Source-linked AI summary
A Survey of Fake News: Fundamental Theories, Detection Methods, and Opportunities
Xinyi Zhou, Reza Zafarani
TL;DR
Fake news threatens democracy, public trust, and economic stability, motivating a systematic review of detection and intervention research. The survey organizes methods across four perspectives and synthesizes interdisciplinary theories and research opportunities, while noting boundaries in available evidence and method scope.
Problem
Fake news has become a serious social and political problem, creating demand for detection and intervention that address its content, style, propagation, and sources.
Method
The survey reviews and evaluates fake-news detection methods across knowledge, style, propagation, and source perspectives, while synthesizing interdisciplinary theories and research tasks.
Results
The survey identifies potential research tasks and connects interdisciplinary theories to fake-news detection and intervention, including explainable models.
Takeaways & Limitations
The review supports interdisciplinary efforts toward fake-news detection that is efficient and, more importantly, explainable.
Takeaways & Limitations
Fast fact-checking must address recent events as well as common knowledge, while some reported propagation patterns are limited to political news articles.
Abstract
from arXiv · showhide
The explosive growth in fake news and its erosion to democracy, justice, and public trust has increased the demand for fake news detection and intervention. This survey reviews and evaluates methods that can detect fake news from four perspectives: (1) the false knowledge it carries, (2) its writing style, (3) its propagation patterns, and (4) the credibility of its source. The survey also highlights some potential research tasks based on the review. In particular, we identify and detail related fundamental theories across various disciplines to encourage interdisciplinary research on fake news. We hope this survey can facilitate collaborative efforts among experts in computer and information sciences, social sciences, political science, and journalism to research fake news, where such efforts can lead to fake news detection that is not only efficient but more importantly, explainable.
1 INTRODUCTION
Fake news has gained influence through social media, motivating a survey that organizes detection methods, clarifies definitions, and connects interdisciplinary theories to explainable detection and intervention.
- Motivation: Fake news is viewed as a major threat to democracy, journalism, freedom of expression, public trust, and economic stability.The survey links fake news to political events, stock-market fluctuations, and large financial losses.
- Motivation: Fake news can spread farther and faster than truth, especially for political news, while high engagement creates political and economic incentives for malicious actors.The survey cites substantially greater discussion and sharing of prominent fake election stories than stories from major news websites.
- Motivation: Social media accelerates fake-news dissemination by enabling rapid sharing, discussion, and amplification within echo chambers.Around 68% of Americans received news from social media as of August 2018.
- Definitions: The survey distinguishes fake-news-related concepts using authenticity, intention, and whether information is news, and defines fake news narrowly as intentionally false news published by a news outlet.It also presents broad and narrow definitions and discusses research problems raised by those definitions.
- Interdisciplinary research: The survey reviews theories across disciplines and proposes using them to develop well-justified, explainable detection and intervention models.It emphasizes collaboration among computer and information sciences, political science, journalism, social sciences, psychology, and economics.
- Research opportunities: The forensic-psychology theories reviewed target deceptive statements or testimonies rather than fake news, so their applicability requires further investigation.The survey identifies testing whether attributes such as sentiment polarity are statistically distinguishable as a research opportunity.
- Research opportunities: The survey highlights open research tasks, including constructing ground-truth datasets and investigating how fundamental theories can support fake-news research.It focuses more on dataset construction and ground-truth sources than on detailing existing datasets.
- Detection perspectives: Automatic detection methods are organized around knowledge, style, propagation, and source perspectives, covering both news content and its social context.This organization supports comparisons of revealed patterns, model architectures, and detection performance.
2 KNOWLEDGE-BASED FAKE NEWS DETECTION
Knowledge-based fake news detection assesses news authenticity by comparing extracted claims or statements with known facts, using manual or automatic fact-checking. The survey describes manual fact-checking sources and an automatic process based on knowledge extraction and comparison, while identifying scalability, timeliness, and knowledge-quality challenges.
- Fact-checking foundations: Fact-checking compares knowledge extracted from news claims or statements with known facts to assess authenticity.
- Manual Fact-checking: Manual fact-checking is broadly divided into expert-based and crowd-sourced approaches.Expert-based approaches rely on domain experts, whereas crowd-sourced approaches rely on large populations of individuals.
- Manual Fact-checking: Expert-based websites provide ground truth and detailed justifications useful for datasets, check-worthy content identification, and explainable detection.These insights have not been well utilized to date.
- Manual Fact-checking: Crowd-sourced fact-checking can scale through collective intelligence but is less credible and accurate, requiring user filtering and conflict resolution.Its management becomes more difficult as the number of fact-checkers grows.
- Automatic Fact-checking: Automatic fact-checking divides into fact extraction, or knowledge-base construction, and fact-checking, or knowledge comparison.The approach relies heavily on Information Retrieval, Natural Language Processing, Machine Learning, and network or graph theory.
- Discussion: Fake news spreads faster than true news, increasing the importance of fast news fact-checking.
- Discussion: Fast news fact-checking requires identifying check-worthy content and simplifying knowledge bases to retain valuable facts rather than merely maximizing fact count.Dynamic knowledge bases are also important because news concerns recent events and may require removing invalid knowledge or introducing new facts.
3 STYLE-BASED FAKE NEWS DETECTION
Style-based fake news detection examines quantifiable content style to distinguish fake from true news and assess possible intent to mislead. It represents articles with content features and applies classification models, with performance depending on both representation quality and classifier choice.
- Style-based detection analyzes news content style to differentiate fake from true news and can assess whether content reflects an intention to mislead.The approach assumes malicious entities may use a distinctive style to attract attention and build trust.
- Fake news style is defined as quantifiable characteristics that represent fake news content and differentiate it from true news content.
- The task is commonly formulated as binary or multi-label classification using a feature vector representing an article’s content style.
- Training data contain feature vectors and known labels, which help estimate the parameters of the style-based prediction function.
- Performance depends on how well textual and visual styles are represented and how effectively classifiers use those representations.
- The survey reviews content-style patterns and discusses style-based detection methods alongside their representation and classification components.
3.1 Style Representation
Style representation converts news text and images into quantifiable features across language and visual levels. The survey covers general and latent representations, including frequency statistics, syntax and discourse structures, semantic attributes, embeddings, and image features.
- Style features are grouped into textual and visual features, representing news text and images respectively.
- Textual features include general and latent representations spanning lexicon, syntax, semantic, and discourse levels.
- Textual features: General textual features use frequency statistics, n-grams, POS tags, and PCFG rewrite rules to represent lexical and syntactic patterns.PCFG representations can encode unlexicalized or lexicalized rules, with or without grandparent nodes.
- Textual features: Discourse and semantic representations count rhetorical relations and psycho-linguistic categories, including ten dimensions such as uncertainty, subjectivity, sentiment, diversity, informality, and readability.
- Textual features: Frequency can be measured as absolute, standardized, or TF-IDF-based relative frequency, with standardization removing content-length effects.
- Latent features: Latent text representations embed words, sentences, or documents as vectors that can feed traditional or neural models.
- Research coverage: The survey notes that matrix or tensor factorization has rarely been considered in style-based fake news detection, while relatively few studies explore news images.
- Visual features: Image representations use hand-crafted visual features or pixel matrices and tensors processed by neural networks.Image tensors are described by width × height × number of channels, including grayscale or RGB channels.
3.2 Style Classification
Style classification uses manually selected or learned representations of text and images with traditional machine-learning or deep-learning models. Reported comparisons favor non-latent and cross-level feature combinations in traditional settings, while multimodal models exploit relationships among content modalities.
- Traditional machine learning: Traditional machine-learning systems use manually selected latent and non-latent features extracted from news text or images.
- Traditional machine learning: Supervised classifiers used for style-based detection include SVMs, Random Forests, and XGBoost, alongside semi-supervised and unsupervised possibilities.
- Traditional machine-learning results: Traditional ML results indicate that non-latent features often outperform latent features, while combining levels can outperform single-level features.
- Traditional machine-learning results: Standardized lexicon and rewrite-rule frequencies better represent fake-news style and perform better than other feature groups, although they take longer to compute.
- Caveat: It is unjustified to identify one universally best algorithm because classifiers perform best in settings for which they were designed.
- Deep learning: Deep-learning systems embed text as word representations or images as pixel tensors, then use CNNs, RNNs, or Transformers to extract latent features.
- Multimodal models: EANN combines multimodal extraction, event-invariant representation learning through an event discriminator, and true-or-fake classification.
- Multimodal models: SAFE detects fake news by modeling relationships between textual and visual features, motivated by gaps between fake stories and their accompanying images.
3.3 Patterns of Fake News Content Style
Studies report distinguishable stylistic differences between fake and true news in both text and images. Fake news text is more informal, diverse, subjective, and emotional, while its images show different clarity, coherence, diversity, and clustering patterns.
- These textual and visual patterns have been reported in recent fake-news detection studies.
- Fake news text has higher informality, diversity, subjectivity, and emotionality than true news text.The measures correspond to swear words, unique verbs, report verbs, and emotional words, respectively.
- Fake news images often have higher clarity and coherence but lower diversity and clustering scores than true news images.
3.4 Discussion
Style-based methods represent and classify textual or visual writing patterns, but their reliance on content makes them vulnerable to manipulated styles and domain or language shifts.
- Style-based fake news detection represents and classifies textual and/or visual features of news content.
- Fake news style patterns vary across domains and languages, evolve over time, and can affect prediction performance.
- Writing style can be manipulated, creating a cat-and-mouse dynamic between style-based detection and future countermeasures.
- Content-based detection can enable identifying fake news before it propagates on social media.
- Propagation-based methods can add social context to improve robustness against writing-style manipulation.
4 PROPAGATION-BASED FAKE NEWS DETECTION
Propagation-based detection classifies news cascades or self-defined propagation graphs using structural, temporal, behavioral, and network features. Studies report distinctive diffusion patterns for fake news, while limited dissemination information constrains early detection.
- Propagation-based detection uses dissemination information to classify news cascades or self-defined propagation graphs.
- News cascades represent propagation as tree-like structures measured through depth, breadth, size, lifetime, and heat.
- Traditional machine-learning methods use engineered cascade features with classifiers including SVMs, decision trees, decision rules, naïve Bayes, and random forests.
- Fake news generally spreads faster, farther, more widely, and with greater depth, breadth, size, and structural virality than true news.
- Fake-news spreader networks show more spreaders, stronger engagement, and denser networks than true-news spreader networks.
- Propagation-based detection is more robust to writing-style manipulation but is inefficient before dissemination or when social information is limited.
- Ground-truth dependence and unclear annotation practices remain concerns for supervised propagation-based detection.
5 SOURCE-BASED FAKE NEWS DETECTION
Source-based detection assesses the credibility of authors, publishers, and social-media accounts across the news life cycle. Network homogeneity, publisher communities, web-spam signals, and bot behavior provide source-level evidence.
- Source-based detection evaluates credibility at creation, publication, and social-media propagation stages.
- Source credibility is efficient to assess but may classify true articles from unreliable sources as fake news.
- Authors and publishers can be assessed through relationships with other authors or publishers and patterns in their networks.
- Coauthorship networks exhibit homogeneity, with denser connections among authors in the same true-news, fake-news, or mixed-news group.
- Content-sharing networks organize publishers into Russian/conspiracy, right-wing/conspiracy, U.S. mainstream, left-wing blog, and U.K. mainstream communities.
- Publisher credibility assessment includes web-ranking methods and content-, link-, and other web-spam detection techniques.
- Bots spread unreliable news earlier, while humans perform most retweeting and retweet bot-posted articles almost as often as human-posted articles.
6 DISCUSSION AND FUTURE WORK
The survey identifies future work spanning joint detection, early detection, check-worthy content, cross-domain analysis, explainability, intervention, and non-traditional fake news. These directions address evolving knowledge and styles, limited early evidence, and the need for interpretable systems.
- The four detection strategies emphasize different aspects and can be combined into more comprehensive detection strategies.
- I. Detection of Non-traditional Fake News: Non-traditional fake news includes outdated knowledge and content that is only partially correct, motivating detection beyond binary false-news labels.
- II. Fake News Early Detection: Early detection must handle outdated knowledge, evolving deceptive styles, and limited information by improving ground-truth timeliness, feature compatibility, and verification efficiency.
- III. Identifying Check-worthy Content: Identifying check-worthy content or topics can prioritize verification, while identifying check-worthy portions can support explainable detection.
- IV. Cross-domain (-topic, -website, -language) Fake News Analysis: Cross-domain, cross-topic, cross-website, and cross-language analysis can reveal non-varying characteristics relevant to early detection and check-worthy identification.
- V. Explainable Fake News Detection: Explainability can draw on social feedback, expert analyses, and interdisciplinary theories, though these resources have been rarely utilized.
- VI. Fake News Intervention: Intervention research considers network structure, influential users, technical innovations, and policies that increase information quality.
7 CONCLUSION
The survey synthesizes fake news research by defining the phenomenon, identifying interdisciplinary theories, reviewing detection methods across four perspectives, and highlighting research challenges and opportunities.
- The survey defines fake news using authenticity, intention, and being news, while distinguishing it from related forms such as deceptive, false, satire, misinformation, and rumors.
- It comprehensively identifies fundamental theories from disciplines including the social sciences to support interdisciplinary fake news research.
- Detection methods are reviewed across false knowledge, writing style, propagation patterns, and source credibility.
- The survey highlights challenges in current research and outlines corresponding research opportunities.
- An online repository provides summaries, timely updates, tutorials, publications, methods, datasets, and related resources.