This Just In: Fake News Packs a Lot in Title, Uses Simpler, Repetitive Content in Text Body, More Similar to Satire than Real News

Authors: Benjamin D. Horne, Sibel Adalı Venue: AAAI, 2017 — arXiv:1703.09398

TL;DR

Fake news differs systematically from real news in stylistic features, especially titles. Contrary to the assumption that fake news mimics real news, it is actually more similar to satire: packed titles with fewer stop words and more proper nouns, combined with shorter, repetitive body text with simpler language. This suggests fake news targets audiences unlikely to read beyond headlines, relying on heuristics rather than argument strength for persuasion.

Contributions

• Empirical characterization of stylistic and linguistic differences between fake, real, and satire news across three datasets using 88 features (stylistic, complexity, psychological).
• Finding that fake news is more similar to satire than to real news, despite the common assumption that fake news is designed to deceive by looking like real news.
• Demonstration that fake titles are the key differentiator between fake and real news (78% classification accuracy on titles alone vs. 71% on body text).
• Evidence supporting the Elaboration Likelihood Model (ELM): real news persuades through central route (argument strength), while fake news targets peripheral route (heuristics, mental associations).
• Release of novel datasets: Buzzfeed's 2016 election data (71 stories) and the authors' own political news dataset (225 stories from real, fake, and satire sources).

Method

Feature engineering: The authors compute 88 hand-crafted features across three categories:

• Stylistic features: Part-of-speech tag frequencies, stop words, punctuation, quotes, negations, informal/swear words, interrogatives, capitalized words (from LIWC dictionaries and NLTK).
• Complexity features: Readability indices (Gunning Fog, SMOG, Flesch-Kincaid), type-token ratio, word-level fluency (frequency of words in COCA corpus), and syntactic tree depth from Stanford Parser.
• Psychological features: LIWC-based counts of analytic, insightful, causal, certain, tentative words, and sentiment strength (SentiStrength) for positive/negative emotion.

Statistical analysis: One-way ANOVA (for normally distributed features) and Wilcoxon rank-sum test (for non-normal) to identify significant differences (p < 0.05) between fake, real, and satire articles.

Classification: Linear kernel SVM with 5-fold cross-validation using top 4 features from hypothesis testing (body: word count, TTR, nouns, quotes; title: stopwords, nouns, avg word length, readability).

Results

Content differences (body text)

Real news articles are significantly longer than fake news and use: - More technical/complex vocabulary - More punctuation and quotes - Fewer adverbs and less lexical redundancy - More nouns, fewer personal pronouns - Higher readability grade level (more demanding to read)

Fake news uses simpler, more redundant language with: - Lower TTR (Type-Token Ratio) — high word repetition - More personal pronouns (we, you, us) - More adverbs and fewer analytical words - Lower readability score (easier to read, but less informative)

SVM classification accuracy: 71% (Dataset 2) and 77% (Dataset 1) on body text alone.

Title differences

Fake titles are longer than real titles but more tightly packed with claims: - Fewer stop words (the, is, on) — proportion of stopwords is low - More proper nouns (NNP) — entities and names emphasized - Fewer nouns overall — fewer generic nouns, focusing instead on named entities - More capitalized words (ALL CAPS emphasis) - Simpler word length and lower readability grade level - More verb phrases and past-tense words

Real titles present a single main claim concisely. Fake titles use many verb phrases and entity names to pack multiple claims into the headline.

Example: - FAKE: "BREAKING BOMBSHELL: NYPD Blows Whistle on New Hillary Emails: Money Laundering, Sex Crimes with Children, Child Exploitation, Pay to Play, Perjury" - REAL: "Preexisting Conditions and Republican Plans to Replace Obamacare"

SVM classification accuracy on titles: 78% (Dataset 2) and 71% (Dataset 1).

Fake vs. satire similarity

Fake news is significantly more similar to satire than to real news. Both use: - Simpler, fewer technical words - More adverbs, fewer nouns - Fewer quotes and punctuation - Higher lexical redundancy (repeated words)

However, fake and satire differ when they diverge — the distinction "simply exaggerates satire's differences with real news further" (i.e., fake is an extreme version of satire's style).

Satire vs. Real classification accuracy: 91% (body), 75% (title). Fake vs. Satire: 67% (body), 55% (title).

Connections

• 2023 — HERO: Hierarchical Linguistic Trees for Fake News Detection extends this work with structure-aware neural networks that model syntactic and discourse hierarchies.
• Linguistic style detection — topic synthesis covering style-based detection methods.
• Content-based detection — broader paradigm of which this work is a foundational example.
• 2017 — Potthast et al.: A Stylometric Inquiry into Hyperpartisan and Fake News — concurrent stylometric analysis using different feature sets.
• 2017 — Shu et al.: Fake News Detection on Social Media: A Data Mining Perspective — contemporaneous survey of detection approaches.
• Satire detection — related topic with overlap in fake/satire distinction.

Notes

Strengths: - Clear motivation: challenges a common misconception that fake news mimics real news. - Large feature set (88 features) capturing multiple linguistic levels; systematic statistical testing. - Inclusion of satire as a third class reveals structural similarities and differences. - Practical insight: fake news targets readers who skim titles; counter-misinformation should emphasize title clarity. - Novel, public datasets for 2016 U.S. election fake news.

Limitations: - Datasets are small (71 and 225 stories in Datasets 1 and 2) and election-focused. The larger Burfoot & Baldwin dataset (4,233 articles) is non-political. - Dataset 1 from Buzzfeed may have selection bias (high-engagement articles only). - Linear SVM with only 4 features is simple; modern deep learning approaches may capture non-linear patterns better. - Features are surface-level (bag-of-words, readability indices); syntactic and semantic structure not fully modeled until later work (Linguistic-style-aware Neural Networks for Fake News Detection). - No cross-domain evaluation; generalization to news domains outside politics/satire untested.

Significance: This is a foundational empirical study demonstrating that fake news has a distinct stylistic fingerprint. The discovery that fake news resembles satire more than real news is theoretically important: it suggests the deception mechanism relies on heuristic persuasion (peripheral route via attention-grabbing titles and entity associations) rather than argumentative content. This framing influenced subsequent work on persuasion, psychological vulnerabilities, and title-focused mitigation strategies.