Hoax: Definition, Detection, and Impact

• Hoax is false or misleading content presented as fact, often blending real details with fabrication to deceive audiences.
• Studies reveal that hoaxes exploit social heuristics and network structures, rapidly spreading across online platforms during high-volume events.
• Advanced computational methods, including content analysis and interaction graphs, have proven effective in detecting hoaxes across diverse domains.

A hoax is false or misleading content, or in some domains a false substitute object, presented as legitimate fact or authenticity. In the false-information literature, it is defined as a news story that contains facts that are false or inaccurate and is presented as legitimate fact; adjacent literatures describe it as false or misleading information presented as fact, often via repurposed or manipulated content, while Wikipedia research treats it as an intentionally false or misleading article about a fictitious topic (Zannettou et al., 2018, Zubiaga et al., 2013, Borkakoty et al., 2024). The term is therefore best understood as a family of operational categories linked by deceptive verisimilitude rather than by a single universal ontology.

1. Definitions, typologies, and boundaries

Within the web false-information ecosystem, hoaxes are positioned between several neighboring categories. They differ from fabricated news in that hoaxes often retain some real-world context or plausible detail; they differ from rumors because rumors may remain unresolved, whereas hoaxes are asserted as facts and, once debunked, are verifiably false; and they differ from clickbait because clickbait may be misleading in presentation while the underlying payload remains true (Zannettou et al., 2018). In the same taxonomy, satire is separated by lack of intent to deceive, whereas propaganda is oriented toward harming or manipulating a target (Ghanem et al., 2019).

Operational usage varies sharply by domain, and that variation matters methodologically because ground truth, error costs, and detection features change with the object being classified.

Domain	Operational meaning	Source
Online false information	False or inaccurate story presented as legitimate fact	(Zannettou et al., 2018)
Social media crisis content	False or misleading information, often including manipulated or out-of-context images	(Zubiaga et al., 2013)
Wikipedia	Intentionally false or misleading article about a fictitious or fabricated topic	(Borkakoty et al., 2024)
E-commerce reviews	Deceptive or illegitimate review misleading customers about product quality	(Dinh et al., 2022)
Treaty verification	Non-authentic object submitted in place of a genuine warhead	(Vavrek et al., 2017)

Several studies make the definition explicitly operational rather than philosophical. In Facebook hoax detection, posts are labeled by page category: scientific pages are assumed to publish reliable content and conspiracy pages hoaxes, yielding a page-based proxy for post-level ground truth without per-post fact-checking (Tacchini et al., 2017). In HoaxItaly, “hoax” is subsumed under a broader disinformation class based on blacklists compiled by Italian fact-checking sites (Pierri et al., 2020). In Wikipedia, known hoaxes are identified from administrative lists and deletion discussions rather than from intrinsic semantic criteria alone (Elebiary et al., 2023, Borkakoty et al., 2024). This suggests that “hoax” in computational work is frequently a task definition tied to annotation protocol, not merely a lexical label.

2. Social, cognitive, and network conditions of hoax diffusion

Hoaxes exploit specific epistemic and network conditions. During fast-paced events, users face high-volume streams, limited verification time, and strong reliance on social heuristics. A study of Hurricane Sandy image credibility found that showing multiple tweets corroborating a fact misled users into trusting what was actually a hoax: 71% of fake items received a higher corroboration rating than their single-text plausibility, and fake pictures were at least as popular as real ones, with median shares 261 for fakes versus 216 for reals (Zubiaga et al., 2013). The same study found that richer author details were more diagnostically useful than presentation quality, while grammar and style were weak cues.

This empirical pattern is consistent with the broader literature on social proof and the illusory truth effect. Survey work on false information reports that hoaxes seed in receptive communities, are boosted by botnets or coordinated networks, exploit algorithmic ranking and low-friction sharing, and can jump from fringe communities to mainstream platforms (Zannettou et al., 2018). The same survey notes that false stories propagate faster, farther, and more broadly than true stories on Twitter, with especially strong effects for political content (Zannettou et al., 2018).

Network structure also matters. A compartmental model of misinformation and fact-checking divides agents into susceptible, hoax-believer, and fact-checker states, with homophily-induced segregation between gullible and skeptic groups (Tambuscio et al., 2016). Its central finding is narrow but important: a more segregated network facilitates the spread of a hoax only at low forgetting rates, but has no effect when agents forget at faster rates (Tambuscio et al., 2016). This formalizes a recurrent intuition in empirical work on echo chambers and polarization.

Real-time observatories reveal a further asymmetry between hoax diffusion and correction. Hoaxy, which tracks misinformation and fact-checking URLs on Twitter, found that fact-checking content typically lags misinformation by 10–20 hours, with the aggregate lagged cross-correlation peaking at approximately $-13$ hours, meaning misinformation leads (Shao et al., 2016). HoaxItaly shows a similar imbalance at ecosystem scale: less than 5% of total tweets contain a link to fact-checking stories and less than 10% of shared articles belong to the fact-checking class (Pierri et al., 2020).

Emotional and semantic framing can reinforce these dynamics. In popular English tweets about COVID-19 vaccines, the semantic neighborhood of “hoax” was described as highly clustered, linked to “censor,” “pandemic,” and “vaccine,” and devoid of negations of meaning, indicating that conspiratorial framing appeared inside high-visibility content rather than only at the fringe (Stella et al., 2021). A plausible implication is that hoax discourse is sustained not only by raw exposure but by stable semantic micro-communities that normalize the claim structure itself.

3. Computational detection and early classification

Automated hoax detection has developed along three principal feature families: content, interaction structure, and temporal behavior. A canonical network-based example is Facebook post classification from user “likes.” In that setting, each post $i$ is represented by binary user-like indicators $x_{iu}$, and logistic regression models the probability that the post is non-hoax as

$$p_i = \frac{1}{1 + e^{-y_i}}, \qquad y_i = \sum_{u \in U} x_{iu} w_u.$$

On 15,500 Facebook posts and 909,236 users, this approach and an adapted boolean crowdsourcing method both achieved accuracies exceeding 99%, even when the training set contained less than 1% of posts (Tacchini et al., 2017). The same study reported that the harmonic boolean crowdsourcing variant reached 99.4% in 80/20, 5-fold cross-validation and transferred across unseen pages with average accuracy 0.991 in one-page-out evaluation (Tacchini et al., 2017).

Early detection has been studied at larger temporal scale on Twitter. For celebrity death reports, a semi-automatically constructed dataset of 4,007 reports and 10,301,996 tweets supported three-way classification into real, commemoration, and fake. Class-specific word embeddings combined with logistic regression reached macro-F1 approximately 0.716 within 10 minutes of the first tweet, 0.707 at 5 minutes, and 0.647 from only the first tweet (Zubiaga et al., 2018). Notably, using the full prefix $[t_0,t]$ consistently outperformed smaller sliding windows, providing no support for a strong self-correction signal dominating later content (Zubiaga et al., 2018).

Multilingual semi-automated verification systems couple retrieval with inferential filtering. FacTeR-Check embeds new claims and known hoaxes using multilingual sentence-transformer ensembles, ranks candidates by cosine similarity,

$$\mathrm{CosSim}(u,v)=\frac{\langle u,v\rangle}{\|u\|\|v\|},$$

and then applies natural language inference to classify entailment, contradiction, or neutrality (Martín et al., 2021). The system reports strong module-level results, including average semantic-similarity correlations $r=84.24$, $\rho=84.71$ for its best ensemble+PCA setting on MSTSB, and NLI accuracy 87.77% on English SICK and 82.92% on a Spanish test translation (Martín et al., 2021). Its tracking workflow was used to analyze 61 COVID-19-related hoaxes over time (Martín et al., 2021).

Content-only models remain viable but domain dependent. An emotionally infused LSTM for false-information subtype classification improved macro-F1 from 72.26 to 79.43 on news articles and from 55.41 to 59.70 on Twitter, with hoax characterized by subtler emotional cues such as hope, anticipation, like, and in some domains disgust (Ghanem et al., 2019). For Vietnamese e-commerce spam reviews, where hoaxing customers is operationalized as deceptive or illegitimate review content, PhoBERT achieved macro-average F1 86.89% on binary spam detection and 72.17% on spam-type classification (Dinh et al., 2022). These results suggest that “hoax detection” is not a single task family but a class of supervised and semi-supervised problems whose discriminative signal may come from users, language, metadata, or graph structure.

4. Wikipedia and the problem of encyclopedic hoaxes

Wikipedia hoaxes are a distinct subtype because they are designed to satisfy institutional style constraints. Hoaxpedia’s surface-level analysis shows that hoax and legitimate articles are strikingly similar on shallow stylistic measures: median article lengths 1,057 versus 1,777 words, median sentence lengths 22.0 versus 21.23 words, median word lengths 4.35 versus 4.36 characters, and median Flesch–Kincaid grade levels 9.5 versus 9.4 (Borkakoty et al., 2024). This explains why content-only detection is difficult.

Nevertheless, the task is feasible. Hoaxpedia compiles 311 verified hoax articles together with semantically similar legitimate negatives and reports that RoBERTa-large on full text achieved hoax-class F1 scores of 0.84, 0.81, and 0.79 for $1\mathrm{H}:2\mathrm{R}$, $1\mathrm{H}:10\mathrm{R}$, and $i$0 settings, respectively (Borkakoty et al., 2024). Definition-only models performed substantially worse, implying that discriminative content is distributed through the article body rather than concentrated in the lead (Borkakoty et al., 2024).

Attention-based analyses approach the problem from the opposite direction: not how to detect a hoax once written, but when a topic is vulnerable to hoax creation. A study of English Wikipedia defined online attention as page views to semantically related “neighbor” pages in a 14-day window around article creation and used

$i$1

to compare the median traffic in the 7 days before and after creation (Elebiary et al., 2023). Relative to legitimate same-day cohorts, hoaxes had more pre-creation attention: $i$2 for 75 out of 83 hoaxes, with sample mean approximately 0.123 and a 95% bootstrapped confidence interval of $i$3 (Elebiary et al., 2023). This suggests a demand-side mechanism in which attention spikes create conditions for disinformation supply.

The same work situates its sample among longer-lived and notable cases: 92% of hoaxes are detected within the first day in prior work, yet one in a hundred can remain undetected for more than a year (Elebiary et al., 2023). That asymmetry clarifies why Wikipedia hoaxes are operationally important despite low prevalence: survivorship and institutional credibility amplify their downstream effect.

5. Extended senses: academic stunts, scientific misconceptions, deceptive artifacts, and review fraud

The term “hoax” also extends beyond online false news to deception against disciplinary or institutional norms. In the Science Wars literature, Alan Sokal’s 1996 submission to Social Text is treated as a deliberate deception staged to expose lax editorial standards; it was published without review and later revealed as a hoax (Boughn, 2023). In that setting, the hoax differs from satire because it seeks to be taken at face value, and from fraud because its stated aim was to stress-test gatekeeping rather than to advance false knowledge for personal gain (Boughn, 2023).

A different extension appears in astrophysics, where the persistent belief that spiral-galaxy rotation curves are universally flat is labeled a hoax in the sense of a widely repeated but false claim that survived after specialists had abandoned it (Salucci, 2010). Here the term names a sociologically entrenched misconception rather than a fabricated story. The underlying argument is evidentiary: thousands of high-quality rotation curves show luminosity-dependent diversity—rising, approximately flat, or mildly declining profiles—rather than universal flatness (Salucci, 2010).

In treaty verification, a hoax is not a claim but a physical object. Nuclear resonance fluorescence work defines a hoax object as a non-authentic item submitted in place of a genuine warhead, including isotopic hoaxes that replace depleted uranium with lead while matching areal density (Vavrek et al., 2017). The proof-of-concept system distinguished genuine and hoax proxy objects with high confidence, reporting discrepancies as large as 10.7 standard deviations in summed U-238 NRF peak rates (Vavrek et al., 2017). This is a materially different usage, but the underlying logic remains deception through plausible substitution.

Consumer-review fraud offers another operationalization. In Vietnamese e-commerce, hoaxing customers is defined as creating untruths and illegitimate reviews that misrepresent product quality or distract from product-specific evidence (Dinh et al., 2022). The dataset partitions spam into fake reviews, brand-only reviews, and non-reviews, showing that “hoax” can be decomposed into subtype-specific annotation rules rather than treated as a unitary label (Dinh et al., 2022).

6. Limitations, ethics, and open problems

The study of hoaxes is constrained by labeling, generalization, and intervention risk. Page-based Facebook labeling assumes all posts from conspiracy pages are hoaxes and all from scientific pages non-hoaxes, which may introduce mislabeling and limit generality (Tacchini et al., 2017). Domain-based systems such as Hoaxy and HoaxItaly inherit biases from curated site lists and miss linkless, screenshot-based, or untracked-domain circulation (Shao et al., 2016, Pierri et al., 2020). Wikipedia datasets are biased toward discovered, administratively notable, or longer-lived hoaxes (Elebiary et al., 2023, Borkakoty et al., 2024). FacTeR-Check depends on the completeness of existing fact-check repositories, so novel hoaxes remain hard cases by design (Martín et al., 2021).

Privacy and fairness concerns recur across platforms. Facebook-like based classifiers use public interaction data but do not describe anonymization beyond aggregate reporting (Tacchini et al., 2017). HoaxItaly releases only tweet IDs in accordance with Twitter’s terms, yet the underlying problem of politically concentrated diffusion remains salient (Pierri et al., 2020). The use of interaction patterns and polarization for classification raises the possibility that demographic or ideological skews become encoded into moderation decisions (Tacchini et al., 2017).

Adversarial robustness remains underexplored. Several studies note that coordinated manipulation or bots could bias systems relying on likes, retweets, or semantic retrieval, but do not provide explicit adversarial evaluations (Tacchini et al., 2017, Zannettou et al., 2018). Cross-platform and multimodal generalization also remain open: social-media hoaxes may be text-light, image-heavy, or driven by screenshots and video, while many strong reported results are platform- or task-specific (Zannettou et al., 2018, Martín et al., 2021).

A plausible synthesis is that reliable hoax analysis requires hybrid architectures: interaction graphs for community structure, content models for semantic and stylistic anomaly, temporal models for early warning, and human review for context-sensitive adjudication. That synthesis is directly suggested by the literature rather than fully realized within it (Tacchini et al., 2017, Zannettou et al., 2018, Borkakoty et al., 2024).