy score. The model exhibits robust dataset generalization potential for the BERT architecture [62].

In the same spirit, a large-scale empirical study has been initiated to examine the factors that influence cross-lingual zero-shot transfer inside multilanguage architectures such as BERT [63]. This study will further impact research involving transformer-based architecture. Additionally, such studies will contribute to adapting models that were trained on high-resource languages in order to operate on low-resource languages by utilizing external dictionaries for tokenizer adaptation, for example, on Silesian and Kashubian languages [64].

4.3. Techniques Relying on Large Language Models (LLMs)

The generation and detection of fake profiles and false content have been significantly influenced by recent advancements in LLMs. Their ability to comprehend and generate text that resembles human writing has been used to enhance the effectiveness of identification methods and to generate complex, deceptive content. The integration of LLMs has considerably improved the methods used to identify fake news.

Supplementary data, including the timestamp, authorship, publication sources, or subject matters, is typically required by traditional and deep learning techniques in addition to the article’s text. Nevertheless, these methodologies are constrained by practical constraints, as secondary data is not always readily accessible. Furthermore, the utilization of online fact-checking outlets for dataset gathering can be tedious, which can impede the capacity for keeping up with the rapid production of fake media [65].

Recent research has shown that fine-tuned LLMs, like LLaMA 3 and GPT-4, can successfully identify false news without the need for large amounts of additional data while employing a variety of approaches [13].

4.3.1. Data Generation

In order to surmount existing constraints, novel methodologies have been developed that integrate a blend of genuine news and accurate data with intentionally deceptive content generated by human operators. MegaFake is the name of a vast-scale, theorized dataset of fake news that was produced by LLMs in a single study. The LLM-false Theory framework, which is rooted in social psychology, was employed to generate this dataset. MegaFake and GossipCop were used to experiment with six natural language understanding (NLU) models like BERT or RoBERTa and eight natural language generation (NLG) models like GPT-4 which were trained and tested on the two datasets. The resulting metrics show that the NLU models outperformed the NLG models by a significant margin. The CT-BERT model achieved the highest metrics (92.28% Acc, 0.9459 F1) [66].

Some examples of fake data generation include instructing LLMs to generate fictitious articles that are based on human-generated summaries of false events. Three scenarios were developed by a research team that employed this methodology: Human Legacy (human written news), Transitional Coexistence (combination of news generated by humans and machines), and Machine Dominance (data dominated by generated fakes). The team also employed human and generated true/false data as part of their analysis. The datasets utilized were PolitiFact++ and GossipCop++, with the addition of LLM-generated and paraphrased samples generated using ChatGPT. Various proportions of machine-generated content have been observed to influence performance in the experiments that were conducted. The results indicate that detectors trained on human written news are more generalizable, whereas those that are trained predominantly on machine generated content become biased. For further scenarios, the recommendation for training balance is prioritized [67].

Other research cases were designed to enhance the credibility of false news articles by utilizing LLMs. Med-MMHL, a multimodal dataset, was presented in a 2023 study. It is intended to identify medical misinformation that is generated by both humans and LLMs in 15 different diseases. The data, which was gathered from 2017 to 2023, comprises news articles, tweets, allegations, and corresponding images. The procedures included the creation of five comparison tasks (oriented on documents, sentences, tweets, and multimodal), structured data crawling, and adversarial false news generation based on ChatGPT. Some of the assessed models were BERT, VisualBERT, BioBERT, FN-BERT, and CLIP. The results outlined that FN-BERT achieved the best metrics (95.78% accuracy and F1 of 95.76%) in textual situations, while CLIP surpassed the other models in multimodal situations. The work emphasizes the persistent challenge of detecting false sentences generated by LLMs, which necessitates future methodological advancements [68].

Deceptive content can be produced by integrating true articles with false events, among other innovative methods. One study employs prompt engineering techniques to generate three datasets: Dgpt std (Standard Prompting Dataset), Dgpt mix (Mixed Prompting Dataset), and Dgpt cot (Chain-of-Thought Prompting Dataset). A standard query technique is utilized for instructing ChatGPT to recompose original human-created news to fakes. Next, a mixed prompt instructs ChatGPT to rewrite articles by mixing true with fake data, which makes detection more challenging. Finally, the chain-of-thought prompting dataset is generated by pairing generated fake news articles with original ones, resulting in pairs of fake and real matching articles used in training and evaluation. A RoBERTa model that was refined using human-written news demonstrated satisfactory performance on Dgpt std false content (1.2% misclassifications). However, it encountered difficulties when confronted with more complex Dgpt mix (15.4%) and Dgpt cot (77.9%) fake news [69].

A similar study was conducted to develop AdStyle, an adversarial-style augmentation method that is designed to improve the robustness of false news detection resistant to style-conversion aggression. The research employed LLM dataset augmentation through GPT-3.5 Turbo, BERT architecture classifiers, and benchmarking using PolitiFact, GossipCop, and Constraint datasets. The primary development of this method is the automated generation/selection of prompts utilized to rephrase real or fake news in a variety of ways, such as adding emotional exaggeration, a humorous tone, a poetic form, or a sarcastic style, while maintaining the semantics of the content. The selection of these prompts is determined by their diversity, coherence, and adversarial nature. The results indicate that AdStyle obtained an AUC of 0.9646 in attack settings and 0.9460 in clear settings [70].

Additional research evaluates the influence of fake information generated by LLMs on Open-Domain Question Answering (ODQA) systems. It creates four scenarios for the generation of misinformation: GENREAD, CTRLGEN, REVISE, and REIT and uses GPT-3.5 to evaluate their effect over the ODQA system. In GENREAD, GPT-3.5 directly generates false text in response to inquiries. CTRLGEN employs controlled generation by using specific false assertions to prompt GPT-3.5. REVISE modifies passages that contain ground-truth information to introduce fakes while maintaining the same context. Reinsertion and Iterative Training (REIT) encompasses numerous cycles of retraining and fake data injection to replicate persistent disinformation scenarios. The effects of this generated data are assessed using Best Matching 25 (BM25) and Dense Passage Retriever (DPR) retrievers combined with Fusion-in-Decoder (FiD) and GPT-3.5 as readers, and they were administered into QA corpora dataset (Natural Questions Subset NQ-1500 and CovidNews). The results indicate a decline in performance of up to 87%, which underscores the susceptibility of the Open-Domain Question Answering systems to disinformation [71]. The large-scale automatic generation of fake or misleading articles is effectively curbed by such approaches, particularly those that focus on manually created false content. Nevertheless, content that is devoid of substance is often the consequence of methods that rely on fabricated summaries, and the maintenance of contextual coherence can be compromised by modifications to specific events or components.

In 2024, a study was conducted to improve the detection of false news in multiple languages by enhancing datasets with LLaMa 3, a large language model (LLM). TALLIP and MM-COVID, two multilingual datasets, were used to evaluate BERT-based classifiers in the experimental environment. Following the experimental procedure, all samples were translated into English, then LLaMa 3 was used to generate news samples using prompts to paraphrase existing news while implementing a variety of augmentation strategies, including augmentation rates, random or similar samples, and class specific augmentation (only fake news, only real news, or binary). The “only fake” strategy at one rate obtained a 7.7% increase in F1-scores for English language, while a 4.4% increase was obtained for Hindi. Nevertheless, a significant concern in this experiment is that the complete nuances of the narrative and factual structure of human created news may not be preserved by fabricated summaries or paraphrased content [72].

A distinct Algerian team studied the efficacy of translation-obtained text augmentation in false content detection, with an emphasis on the Algerian dialect. BLEU, Chrf++, COMET, and expert human judgments were employed to evaluate GPT-4 in the experimental environment. The transformer architectures AraBERTv02, MARBERTv2, and DziriBERT were fine-tuned using a translated Modern Standard Arabic (MSA) Khouja fake news dataset. Various augmentation quantities and manual versus automatic translation were evaluated in experiments. The results indicated that automatic translation improved recall, but it also reduced precision as a result of noise (irrelevant, incorrect, or misleading information introduced during the automatic translation process). The AraBERTv02 model obtained the highest metrics (0.67 F1). This study reaffirmed the drawbacks of augmentation in dialect-sensitive languages that can obscure subtleties, misunderstand, or warp rhetorical devices, colloquial idioms, or cultural significance found inside the source material [73].

4.3.2. Text Classification

LLMs are refined on labeled datasets that contain both authentic and fraudulent news, in a manner similar to previous models. This method capitalizes on the language comprehension capabilities of LLMs to detect indicators of disinformation. In past years, there has been a growing emphasis on the fine-tuning of pretrained architectures using datasets that contain both real and false news in order to enhance the accuracy of misinformation recognition. Utilizing the robust capabilities of PLMs, the classification approach is specifically designed to identify disinformation, whether it is within multiple categories or between true and false classifications. This results in significant improvements in performance metrics.

Upon analyzing the 2024 style-agnostic model SheepDog, intended to resist against adversarial stylistic aggression generated by large language models, we observe that so-called traditional detectors experience a performance loss of up to 38% while in a stylistic attack setting. In contrast, SheepDog significantly outperforms them across the tested datasets, achieving a maximum F1 of 93.04% against the LUN dataset. The fake news was restyled (paraphrased) using GPT-3.5 and LLaMA-2 and the benchmark datasets were LUN, PolitiFact, and GossipCop. The procedures comprised content-focused veracity attribution, style-agnostic training, and LLM-based news reframing [74].

Recent work reframes exposure bias in LLM distillation through the lens of imitation learning. A 2025 research team [75] proposed DaD-DAgger, an iterative scheme that treats both ground-truth data (“hard” labels) and the teacher’s predicted distributions (“soft” labels) as dual expert demonstrations. At each iteration, the student generates free-running continuations that are then corrected using both sources and re-added to training, directly countering the train–test mismatch that causes error compounding. Empirically, with a GPT-Neo-1.3B teacher and GPT-Neo-125M student, the method improved next-token perplexity, free-running cross-entropy (multistep generation), and G-Eval text-quality scores across five datasets, with the best results when mixing hard and soft labels (0.25–0.75), outperforming both next-token (teacher forcing) baselines and an imitation-learning KD variant that uses only soft labels (ImitKD). These findings suggest that misinformation-detection systems that distill LLMs, especially those relying on generative rationales or sequence-level labeling, should train with student-generated trajectories and evaluate with sequence-level metrics, not only token-level likelihoods.

Another method involves fine-tuning ChatGPT, Gemini, and similar models, using the LIAR dataset for both training and testing. This resulted in high metrics of 89% accuracy for GPT and 91% accuracy for Google’s Gemini [76].

FakeNewsGPT4, a cross-domain framework for detecting false news, was introduced in 2024. This framework improves manipulation reasoning by incorporating forgery-tailored data into large vision–language models while also utilizing general knowledge as a plus. ImageBind and Vicuna LLMs were employed in the experimental environment, where they were fine-tuned with the DGM4 and NewsCLIPpings datasets in different settings (single or multidomain and cross-dataset). This process entailed the injection of two knowledge types (visual artifact traces and semantic correlation) through dual lightweight modules, which was subsequently followed by instruction-tuned alignment. The findings demonstrated substantial enhancements over the baselines, with the multimodal FKA-Owl method surpassing state-of-the-art methods such as HAMMER and PandaGPT, particularly under domain shift, with an AUC of up to 89.61% [77].

Nuanced false news detection was achieved through the utilization of ChatGPT, Gemini, and Meta LLaMA 3 models. A study conducted in 2024 examined the efficacy of a fine-tuned LLaMA 3 (8B) model in the detection of multiclass false news across bilingual datasets (English and Romanian). The procedure involved dataset tokenization, text normalization, and LoRa tuning for the proposed. The benchmarking included GPT-4, Gemini, and LLaMA-2. The proposed model outperformed all LLaMA versions on a Romanian dataset, achieving an accuracy of 39% [78].

In a comparable experiment, the RUN-AS dataset was trained on a fine-tuned LLaMA 3 model over the Iberian Languages Evaluation Forum (IberLEF) 2024 with multilanguage datasets annotated by the 5W1H journalistic method. The proposed system obtained a macro F1-score of 0.59658, placing it second in the competition [79], demonstrating superior performance against other frameworks.

4.3.3. Fact-Checking

Large language models are able to streamline fake data detection by comparing statements to reliable sources with verified data. This method also identifies factual discrepancies or stylistic cues which indicate the existence of falsehoods. The accuracy of the verification process is improved by the ability of more advanced models to analyze complex sentences and comprehend context.

Two methods utilizing LLMs were proposed by a research team. The initial approach, Reinforcement Retrieval Leveraging Fine-grained Feedback (FFRR), aims to boost both the precision and informativeness of the data obtaining process [80]. FFRR exhibited substantial enhancements in comparison to both LLM and non-LLM models. FFRR employs prompting techniques to generate subquestions that analyze a variety of aspects of a claim in order to collect pertinent data. FFRR subsequently extracts comprehensive outputs from the LLM that reflect both document and question perspectives of the retrieved content. This feedback is employed as an incentive to improve the retrieval strategy for intermediate queries and refine the document retrieval list. The RAWFC and LIAR-RAW datasets are used to evaluate the experimental environment, which is based on RoBERTa. The results demonstrate that FFRR, incorporating both document and question level representations, significantly surpasses all baseline models, achieving macro F1-scores of 57% on the RAWFC dataset and 33.5% against the LIAR-RAW dataset.

Hierarchical Step-by-Step (HiSS) is a second fact-checking method that instructs LLMs to deconstruct a claim into multiple subclaims and subsequently validates them through a succession of question-and-answer rounds [81]. This process entails the deconstruction of the claim, the verification of each subclaim, and the subsequent provision of a final prediction. The large language model initially deconstructs the claim into subclaims to guarantee that no part is neglected. Then, the model verifies each subclaim by generating and answering to series of detailed inquiries, utilizing external sources as necessary. The model gives a definitive classification of the initial data after verification. Experimental results on two fake and real news datasets, RAW-Fact-Checking (RAWFC) and LIAR, show that HiSS prompting surpasses both fully supervised baselines and advanced few-shot contextual learning methods. Pitted against standard prompting, basic Chain-of-Thought (CoT), and ReAct prompting methods, Hierarchical Step-by-Step demonstrates the highest F1-score compared to other supervised methods, achieving a 53.9% score against the RAWFC dataset and a 37.5% score on the LIAR dataset. Notably, the method generated high-quality, detailed, and interpretable reasoning outputs. The main two issues about the evaluation of news claims that HiSS addresses are the introduction of “hallucinations” as facts and the omission of critical details.

An alternative approach entails model fine-tuning for generating clear reasoning that either validates or critiques news titles, thus establishing a transparent justification for its labeling [82]. The approach uses model distillation on FLAN-T5 and LLaMa-2 models, then combines the Chain of Thought (CoT) reasoning method with current large language models, allowing for step-by-step reasoning that replicates the human thought process. The proposed models surpassed similar models by 11.9% on all metrics.

Prompted Weak Supervision with Credibility Signals (PASTEL) is another pertinent study that prompts LLMs to produce weak labels for different credibility indicators [83]. First, open-ended questions are used to obtain responses from claims that are subsequently categorized using generic prompts. The claim is paired with an instruction prompt, and each signal is sequentially subjected to a particular prompt in order to assess believability signals. In a zero-shot prompt, fakes are binary classified (fake vs. genuine news). A task-neutral mapping prompt is implemented when string-matching is incapable of assigning a class. This method, which integrates weak supervision with zero-shot labeling, outperforms other state-of-the-art models by 38.3% accuracy on four datasets (PolitiFact, FakeNewsAMT, Celebrity, and GossipCop) and eliminates the necessity of human annotated class labeling during model training. PASTEL achieved an 86.7% performance when compared to a RoBERTa supervised architecture and a 63% superior cross-domain generalization.

4.3.4. Contextual Investigation

Large language models can use their training data to analyze the context in which information is presented. This process entails the evaluation of the logical development of information and the identification of discrepancies and manipulative language cues or rhetorical techniques that are frequently encountered in false news. Furthermore, semantic analysis assists in the discovery of the text’s fundamental intentions and meanings.

A network that was designed for contextual investigation was introduced in a recent study about the Adaptive Rationale Guidance (ARG) [84]. The method pairs small and large language models, allowing the small ones to select useful arguments in their classification decision. ARG encodes inputs with small models that use the reasoning from the large ones through news–rationale interactions. The final judgments are established upon aggregated interactive features. Additionally, ARG-D, a rationale-free, cost-sensitive variant of ARG, is claimed to be functional without LLM queries. The claimed results indicate that ARG outperforms all base models, with a macro F1-score of up to 0.784 for the Chinese language and 0.790 for the English language. Additionally, in situations where computational cost is a concern, the distilled ARG-D model retains good accuracy.

SheepDog [74], a methodology that was previously investigated, employs LLM news paraphrasing to customize news in various stylistic modes using tailored prompts. This method guarantees that the detector is resistant to stylistic attacks by prioritizing content over stylistic cues.

DELL, a three stage framework for misinformation detection, was devised by a distinct research team [85]. Initially, LLMs produce fabricated reactions to news in order to represent a variety of perspectives. Second, they offer reasoning for proxy tasks, which refine feature embedding. Next, the method employs three LLM-based strategies that combine classifications taken from the specialized models in order to enhance calibration. The method achieved a macro F1 metric that surpasses other state-of-the-art methods by 16.8% in experiments conducted on a variety of datasets by DELL.

4.3.5. Fake Profile Identification

By examining linguistic patterns, behavioral anomalies, and contextual inconsistencies, research teams have developed methodologies to identify false profiles by using the LLMs’ advanced natural language understanding (NLU) capabilities.

MedGraph, a model that was developed for online dating platforms, serves as an appropriate illustration [86]. MedGraph utilizes a Graph Neural Network (GNN) framework that is founded on embedding to detect deceptive behavior, especially edges inside a temporal mutual graph. The method incorporates both the interaction patterns of users and their attributes. Its methodology comprises the following components: a Motif-Based GNN which identifies corresponding user attributes with a bipartite graph, another Motif-Based GNN is used for neighbor selection; the unusual interaction between users is revealed using a Temporal Behavior Embedding for historical interaction analysis; the probability for a certain user action to be abnormal is achieved by a co-attention mechanism; finally, a prediction layer evaluates the likelihood of a specific interaction to be malicious. Collectively, these elements establish an efficient framework for identifying malevolent conduct in online dating media.

The LLM-empowered Rumor Detection (LeRuD) is an alternate method that employs curated prompting cues to evaluate credibility signs, including source credibility and factual accuracy [87]. The method identifies fake content without labeled datasets tuning by integrating weak supervision with zero-shot classification. The proposed LeRuD method, applied on Twitter15, Twitter16, and Weibo datasets, with task-agnostic mapping and synthetic augmentation, outperformed other SOTA rumor detection models by scores between 3.2% and 7.7%, proving its ability for zero-shot fake content detection without using training data.

Additionally, a hybrid method was devised to detect false and AI-generated LinkedIn profiles within the registration phase, prior to establishing connections between others [88]. Section and Subsection Tag Embedding (SSTE) is a method that prioritizes the consistency of textual and metadata information in profiles submitted during registration. The datasets consisted of data from 3600 LinkedIn profiles, including 1800 legitimate users, 600 human-generated false profiles, and 1200 GPT-generated fakes. The data was tagged and preprocessed using GloVe and Flair. The BERT, RoBERTa, LR, RF, and SVM classifiers were trained and benchmarked within a variety of scenarios with this dataset. The results indicated that early-stage detection performance in the absence of dynamic user activity data is viable in a devised scenario. RoBERTa achieved the best metric (96.3% accuracy) for false profile detection, while LLM-generated profile accuracy ranged from 70 to 90%.

5. Critical Discussion

This review follows the development of methods for detecting fake news or fake content from the early AI-driven models to current large-language-driven usage. The paper distinguished four different methodological positions among them: knowledge-based (by using databases and outside facts), style-based (by looking at linguistic writing cues), propagation-based (by mimicking the spread of false information in networks), and source-based (that assesses source credibility).

As shown in Table 1, this work contributes a unified curation of 31 publicly available misinformation datasets, harmonized into a common schema (name, modality, label granularity, description, URL/DOI, and reference IDs). By explicitly annotating modality (text-only vs. text+image) and per-modality supervision, the table surfaces the field’s heavy reliance on text-only corpora and shallow media labels. Our normalization of label schemes (binary vs. 3/4/6/9-class; multiclass vs. multilabel) enables fairer cross-paper comparisons and shows that most benchmarks remain binary, limiting realism. The table also maps topic, language, and provenance (e.g., political fact-checking, COVID-19, multilingual and cross-lingual splits) and flags synthetic/LLM-augmented resources, highlighting potential distribution shift.

Table 1. Datasets used in this review: type, labeling scheme, description, and location.

Reference	Dataset	Data Type	Labels	Description	URL/DOI
[47,57,76]	LIAR	Text	6-class	PolitiFact statements.	https://www.kaggle.com/datasets/doanquanvietnamca/liar-dataset (access on 22 September 2025)
[80,81]	LIAR-RAW	Text	6-class	12,590 PolitiFact statements.	https://opendatalab.com/OpenDataLab/LIAR-RAW/tree/main (access on 22 September 2025)
[30,41,42,49,50]	COVID-19	Text	Binary	Claims made during the COVID-19 pandemic.	https://www.kaggle.com/datasets/arashnic/covid19-fake-news (access on 22 September 2025)
[72]	MM-COIVD	Text	Binary	3581 fake and 10,824 real lines for a COVID-19-themed multilingual dataset (English, Spanish, French, Portuguese Hindi and Italian).	https://github.com/bigheiniu/MM-COVID (access on 22 September 2025)
[80,81]	RAWFC	Text	Binary	2012 Claims sourced from Snopes Media Group.	https://www.dropbox.com/sh/1w7crp3hauoec5m/AABJpG6YWbqrumypBpHJEDnSa?dl=0 (access on 22 September 2025)
[70,83]	PolitiFact	Text	6-class	PolitiFact statements.	https://github.com/deu30303/AdStyle?tab=readme-ov-file (access on 22 September 2025)
[67]	PolitiFact++	Text	6-class	520 lines of PolitiFact statements augmented by LLM.	https://doi.org/10.48550/arXiv.2408.11871 (access on 22 September 2025)
[66,70]	GossipCop	Text	Binary	Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://www.kaggle.com/datasets/akshaynarayananb/gossipcop (access on 22 September 2025)
[67]	GossipCop++	Text	Binary	Over 20,500 lines of LLM augmented Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://doi.org/10.48550/arXiv.2408.11871 (access on 22 September 2025)
[51,60,61]	Fakeddit	Text and image	Binary/3-class/6-class	Over 1 million samples for 2-way, 3-way, and 6-way classification categories through distant supervision,	https://github.com/entitize/Fakeddit (access on 22 September 2025)
[85]	PHEME	Text	9-class, 3 veracity categories	Collection of Twitter rumors and nonrumors posted during breaking news.	https://www.kaggle.com/datasets/usharengaraju/pheme-dataset (access on 22 September 2025)
[34,58,87]	Twitter15	Text	4-class (True/False/Non-rumor/Unverified)	Rumour verification.	https://www.kaggle.com/datasets/syntheticprogrammer/rumor-detection-acl-2017 (access on 22 September 2025)
[34,58,87]	Twitter16	Text	4-class (True/False/Non-rumor/Unverified)	Rumour verification.	https://www.kaggle.com/datasets/syntheticprogrammer/rumor-detection-acl-2017 (access on 22 September 2025)
[31,36,39,48,59]	ISOT Fake News	Text	Binary	Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://www.kaggle.com/datasets/csmalarkodi/isot-fake-news-dataset (access on 22 September 2025)
[59]	MediaEval 2016	Text and image	Binary	Verifying multimedia use: tweets/posts tagged real vs. fake/misleading.	http://www.multimediaeval.org/mediaeval2016/ (access on 22 September 2025)
[56]	CheckThat-2022	Text	4-class	Claim verification, English claim, German test.	https://sites.google.com/view/clef2022-checkthat (access on 22 September 2025)
[77]	DGM4	Text and image	Binary classes for both media	Large-scale dataset for studying machine-generated multimodal media manipulation.	https://github.com/rshaojimmy/MultiModal-DeepFake (access on 22 September 2025)
[77]	NewsCLIPpings	Text and image	Binary for each media	Automatically generated out-of-context image-caption pairs in the news media.	https://github.com/g-luo/news_clippings (access on 22 September 2025)
[62]	TR_FaRe_News	Text	Binary	Turkish news articles and columns.	https://doi.org/10.1109/ACCESS.2024.3354165
[66]	MegaFake	Text	Binary	Fake news generated by LLM based on GossipCop, which is proposed by FakeNewsNet.	https://github.com/zhe-wang0018/MegaFake (access on 22 September 2025)
[68]	Med-MMHL	Text and image	Binary for each media	A total of 12,968 articles, 27,633 tweets, including article images and tweet images.	https://github.com/styxsys0927/Med-MMHL (access on 22 September 2025)
[83]	FakeNewsAMT	Text	Binary	480 articles for automated evaluation of misinformation.	https://github.com/ComplexData-MILA/misinfo-datasets (access on 22 September 2025)
[83]	Celebrity	Text	Binary	Article-level veracity dataset of celebrity news (actors, singers, socialites, politicians).	https://doi.org/10.48550/arXiv.2309.07601
[74]	LUN	Text	4 class	Labeled unreliable news.	https://github.com/BUPT-GAMMA/CompareNet_FakeNewsDetection/releases/tag/dataset (access on 22 September 2025)
[87]	Weibo	Text	Binary	Rumor detection based on Twitter15/16 and Weibo, uses news and comments propagation.	https://doi.org/10.48550/arXiv.2402.03916
[70]	Constraint	Text	Binary	Dataset of social media posts on COVID-19.	https://drive.google.com/file/d/16CtbPjyILGc1P1EcDDXhrckw22mKQ5aT/view (access on 22 September 2025)
[69]	Dgpt_human/Dgpt_std/Dgpt_mix/Dgpt_cot	Text	Binary	23,278 for std, 1000 for mix, and 1737 for cot. ChatGPT-generated fake-news sets built from a human news corpus of 23,525 samples.	https://doi.org/10.1137/1.9781611978032.50
[45]	Curpos	Text	Binary	Article classification.	https://doi.org/10.1007/s13278-024-01198-w
[78]	Ro-En Dataset	Text	4-class	200 (100 vs. 100) testing Romanian and English language claims.	https://www.kaggle.com/datasets/restem/en-ro-datasets-for-llm-testing (access on 22 September 2025)
[31,38,40,46]	Kaggle 2016 (US election news)	Text	Binary	Data relevant for the 2016 US Presidential Election, including up-to-date primary results.	https://www.kaggle.com/datasets/benhamner/2016-us-election (access on 22 September 2025)
[88]	LinkedIn profile Dataset	Text	Binary	1800 legitimate LinkedIn profiles (LLPs), 600 fake LinkedIn profiles (FLPs), and 1200 profiles generated by ChatGPT (CLPs).	https://doi.org/10.1145/3603163.3609064

Table 1. Datasets used in this review: type, labeling scheme, description, and location.

Reference	Dataset	Data Type	Labels	Description	URL/DOI
[47,57,76]	LIAR	Text	6-class	PolitiFact statements.	https://www.kaggle.com/datasets/doanquanvietnamca/liar-dataset (access on 22 September 2025)
[80,81]	LIAR-RAW	Text	6-class	12,590 PolitiFact statements.	https://opendatalab.com/OpenDataLab/LIAR-RAW/tree/main (access on 22 September 2025)
[30,41,42,49,50]	COVID-19	Text	Binary	Claims made during the COVID-19 pandemic.	https://www.kaggle.com/datasets/arashnic/covid19-fake-news (access on 22 September 2025)
[72]	MM-COIVD	Text	Binary	3581 fake and 10,824 real lines for a COVID-19-themed multilingual dataset (English, Spanish, French, Portuguese Hindi and Italian).	https://github.com/bigheiniu/MM-COVID (access on 22 September 2025)
[80,81]	RAWFC	Text	Binary	2012 Claims sourced from Snopes Media Group.	https://www.dropbox.com/sh/1w7crp3hauoec5m/AABJpG6YWbqrumypBpHJEDnSa?dl=0 (access on 22 September 2025)
[70,83]	PolitiFact	Text	6-class	PolitiFact statements.	https://github.com/deu30303/AdStyle?tab=readme-ov-file (access on 22 September 2025)
[67]	PolitiFact++	Text	6-class	520 lines of PolitiFact statements augmented by LLM.	https://doi.org/10.48550/arXiv.2408.11871 (access on 22 September 2025)
[66,70]	GossipCop	Text	Binary	Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://www.kaggle.com/datasets/akshaynarayananb/gossipcop (access on 22 September 2025)
[67]	GossipCop++	Text	Binary	Over 20,500 lines of LLM augmented Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://doi.org/10.48550/arXiv.2408.11871 (access on 22 September 2025)
[51,60,61]	Fakeddit	Text and image	Binary/3-class/6-class	Over 1 million samples for 2-way, 3-way, and 6-way classification categories through distant supervision,	https://github.com/entitize/Fakeddit (access on 22 September 2025)
[85]	PHEME	Text	9-class, 3 veracity categories	Collection of Twitter rumors and nonrumors posted during breaking news.	https://www.kaggle.com/datasets/usharengaraju/pheme-dataset (access on 22 September 2025)
[34,58,87]	Twitter15	Text	4-class (True/False/Non-rumor/Unverified)	Rumour verification.	https://www.kaggle.com/datasets/syntheticprogrammer/rumor-detection-acl-2017 (access on 22 September 2025)
[34,58,87]	Twitter16	Text	4-class (True/False/Non-rumor/Unverified)	Rumour verification.	https://www.kaggle.com/datasets/syntheticprogrammer/rumor-detection-acl-2017 (access on 22 September 2025)
[31,36,39,48,59]	ISOT Fake News	Text	Binary	Reuters truthful news, Wikipedia, and unreliable websites flagged by PolitiFact.	https://www.kaggle.com/datasets/csmalarkodi/isot-fake-news-dataset (access on 22 September 2025)
[59]	MediaEval 2016	Text and image	Binary	Verifying multimedia use: tweets/posts tagged real vs. fake/misleading.	http://www.multimediaeval.org/mediaeval2016/ (access on 22 September 2025)
[56]	CheckThat-2022	Text	4-class	Claim verification, English claim, German test.	https://sites.google.com/view/clef2022-checkthat (access on 22 September 2025)
[77]	DGM4	Text and image	Binary classes for both media	Large-scale dataset for studying machine-generated multimodal media manipulation.	https://github.com/rshaojimmy/MultiModal-DeepFake (access on 22 September 2025)
[77]	NewsCLIPpings	Text and image	Binary for each media	Automatically generated out-of-context image-caption pairs in the news media.	https://github.com/g-luo/news_clippings (access on 22 September 2025)
[62]	TR_FaRe_News	Text	Binary	Turkish news articles and columns.	https://doi.org/10.1109/ACCESS.2024.3354165
[66]	MegaFake	Text	Binary	Fake news generated by LLM based on GossipCop, which is proposed by FakeNewsNet.	https://github.com/zhe-wang0018/MegaFake (access on 22 September 2025)
[68]	Med-MMHL	Text and image	Binary for each media	A total of 12,968 articles, 27,633 tweets, including article images and tweet images.	https://github.com/styxsys0927/Med-MMHL (access on 22 September 2025)
[83]	FakeNewsAMT	Text	Binary	480 articles for automated evaluation of misinformation.	https://github.com/ComplexData-MILA/misinfo-datasets (access on 22 September 2025)
[83]	Celebrity	Text	Binary	Article-level veracity dataset of celebrity news (actors, singers, socialites, politicians).	https://doi.org/10.48550/arXiv.2309.07601
[74]	LUN	Text	4 class	Labeled unreliable news.	https://github.com/BUPT-GAMMA/CompareNet_FakeNewsDetection/releases/tag/dataset (access on 22 September 2025)
[87]	Weibo	Text	Binary	Rumor detection based on Twitter15/16 and Weibo, uses news and comments propagation.	https://doi.org/10.48550/arXiv.2402.03916
[70]	Constraint	Text	Binary	Dataset of social media posts on COVID-19.	https://drive.google.com/file/d/16CtbPjyILGc1P1EcDDXhrckw22mKQ5aT/view (access on 22 September 2025)
[69]	Dgpt_human/Dgpt_std/Dgpt_mix/Dgpt_cot	Text	Binary	23,278 for std, 1000 for mix, and 1737 for cot. ChatGPT-generated fake-news sets built from a human news corpus of 23,525 samples.	https://doi.org/10.1137/1.9781611978032.50
[45]	Curpos	Text	Binary	Article classification.	https://doi.org/10.1007/s13278-024-01198-w
[78]	Ro-En Dataset	Text	4-class	200 (100 vs. 100) testing Romanian and English language claims.	https://www.kaggle.com/datasets/restem/en-ro-datasets-for-llm-testing (access on 22 September 2025)
[31,38,40,46]	Kaggle 2016 (US election news)	Text	Binary	Data relevant for the 2016 US Presidential Election, including up-to-date primary results.	https://www.kaggle.com/datasets/benhamner/2016-us-election (access on 22 September 2025)
[88]	LinkedIn profile Dataset	Text	Binary	1800 legitimate LinkedIn profiles (LLPs), 600 fake LinkedIn profiles (FLPs), and 1200 profiles generated by ChatGPT (CLPs).	https://doi.org/10.1145/3603163.3609064

The dataset audit reveals three consistent patterns. First, there is a strong text-only and binary bias: of the 31 resources, approximately 84% are text-only (26/31) and 65% are binary-only (20/31). Although multiclass options exist (e.g., LIAR/PolitiFact 6-class; Twitter15/16 and LUN 4-class; PHEME 9-class; Fakeddit up to 6-class), many studies still collapse them to binary.

Second, multimodality is limited, and media labels are shallow: only 5/31 pair text with images (Fakeddit, MediaEval 2016, DGM4, NewsCLIPpings, Med-MMHL), often with per-modality binary tags rather than rich cross-modal alignment; audio/video are largely absent, misaligning with modern formats such as short video and deepfakes.

Third, the analysis shows a topical and linguistic skew alongside a rising synthetic trend: political fact-checking and COVID-19 dominate, multilingual coverage remains sparse (notable exceptions include MM-COVID, TRFaReNews, Weibo, Ro-En, and Check-That cross-lingual), and LLM-augmented sets are increasing (PolitiFact++, GossipCop++, MegaFake, DGPT variants, LinkedIn–ChatGPT), which aids scale but risks distribution shift and overfitting to generator artifacts unless balanced with real, event-split data.

For a better critical analysis, Table 2 was compiled as a consolidated research map that standardizes recent studies into a single table with consistent fields, references, methodological perspectives, datasets, classifiers, reported accuracy, and notes on LLM involvement and other particulars.

Table 2. Research papers presented in this review, used datasets, classifiers, claimed results, and particularities.

Ref.	Perspective	Datasets	Classifiers	Accuracy %	Type. LLM Enhancing. Other Considerations
[30]	StyleBased	COVID-19	SVM	98	Binary
[31]	StyleBased	ISOT	SVM	99	Binary.
[32]	StyleBased	BuzzFeed	SVM+KNN	99	Binary.
[33]	StyleBased	BuzzFeed	NB+SVD	99	Binary.
[34]	StyleBased	Twitter	NB + CountV	94.7	Binary.
[35]	StyleBased	JobPostings	LR	97.92	Binary.
[36]	StyleBased	ISOT	SVM + LR + LSTM	95.62	Binary.
[37]	StyleBased	Kaggle 2016	DT (XGBoost)	97	Binary.
[38]	StyleBased	Kaggle 2016	DT+RF	99.64	Binary.
[39]	StyleBased	ISOT	DT + IBAVO-AO	92.5	Binary.
[40]	StyleBased	Kaggle 2016	RF	99.32	Binary.
[41]	StyleBased	COVID-19	LCRF-HB	98.3	Binary.
[42]	StyleBased	COVID-19	CNB, NB	92	Binary.
[45]	StyleBased	Curpos	GloVe +BiLSTM	98.97	Binary.
[46]	StyleBased	Kaggle	LSTM + Word2Vec	94	Binary.
[47]	StyleBased	LIAR	RNN-LSTM	99.1	Binary.
[48]	StyleBased	ISOT	SA-BiLSTM	99.98	Binary.
[49]	StyleBased	COVID-19	CNN	96.19	Binary.
[50]	StyleBased	COVID-19	CNN–Bi-LSTM–SelfAttention	98.71	Binary.
[51]	StyleBased	Fakeddit	TextGCN+ViT	94.17 binary 90.14 3-class 75.91 6-class	Binary, 3-class, 6-class
[56]	StyleBased	CheckThat-2022	mBERT	51 binary 34 4-class	Binary. 4-class.
[57]	StyleBased	LIAR	LightGBM	41	6-class.
[58]	StyleBased	Twitter15, Twitter16, PHEME	GBCA	92.6 4-class 70.8 9-class	4-class. 9-class. Best scores shown.
[59]	StyleBased	ISOT, RESNET MediaEval 2016	BERT + VGG-19 with RF and SVM	99 textual data 94.4 MediaEval	Multimodal binary. Only best scores shown.
[60]	StyleBased	Fakeddit	BETR + VGG-19	92	Binary.
[61]	StyleBased	Fakeddit	ELD-FN	88.83	Binary.
[62]	StyleBased	TR_FaRe_News	BERTurk + CNN	94	Binary, Turkish Language.
[66]	StyleBased	MegaFake, GossipCop	LLaMa 3-70b,	46.79 82.59	Binary. For human and LLM enhanced data.
[67]	StyleBased	GossipCop++ PolitiFact++	Bert Variants trained on human and LLM generated data	98	Binary. GPT used to enhance datasets Best result shown
[68]	StyleBased	Med-MMHL	FN-BERT CLIP	98.60 97.95	Binary. Metrics shown for human datasets.
[69]	StyleBased	Dgpt variants	RoBERTa	95.3	Binary. GPT-4 used to enhance datasets. Results for human variant
[70]	StyleBased	PolitiFact, GossipCop, Constraint	AdStyle	94.6, 87.97, 98.49.	Binary. Only clear settings scores shown
[72]	StyleBased	MM-COVID	LLaMa 3	95.1	Binary. LLaMa 3 enhanced datasets. Only English clear settings score shown.
[74]	StyleBased	LUN	SheepDog	93.04	Binary.
[76]	StyleBased	LIAR	GPT-4, GEMINI	89, 91	Binary.
[77]	StyleBased	DGM, NewsCLIPpings,	FakeNewsGPT4	89.6	Binary. Multimodal datasets used.
[78]	StyleBased	Ro-En dataset	LLaMa 3	39	4-class, Romanian Dataset
[79]	StyleBased and Source Reliability	RUN-AS	LLaMa 3	59.68	6-class. Spanish language
[80]	StyleBased	RAWFC LIAR-RAW	RoBERTa	52.8 29	Binary. LLM augmented datasets used. Only SOTA model presented.
[81]	StyleBased	RAWFC LIAR-RAW	GPT-3.5 HiSS	53.9 37.5	Binary. LLM augmented datasets used.
[83]	StyleBased	PolitiFact	PASTEL	86.7	Binary. LLM augmented datasets used.
[85]	StyleBased	Pheme	LLaMa 2-70B DELL Selective	82	Binary. LLM augmented datasets used. Only result in unaltered environment included.
[87]	Content propagation	Twitter Weibo	GPT 3.5 Graph based LeRuD	94.01 98.09	Binary. Multiple language content propagation datasets.
[88]	Source Reliability	LinkedIn profile Dataset	BERT + SSTE	95.67	Binary. LLM enhanced dataset. Result shown on numerical and textual data.

Table 2. Research papers presented in this review, used datasets, classifiers, claimed results, and particularities.

Ref.	Perspective	Datasets	Classifiers	Accuracy %	Type. LLM Enhancing. Other Considerations
[30]	StyleBased	COVID-19	SVM	98	Binary
[31]	StyleBased	ISOT	SVM	99	Binary.
[32]	StyleBased	BuzzFeed	SVM+KNN	99	Binary.
[33]	StyleBased	BuzzFeed	NB+SVD	99	Binary.
[34]	StyleBased	Twitter	NB + CountV	94.7	Binary.
[35]	StyleBased	JobPostings	LR	97.92	Binary.
[36]	StyleBased	ISOT	SVM + LR + LSTM	95.62	Binary.
[37]	StyleBased	Kaggle 2016	DT (XGBoost)	97	Binary.
[38]	StyleBased	Kaggle 2016	DT+RF	99.64	Binary.
[39]	StyleBased	ISOT	DT + IBAVO-AO	92.5	Binary.
[40]	StyleBased	Kaggle 2016	RF	99.32	Binary.
[41]	StyleBased	COVID-19	LCRF-HB	98.3	Binary.
[42]	StyleBased	COVID-19	CNB, NB	92	Binary.
[45]	StyleBased	Curpos	GloVe +BiLSTM	98.97	Binary.
[46]	StyleBased	Kaggle	LSTM + Word2Vec	94	Binary.
[47]	StyleBased	LIAR	RNN-LSTM	99.1	Binary.
[48]	StyleBased	ISOT	SA-BiLSTM	99.98	Binary.
[49]	StyleBased	COVID-19	CNN	96.19	Binary.
[50]	StyleBased	COVID-19	CNN–Bi-LSTM–SelfAttention	98.71	Binary.
[51]	StyleBased	Fakeddit	TextGCN+ViT	94.17 binary 90.14 3-class 75.91 6-class	Binary, 3-class, 6-class
[56]	StyleBased	CheckThat-2022	mBERT	51 binary 34 4-class	Binary. 4-class.
[57]	StyleBased	LIAR	LightGBM	41	6-class.
[58]	StyleBased	Twitter15, Twitter16, PHEME	GBCA	92.6 4-class 70.8 9-class	4-class. 9-class. Best scores shown.
[59]	StyleBased	ISOT, RESNET MediaEval 2016	BERT + VGG-19 with RF and SVM	99 textual data 94.4 MediaEval	Multimodal binary. Only best scores shown.
[60]	StyleBased	Fakeddit	BETR + VGG-19	92	Binary.
[61]	StyleBased	Fakeddit	ELD-FN	88.83	Binary.
[62]	StyleBased	TR_FaRe_News	BERTurk + CNN	94	Binary, Turkish Language.
[66]	StyleBased	MegaFake, GossipCop	LLaMa 3-70b,	46.79 82.59	Binary. For human and LLM enhanced data.
[67]	StyleBased	GossipCop++ PolitiFact++	Bert Variants trained on human and LLM generated data	98	Binary. GPT used to enhance datasets Best result shown
[68]	StyleBased	Med-MMHL	FN-BERT CLIP	98.60 97.95	Binary. Metrics shown for human datasets.
[69]	StyleBased	Dgpt variants	RoBERTa	95.3	Binary. GPT-4 used to enhance datasets. Results for human variant
[70]	StyleBased	PolitiFact, GossipCop, Constraint	AdStyle	94.6, 87.97, 98.49.	Binary. Only clear settings scores shown
[72]	StyleBased	MM-COVID	LLaMa 3	95.1	Binary. LLaMa 3 enhanced datasets. Only English clear settings score shown.
[74]	StyleBased	LUN	SheepDog	93.04	Binary.
[76]	StyleBased	LIAR	GPT-4, GEMINI	89, 91	Binary.
[77]	StyleBased	DGM, NewsCLIPpings,	FakeNewsGPT4	89.6	Binary. Multimodal datasets used.
[78]	StyleBased	Ro-En dataset	LLaMa 3	39	4-class, Romanian Dataset
[79]	StyleBased and Source Reliability	RUN-AS	LLaMa 3	59.68	6-class. Spanish language
[80]	StyleBased	RAWFC LIAR-RAW	RoBERTa	52.8 29	Binary. LLM augmented datasets used. Only SOTA model presented.
[81]	StyleBased	RAWFC LIAR-RAW	GPT-3.5 HiSS	53.9 37.5	Binary. LLM augmented datasets used.
[83]	StyleBased	PolitiFact	PASTEL	86.7	Binary. LLM augmented datasets used.
[85]	StyleBased	Pheme	LLaMa 2-70B DELL Selective	82	Binary. LLM augmented datasets used. Only result in unaltered environment included.
[87]	Content propagation	Twitter Weibo	GPT 3.5 Graph based LeRuD	94.01 98.09	Binary. Multiple language content propagation datasets.
[88]	Source Reliability	LinkedIn profile Dataset	BERT + SSTE	95.67	Binary. LLM enhanced dataset. Result shown on numerical and textual data.

By harmonizing heterogeneous reports, the table enables fairer cross-paper comparisons (e.g., binary vs. multiclass settings, unimodal vs. multimodal inputs), surfaces performance cliffs tied to dataset choice and label granularity, and distinguishes classical from LLM-augmented pipelines.

This table further underscores four constraints that inform future work: a strong text-only and binary bias (84% text-only; 65% binary-only), which reduces label granularity and hinders real-world generalization; underrepresented and shallow multimodality, only 5/31 pair text with images, typically with per-modality binary tags, while audio/video are largely absent, misaligned with short-video and deepfake formats; narrow label taxonomies often collapsed to binary, with few resources supporting sequence/story-level or event/time-split evaluation to control temporal leakage; topical/linguistic skew with rising synthetic content, political fact-checking, and COVID-19 dominate; multilingual/low-resource coverage is thin; and LLM-augmented sets, while useful for scale, risk distribution shift and overfitting unless balanced with real, event-split media and clear provenance.

As seen in Table 2, current fake news detection uses different models, but the vast majority focus on variants of stylistic analysis and achieve a far from standardized result. This lack of standardized performance benchmarking has its roots in a problematic fake news dataset gathering. The existence of widely different and mostly imbalanced datasets makes objective performance comparison challenging. This problem has plagued fake news research from its beginnings, and if future research does not insist on the building of more standardized datasets that encumber human annotation, the same issue will persist.

The current analysis illustrates how current research systems suffer from an immense overreliance on textual data that has been gathered from a limited number of specialized fact-checking projects, is often binary-labeled, and is presented only in English. Audio and video are typically absent, and images, when included, are usually only tagged as “real” or “generated”. Transformer models, LLMs, and hybrid approaches have begun to address these limits, with the most promising results so far, but progress remains modest.

As work on these “traditional” datasets approaches saturation, real-world generalization is doubtful: misinformation today is not confined to blogs and articles but spreads through YouTube, TikTok, and other platforms as images, deepfakes, and short videos. Effective detectors must therefore analyze all modalities: text, image, audio, and video. Early steps in this direction are visible in vision–language models and multimodal datasets such as Fakeddit.

In the same manner, Table 2 illustrates how earlier methods, with the exception of transformer, LLMs, and combined models, still tackle, from a stylistic point of view, mostly binary, human annotated, supervised or unsupervised, English language and textual datasets, which is a clear sign that their usage needs to be upgraded with the newer classes of models.

Another key gap is the narrow use of word-embedding techniques. Most systems rely on a small set of standard embeddings, even though more than thirty alternatives have shown strong results in other domains. Broadening the embedding toolkit for fake news detection could boost accuracy by capturing finer-grained semantic and syntactic signals in text.

For future development, the research community will have to focus on hybrid systems that combine the best aspects of multiple approaches in order to capture the many facets of fake news.

6. Conclusions

The current review shows that even if large language models have seen a rise in usage for fake news detection, more traditional and established models, especially the transformer-based architecture, are still widely used, with noticeable results in various scenarios, but mostly as classifiers.

Large language models, on the other hand, have a mixed role in the current misinformation detection field, showing mixed results in occupying more than one role in this ecosystem.

Firstly, to help with model training, LLMs have been used to generate fictitious content and adversarial instances. For the MegaFake project, for example, a sizable dataset of fake news generated with the help of LLMs was produced, and it illustrated how detectors become more potent after being trained on both human and AI-generated news. In a similar vein, the AdStyle technique enhanced the adversarial by automatically rewriting news articles using GPT-3.5 in a variety of styles, such as humorous, sarcastic, poetic, and so on. This strengthened detectors and tested them against redesigned fake news. The improved model remained functional even after it was styled. Balancing datasets must be a future concern for researchers because relying too much on training data produced by AI can lead to biases.

Secondly, thanks to LLM-based methods, it is now easier to understand how fake news detection works. One method for helping large models break down claims and validate each element is HiSS prompting, which makes the reasoning more understandable. HiSS achieved higher accuracy than other state-of-the-art methodologies on complex datasets. By giving clear and thorough explanations for its findings, the fact-checking process was able to address problems like a lack of evidence or AI “hallucinations”. This is an important step toward detectors that give a conclusion and a clear explanation.

Furthermore, LLM-driven multimodal frameworks are being developed. The FakeNewsGPT4 system, for instance, uses specialized knowledge modules to improve a vision–language model in order to analyze visual artifacts and textual context at the same time. By integrating visual anomaly detection and semantic world knowledge into a GPT-4 architecture, a research team has achieved superior results to previous multimodal false news detectors when tested across domains.

Such new features show how flexible LLMs can be. False news detection systems can be strengthened by their ability to support more flexible learning methods (e.g., few-shot prompts and synthetic data generation) as well as comprehension and reasoning in a range of contexts. The methods for spotting fake news have steadily improved over time.

Prior to deep neural networks, machine learning classifiers, and increasingly potent LLM-based systems, human specialists could only use manual methods for fact-checking but currently gain more tools for this task since detectors have become more accurate, generalizable, and able to handle complex, real-world disinformation with each new wave of invention.

The review shows that the fight against false information is far from over. The lack of complete, qualitative datasets (especially for rare languages or text-vision data) and the dynamic nature of fake news dissemination and creation represent two problems that still plague state-of-the-art automated methods available today.

Fake news detection in real-life scenarios is a multiclass classification problem and involves mixed media. This study barely identified a couple of datasets designed solely for multilabel classification. In multilabel classification, a single instance can have multiple labels. This assumption is not applicable to fake news detection (e.g., an article cannot be both true and mostly true). In contrast, in multiclass classification, each instance can have only one class assigned to it. A special case of multiclass classification is binary classification with only two classes. Among the over 30 publicly available datasets gathered by this study and used in the most recent experiments, over two-thirds are binary. There are some examples of more than two classes (LIAR, CheckThat, LUN, Fakeddit, PHEME, and the Twitter collections), but they are mostly actually used selectively for binary classification; thus, this field needs more diversity. The [79] project uses a known journalistic methodology involving the 5W1H labels and tackles both style based and source reliability methods and merits future attention.

The substantial language disparity in fake news detection presents another significant challenge. Current detection approaches still focus mainly on English content, leaving numerous widely spoken languages with minimal detection capabilities. This limitation stems primarily from the scarcity of publicly available datasets for these languages. Advanced generative language models have already generated context data but are at an initial stage and need further methodological improvement, such as the usage of the CoT prompting strategy, mixed prompting, or variants of the teacher–student methods.

Detection also needs to be faster and more scalable. Large transformers and LLMs work well, but they demand a lot of processing power. Future studies could look into federated approaches, distillation, or model compression to make models accessible without compromising on their accuracy.

Another important issue is making sure that AI-driven detectors are transparent and dependable. Explainable AI (XAI) techniques like reasoning traces (HiSS) or rationale-guided frameworks will help users and domain experts understand and validate the system’s results.

Finally, cross-disciplinary collaboration will become more and more important. Combining ideas from computer science, psychology, social science, and journalism can lead to better false news detection methods and more thorough solutions. These steps could lead to the development of next-gen fake news detection tools that are not only strong and extremely accurate, but also faster, able to process a wider range of information types, and transparent in their reasoning. These kinds of developments are essential to maintaining people’s trust in the information ecosystem and fending off the constantly changing threat of misinformation.