The fingerprints of misinformation: how deceptive content differs from reliable sources in terms of cognitive effort and appeal to emotions

Not all misinformation is created equal. It can adopt many different forms like conspiracy theories, fake news, junk science, or rumors among others. However, most of the existing research does not account for these differences. This paper explores the characteristics of misinformation content compared to factual news—the “fingerprints of misinformation”—using 92,112 news articles classified into several categories: clickbait, conspiracy theories, fake news, hate speech, junk science, and rumors. These misinformation categories are compared with factual news measuring the cognitive effort needed to process the content (grammar and lexical complexity) and its emotional evocation (sentiment analysis and appeal to morality). The results show that misinformation, on average, is easier to process in terms of cognitive effort (3% easier to read and 15% less lexically diverse) and more emotional (10 times more relying on negative sentiment and 37% more appealing to morality). This paper is a call for more fine-grained research since these results indicate that we should not treat all misinformation equally since there are significant differences among misinformation categories that are not considered in previous studies.


Introduction
H ow can we mitigate the spread of disinformation and misinformation? This is one of the current burning questions in social, political and media circles across the world (Kietzmann et al., 2020). The answer is complicated because the existing evidence shows that the prominence of deceptive content is driven by three factors: volume, breadth, and speed. While the access to information has been dramatically increasing since the advent of internet and social networks, the volume of misleading and deceptive content is also on the rise (Allcott et al., 2018); that is the volume problem. Also, deceptive content can adopt different forms. Misinformation can appear as rumors, clickbait or junk science (trying to maximize visitors to a webpage or selling "miraculous" products) or in the form of fake news or conspiracy theories (false information spread deliberately to affect political or social institutions) (Scheufele & Krause, 2019); that is the breadth problem. Finally, misinformation spreads six time faster than factual information (Vosoughi et al., 2018), that is the speed challenge. These three factors contribute to making the tackling of both disinformation and misinformation one of the biggest problems in our times.
Some of the prominent social media platforms have intended to stop the proliferation of misinformation with different features like relying on user's reporting mechanisms (Chan et al., 2017;Lewandowsky et al., 2012) or using fact-checkers to analyze content that already went viral through the network (Chung & Kim, 2021;Tambuscio et al., 2018;Tambuscio et al., 2015). However, both approaches have some shortcomings. The assessment of third-party fact-checkers or user's reports is sometimes a slow process and misinformation often persist after being exposed to corrective messages (Chan et al., 2017). Therefore, a purely human-centered solution is ineffective because misinformation is created in more quantity (volume), in more forms (breadth) and faster (speed) than the human ability to fact-check everything that is being shared in a given platform. To address this issue, I propose to explore what I call "the fingerprints of disinformation". That is, how factual news differ from different types of misinformation in terms of (1) evocation to emotions (sentiment analysis-positive, neutral, negative-and appeal to moral language as a challenge to social identity), and (2) cognitive effort needed to process the content (both in terms of grammatical features-readability-and lexical featuresperplexity).
Regarding cognitive effort, extant research in the human cognition and behavioral sciences can be leveraged to identify misinformation through quantitative measures. For example, the information manipulation theory (McCornack et al., 2014) proposes that misinformation is expressed differently in terms of writing style. The main intuition is that misinformation creators have a different writing style seeking to maximize reading, sharing and, in general, maximizing virality. The limited capacity model of mediated motivated message processing (Lang, 2000(Lang, , 2006 states that in information sharing, structural features and functional characteristics require different cognitive efforts to be processed (Kononova et al., 2016;Leshner & Cheng, 2009;Leshner et al., 2010) and, since humans attempt to minimize cognitive effort when processing information, content that requires less effort to be processed is more engaging and viral (Alhabash et al., 2019).
Moreover, another fingerprint of misinformation is its reliance on emotions (Bakir & McStay, 2018;Kramer et al., 2014;Martel et al., 2019;Taddicken & Wolff, 2020). In general, content that evokes high-arousal emotions is more viral (Berger, 2011;Berger & Milkman, 2009;Berger & Milkman, 2013;Goel et al., 2015;Milkman & Berger, 2014), which explains why social networks are a source of massive-scale emotional contagion (Fowler & Christakis, 2009;Kramer et al., 2014;Rosenquistet al., 2011). One of the main reasons proposed to explain this behavior is the dualprocess theory of judgment stating that emotional thinking (in contrast to a more analytical thinking) hinders good judgment (Evans, 2003;Stanovich, 2005). Indeed, there is experimental evidence that engaging in analytic thinking reduces the propensity to share fake news (Bago et al., 2020;Gordon Pennycook et al., 2015;Gordon Pennycook & Rand, 2019). For example, encouraging people to think analytically, in contrast to emotionally, decreases likelihood of "liking" or sharing fake news (Effron & Raj, 2020). On the contrary, reliance on emotion is associated with misinformation sharing (Weeks, 2015) or believing in conspiracy theories (Garrett & Weeks, 2017;Martel et al., 2019). Similarly, tweets with negative content are retweeted more rapidly and frequently than positive or neutral tweets (Tsugawa & Ohsaki, 2017) and sentiment rather than the actual information content predicts engagement (Matalon et al., 2021). Moreover, beyond sentiment, there are other measures for evocation to emotions: morality. Although there is little research on how moral content contributes to online virality (Rathje et al., 2021), the mechanism is grounded in social identity theory (Tajfel & Turner, 1979) and self-categorization theory (Turner et al., 1987) and lies in the idea that group identities are hyper salient on social media (Brady et al., 2017) because they act as a form of self-conscious identity representation (Kraft et al., 2020;van Dijck, 2013). Therefore, content that appeals to what individuals think is moral (or not moral), is likely to be more viral (Brady et al., 2020), especially in conjunction with negative emotions that challenge their social identity (Brady et al., 2017;Horberget al., 2011). In contrast, reliable sources are forced to follow the principle of objectivity typical of good journalistic practice (Neuman et al., 1992).
This unprecedented fingerprint of misinformation provides evidence that content features differ significantly between factual news and different types of misinformation and therefore can facilitate early detection, automation, and the use of intelligent techniques to support fact-checking and other mitigation actions. More specifically, the novelty and benefits of the paper are fourfold: 1. Volume benefits: A solution must be scalable. This proposal is a highly scalable technique that relies on psychological theories and Natural Language Processing methods to discern between factual news and misinformation types of content. Summarizing, I propose that misinformation has different complexity levels (in terms of lexical and grammatical features) that require different levels of cognitive effort and that misinformation evokes higharousal emotions and a higher appeal to moral values. Using the mentioned variables, we can quickly classify every content shared in social networks at the exact moment when it is posted with little human intervention, helping to mitigate the volume challenge. 2. Breadth benefits: Not all disinformation is created equal.
While extant research has been dedicated to differentiating between factual and fake news aiming at binary classification (Choudhary & Arora, 2021;de Souza et al., 2020), there is not much evidence regarding rumors, conspiracy theories, junk science or hate speech altogether. In this paper, I propose a model that can differentiate between 7 different categories of content: clickbait, conspiracy theories, fake news, hate speech, junk science, rumors, and finally, factual sources. To my knowledge, this is the paper with higher number of categories and higher number of news articles.
3. Time benefits: Detecting misinformation before it is too late with on-spot interception. Usually, fact-checkers and platform create lists and rankings of content (usually URLs) that are viral to assess its veracity. However, this procedure has a problem: the content is fact-checked once it has gone viral. In this paper I propose a method that is independent of network behavior, information cascades or their virality. Therefore, it allows to identify misinformation before it spreads through the network. 4. Explainability: Avoiding inmates running the asylum. In contrast to previous approaches (Hakak et al., 2021;Mahir et al., 2019;Manzoor et al., 2019), I provide an explainable model that is well-justified and grounded in common methods. In other words, I employ a set of mechanics that are easily explainable in human terms. This is important because this type of model have been rarely available (Miller et al., 2017). This model allows not only researchers but also practitioners and non-technical audiences to understand, and potentially adopt the model.
In particular, I analyze 92,112 news articles with a median of 461 words per article (with a range of 201 to 1961 words and 58,087,516 total words) to uncover the features that differentiate factual news and six types of misinformation categories using a multinominal logistic regression (see Methods). For this, I assess the importance of quantitative features grounded in psychological and behavioral sciences through an operationalization based on Natural Language Processing. Specifically, I assess the readability, perplexity, evocation to emotions, and appeal to morality of the 92,112 articles. Among others, these results show that fake news and conspiracy theories are, on average, 8% simpler in terms of readability and 18% simpler in terms of perplexity, and 18% more reliant on negative emotions and 45% more appealing to morality than factual news.

Data.
In order to carry out the analysis I use the Fake News Corpus (Szpakowski, 2018), comprised of 9.4 million news items extracted from 194 webpages. Beyond the title and content for each item, the corpus also categorizes each new into one of the following categories: clickbait, conspiracy theories, fake news, hate speech, junk science, factual sources, and rumors. The category for each website is extracted from the OpenSources project (Zimdars, 2017). In particular, the definitions for each category are: Clickbait. Sources that provide generally credible content, but use exaggerated, misleading, or questionable headlines, social media descriptions, and/or images.
Conspiracy Theory. Sources that are well-known promoters of kooky conspiracy theories.
Fake News. Sources that entirely fabricate information, disseminate deceptive content, or grossly distort actual news reports.
Hate News. Sources that actively promote racism, misogyny, homophobia, and other forms of discrimination.
Junk Science. Sources that promote pseudoscience, metaphysics, naturalistic fallacies, and other scientifically dubious claims.
Reliable. Sources that circulate news and information in a manner consistent with traditional and ethical practices in journalism.
Rumor. Sources that traffic in rumors, gossip, innuendo, and unverified claims.
The labeling of each website was done through crowdsourcing following the instructions as follows (Zimdars, 2017): Step 1: Domain/Title analysis. Here the crowdsourced participants look for suspicious domains/titles like "com.co".
Step 2: About Us Analysis. The crowdsourced participants are asked to Google every domain and person named in the About Us section of the website or whether it has a Wikipedia page with citations.
Step 3: Source Analysis. If the article mentions an article or source, participants are asked to directly check the study or any cited primary source. Then, they asked to assess if the article accurately reflects the actual content.
Step 4: Writing Style Analysis. Participants are asked to check if there is a lack of style guide, a frequent use of caps, or any other hyperbolic word choices.
Step 5: Esthetic Analysis. Similar to the previous step, but focusing in the esthetics of the website, including photo-shopped images.
Step 6: Social Media Analysis. Participants are asked to analyze the official social media users associated with each website to check if they are using any of the strategies listed above.
Here, it is important to note that I follow a similar approach as previous studies to classify sources (Broniatowski et al., 2022;Cinelli et al., 2020;Cox et al., 2020;Lazer et al., 2018;Singh et al., 2020). For example, Bovet and Makse (2019) use Media Bias Fact Check to classify tweets according to their crowdsourced classification instead of manually classifying tweet by tweet. This approach has two advantages. First, that it is scalable (Broniatowski et al., 2022) and, secondly, misinformation intent is better captured at the source level than at the article level (Grinberg et al., 2019). Being aware of the potential limitations of this method, the approach offers a benefit: being able to tackle the breadth problem.
Moreover, the fact that I include several misinformation categories is important because most of the existing research has a strong emphasis on distinguishing between fake news and factual news (de Souza et al., 2020;Helmstetter & Paulheim, 2018;Masciari et al., 2020;Zervopoulos et al., 2020). However, not all misinformation is created equal. In general, it is accepted that there are several categories of misinformation delimited by its authenticity and intent. Authenticity is related to the possibility of fact-checking the veracity of the content (Appelman & Sundar, 2016). For example, a statistical fact is easily checkable. However, conspiracy theories are non-factual, meaning we are not able to fact-check their veracity. On the other hand, intent can vary between mislead the audience (fake news or conspiracy theories), attract website traffic (clickbait) or undefined intent (rumors) (Tables 1 and 2).
In the Fake News Corpus, each website is categorized among one of the options and all their articles have the corresponding category. From there, I extracted 30,000 random articles from each category, generating a dataset of 210,000 misinformation articles. For factual news, I used Factiva to download articles from The New York Times, The Wall Street Journal and The Guardian. This resulted in 3177 articles. In total, the database consists of 213,177 articles. I filtered the articles with less than 200 words and those with more than 2000 words, ending up with a database of 147,550 articles. After calculating all the measures (readability scores, perplexity, appeal to morality and sentiment analysis), I deleted all the outliers (lower bound quantile = 0.025 and upper bound quantile = 0.975). This resulted in the final dataset consisting of 92,112 articles with the following distribution by type: clickbait (12,955 articles), conspiracy theories ( In comparison to other datasets, the one used in this paper includes more sources and more articles than any other: Computational linguistics. Extant research in the human cognition and behavioral sciences can be leveraged to identify misinformation online through quantitative measures. For example, the information manipulation theory (McCornack et al., 2014) or the four-factor theory (Zuckerman et al., 1981) propose that misinformation is expressed differently in terms of arousal, emotions or writing style. The main intuition is that misinformation creators have a different writing style seeking to maximize reading, sharing and, in general, maximizing virality. This is important because views and engagement in social networks are closely related to virality, and being repeatedly exposed to misinformation increases the likelihood of believing in false claims (Bessi et al., 2015;Mocanu et al., 2015).
Based on the information manipulation theory and the fourfactor theory, I propose several parametrizations that can allow to statistically distinguish between factual news and a myriad of misinformation categories. To do so, I calculate a set of quantifiable characteristics that represent the content of a written text and allow us to differentiate it across categories. More specifically, following the definition proposed by (Zhou & Zafarani, 2020) the style-based categorization of content is formulated as a multinominal classification problem. In this type of problem, each text in a set of news articles Ν can be represented as a set of k features denoted by the feature vector f 2 R k . Through Natural Language Processing, or computational linguistics, I can calculate this set of k features for N texts. In the following subsection I describe these features and their theoretical grounding.
Measuring cognitive effort through grammatical features: readability. Using sentence length or word syllables as a measure of text complexity has a long tradition in computational linguistics (Afroz et al., 2012;Fuller et al., 2009;Hauch et al., 2015;Monteiro et al., 2018). Simply put, the longer a sentence or word is, the more complex it is to read. Therefore, longer sentences and longer words require more cognitive effort to be effectively processed. The length of sentences and words are precisely the fundamental parameters of the Flesch-Kincaid index, the readability variable.
Flesch-Kincaid is used in different scientific fields like pediatrics (D'Alessandro et al., 2001), climate change (De Bruin & Granger Morgan, 2019), tourism (Liu & Park, 2015) or social media (Rajadesingan et al., 2015). This measure estimates the educational level that is needed to understand a given text. The Flesch-Kincaid readability score is calculated with the following formula (Kincaid et al., 1975 where n w is the number of words, n st is the number of sentences, and n sy is the number of syllables. In this case, n w and n st act as a proxy for syntactic complexity and n sy acts as a proxy for lexical difficulty. All of them are important components of readability (Just & Carpenter, 1980). While the Flesch-Kincaid score measured the cognitive effort needed to process a text based on grammatical features (the number of words, sentences, and syllables), it does not account for another source of cognitive load: lexical features.
Measuring cognitive effort through lexical features: perplexity. Lexical diversity is defined as a measure of the number of different words used in a text (Beheshti et al., 2020). In general, more advanced and diverse language allows to encode more complex ideas (Ellis & Yuan, 2004), which generates a higher cognitive load (Swabey et al., 2016). One of the most obvious measures for lexical diversity is using the ratio of individual words to the total number of words (known as the type-token ratio or TTR). However, this measure is extremely influenced by the denominator (text length). Therefore, I calculate the uncertainty in predicting each word appearance in every text through perplexity (Griffiths & Steyvers, 2004), a measure that has been used for language identification (Gamallo et al., 2016), to discern between formal and informal tweets (Gonzàlez, 2015), to model children's early grammatical knowledge (Bannard et al., 2009), measuring the distance between languages (Gamallo et al., 2017) or to assess racial disparities in automated speech recognition (Koenecke et al., 2020).
For any given text, there is a probability p for each word to appear. Lower probabilities indicate more information while higher probabilities indicate less information. For example, the word "aerospace" (low probability) has more information than "the" or "and" (high probability). From here, I can calculate how "surprising" each word x is by using log(p(x)). Therefore, words that are certain to appear have 0 surprise (p = 1) while words that will never appear have infinite surprise (p = 0). Entropy is the average amount of "surprise" per word in each text, therefore serves as a measure of uncertainty (higher lexical diversity) and it is calculated with the following formula: where p(x) and q(x) are the probability of word x appearing in each text. The negative sign ensures that the result is always positive or zero. For example, a text containing the string "bla bla bla bla bla" has an entropy of 0 because p(bla) = 1 (a certainty), while the string "this is an example of higher entropy" has an entropy of 2.807355 (higher uncertainty). Building upon entropy, perplexity measures the amount of "randomness" in a text: where τ in our case is 2, and the exponent is the cross-entropy. All else being equal, a smaller vocabulary generally yields lower perplexity, as it is easier to predict the next word in a sequence  (Koenecke et al., 2020). The interpretation is as follows: If perplexity equals 5, it means that the next word in a text can be predicted with an accuracy of 1-in-5 (or 20%, on average). Following the previous example, the string "bla bla bla bla bla" has a perplexity of 1 (no surprise because all words are the same and, therefore, predicted with a probability of 1), while the string "this is an example of higher entropy" has a perplexity of 7 (since there are 7 different words that appear 1 time each, yielding a probability of 1-in-7 to appear).
Measuring emotions through polarity: sentiment analysis. The usual way of measuring polarity in written texts is through sentiment analysis. For example, this technique has been used to analyze movie reviews (Bai, 2011), to improve ad relevance (Qiu et al., 2010), to quantify consumers' ad sharing intentions (Kulkarni et al., 2020), to explore customer satisfaction (Ju et al., 2019) or to predict election outcomes (Tumasjan et al., 2010).
Mining opinions in texts is done by seeking content that captures the effective meaning of sentences in terms of sentiment. In this case, I am interested in the determination of the emotional state (positive, negative, or neutral) that the text tries to convey towards the reader. To do so, I employ a dictionary to help in the achievement of this task. More specifically, I employ the AFINN lexicon developed by Finn Årup Nielsen (Nielsen, 2011) one of the most used lexicons for sentiment analysis (Akhtar et al., 2020;Chakraborty et al., 2020;Hee et al., 2018;Ragini et al., 2018). In this dictionary, 2,477 coded words have a score between minus five (negative) to plus five (positive). The algorithm matches words in the lexicon in each text and adds/subtracts points as it effectively finds positive and negative words in the dictionary that appear in the text. If a text has a neutral evocation to emotions will have a value around 0, if a text is evocating positive emotions will have a value higher than 0 and if a text is evocation a negative emotion will have a value below 0. For example, in my sample, one of the highest values in negative emotions (emotion = −32) is the following news from the fake news category reporting about a shooting against two police officers in France: "(…) Her candidacy [referring to Marine Le Pen] has been an uprising against the globalist-orchestrated Islamist invasion of the EU and the associated loss of sovereignty. The EU is responsible for the flood of terrorist and Islamists into France (…)". In contrast, the following junk science article has one of the highest positive values (emotion = +28): "A new bionic eye lenses currently in development would give humans 3x vision, at any age. (…) Even better is the fact that people who get the lens surgically inserted will never develop cataracts".
Measuring emotions through social identity: morality. To measure morality, I will use a previously validated dictionary (Graham et al., 2009). This dictionary has been used to measure polarizing topics like gun control, same-sex marriage or climate change (Brady et al., 2017), to study propaganda (Barrón-Cedeño et al., 2019) or to measure responses to terrorism (Sagi & Dehghani, 2014) and social distance (Dehghani et al., 2016). Like in sentiment analysis, morality is measured by counting the frequency of moral words in each text. The dictionary contains 411 words like "abomination", "demon", "honor", "infidelity", "patriot" or "wicked". In contrast to previous measures, the technique employed to quantify morality is highly sensitive to text length (with longer texts having higher probabilities of containing "moral" words), therefore, I calculate the morality measure as moral words per 500 words in each text. In addition, since negative news spreads farther (Hansen et al., 2011;Vosoughi et al., 2018), I add an interaction term between morality and negativity by multiplying the morality per 500 words and the negativity per 500 words for each text, this being the main measurement for morality: where mor i is the overall number of moral words in text i, neg i is the absolute number of negative words in text i and n w,i is the total number of words in text i.
Similarities between misinformation and factual news: distance and clustering. For the distance metric I use the Euclidean distance that can be formulated as follows: Regarding the method to merge the sub-clusters in the dendrogram, I employed the unweighted pair group method with arithmetic mean. Here, the algorithm considers clusters A and B and the formula calculates the average of the distances taken over all pairs of individual elements a ∈ A and b ∈ B. More formally: To add robustness to the results, I also used the k-means algorithm, which calculates the total within-cluster variation as the sum of squared Euclidean distances (Hartiga & Wong, 1979): where x i is a data point belonging to the cluster C k , and μ k is the mean value of the points assigned to the cluster C k . With this, the total within-cluster variation is defined as: To select the number of clusters, I use the elbow method with the aim to minimize the intra-cluster variation: Differentiating misinformation from factual news: multinominal logistic regression. The main objective of this paper is not just to report descriptive differences between reliable news and misinformation sources, but to look for systemic variance among their structural features measured through the four variables. Therefore, it is not enough to report the averages and confidence intervals of each variable for each category, but also analyzing differences and similarities in the light of all variables altogether. This is why I will employ a multinominal logistic regression, a technique suitable for mutually exclusive variables with multiple discrete outcomes. I employ a multinominal logistic regression model with K classes using a neural network with K outputs and the negative conditional log-likelihood (Venables & Ripley, 2002). This logistic model is generalizable to categorical variables with more than two levels namely {1,…,J}{1,…,J}. Given the predictors X 1 ; ; X p X 1 ; ; X p . In this multinominal logistic regression model, the probability of each level j of Y is calculated with the HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-022-01174-9 ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2022) 9:162 | https://doi.org/10.1057/s41599-022-01174-9 following formula (García-Portugés, 2021): for j = 1,…,J − 1 j = 1,…,J − 1 and (for the reference level J = factual news): As a generalization of a logistic model, it can be interpreted in similar terms if we take the quotient between (A.1) and (A.2): for j = 1,…,J − 1 j = 1,…,J − 1. If we apply a logarithm to both sides, we obtain: Therefore, multinominal logistic regression is a set of J -1 independent logistic regressions for the probability of Y = j versus the probability of the reference Y = J. In this case, I used factual news as the level of the outcome since I am interested how misinformation differs from this baseline using the following formula: where s = sentiment, m = morality, r = readability and p = perplexity. This method will allow us, beyond the fingerprints of misinformation described before, to quantify the differences between factual news and non-factual content.

Results
To investigate the differences between factual news and misinformation, I analyze 92,112 news articles classified into seven categories: clickbait (n = 12,955), conspiracy theories (n = 15,493), fake news (n = 16,158), hate content (n = 15,353), junk science (n = 16,252), rumors (n = 14,158), and factual information (n = 1743) (see Methods for a detailed description of the database). For each article, I calculated measures of cognitive effort (readability and perplexity) and evocation to emotions (sentiment analysis and appeal to morality). Figure 1 shows the average values and confidence intervals for each measure and each category.
With regard to cognitive effort, I obtained the following results.  (Neuman et al., 1992). On the side of more positive language, I find rumors (sentiment = 0.86, [CI = 0.65, 1.08]) and junk science (sentiment = 2.52, [CI = 2.31, 2.72]). Looking at categories with a negative prominence in their content, I find that hate has the highest negative value (sentiment = −6.01, [CI = −6.21, −5.82]), Quantitative similarities among factual news and misinformation categories. After calculating the factual and misinformation categories profiles, I calculate similarities among them through clustering analysis. This is important because results will indicate how close each misinformation category is to reliable news, allowing us to refine the previously presented results. First, I calculated the Euclidean distance (sum of squared distances and taking the square root of the resulting value), forcing values that are very different to add a higher contribution to the distance among observations. After calculating the distance matrix, I employed a hierarchical clustering technique using the unweighted pair group method with arithmetic mean. The results are the following: In Fig. 2A, B one can observe that the Euclidean distance between factual news and rumors is 37.68, with fake news is 35.69, with hate content is 34.84, with conspiracy theories is 34.33, with clickbait is 21.71 and with junk science is 12.00. Looking at the clustering resulting from these distances, one can see that there are two big clusters (height = 24.90): rumors, hate speech, conspiracy theories and fake news on one side and factual content, clickbait, and junk science on the other. This result indicates that the misinformation categories that are more similar to reliable news are click bait and junk science. However, looking at the resulting clusters for a lower height, factual content is the first category to be separated from the others (height 16.86). In other words, at height = 16.86, the biggest difference is between factual content and misinformation sources. From there, the next separation is between clickbait and rumors (height = 11.85). On the other hand, the separations in the first group appear at lower heights (meaning, more similarities among these categories). For example, rumors split from hate speech, conspiracy theories and fake news at height = 6.68. Next, hate speech separates from conspiracy theories and fake news at height = 3.30. Finally, the most similar categories are conspiracy theories and fake news (height = 1.84).
Using the total within-clusters sum of squares in the sample, one can observe that the optimal number of clusters are two. However, I also report all the other clustering possibilities as a robustness check.
In Fig. 3A, junk science and factual content are the most similar categories, while all the rest pertain to a single big cluster. However, if one opt for explore the results increasing the number of clusters (Fig. 3B), they follow the same behavior as the hierarchical clustering. After three clusters, reliable news is the first category to be isolated from all the others, revealing its distinctive nature in terms of linguistic characteristics.
Quantitative differences among factual news and misinformation categories. In Table 3 and Fig. 4 one can see the results of the multinominal logistic regression model with each column comparing the corresponding misinformation category to the baseline (factual news). This method allows us to use reliable news as a "role model" of information and see how misinformation categories diverge from this baseline. The results show that a one-unit increase in the readability score is associated with a decrease in the log odds of categorizing content in the clickbait (β readability; clickbait ¼ À0:06, p < 0.001) conspiracy theories (β readability; conspiracy ¼ À0:05, p < 0.001), fake news (β readability; fake news ¼ À0:21, p < 0.001), and rumor (β readability; rumor ¼ À0:04, p < 0.001) categories. In other words, the Fig. 1 The fingerprints of misinformation. A shows the cognitive effort needed to process a text using the Flesch-Kincaid readability score; B the cognitive effort as measured by perplexity; C plots the sentiment (positive, neutral, negative) of each category; D shows the appeal to morality in each category. easier to read a text is, the more likely it pertains to the clickbait, conspiracy theories, fake news, or rumor categories. On the other side, a one-unit increase in the Flesch-Kincaid score is associated with an increase in the log odds of hate speech (β readability; hate ¼ 0:01, p < 0.1) and junk science categories (β readability; junk science ¼ 0:04, p < 0.001).
As for the sentiment of the content, a one-unit increase in negative sentiment is associated with an increase in the log odds of a content being clickbait β sentiment; clickbait ¼ À0:01, p < 0.001), conspiracy theory (β sentiment; conspiracy ¼ À0:02, p < 0.001), fake news (β sentiment;fake news ¼ À0:02, p < 0.001) or hate speech (β sentiment;hate ¼ À0:03, p < 0.001), indicating that these categories tend to employ a highly negative and sentimental language; while a one-unit increase in the positive sentiment score is associated with an increase in the log odds of a content being junk science (β sentiment;junk science ¼ 0:02, p < 0.001) or a rumor (β sentiment; rumor ¼ 0:01, p < 0.001). Therefore, misinformation tends to rely on emotional language. However, the polarity of these emotions varies across misinformation categories. Both junk science and rumors tend to be significantly more positive than reliable news.
Finally, an increase by one-unit in the morality appealing in a given text is associated with an increase in the log odds of this content being misinformation (β morality; clickbait ¼ 0:18, p < 0.001; β morality; conspiracy ¼ 0:18, p < 0.001; β morality; fake news ¼ 0:20, p < 0.001; β morality; hate ¼ 0:16, p < 0.001; β morality; junk science ¼ 0:16, p < 0.001; β morality; conspiracy ¼ 0:14, p < 0.001). In addition to lexical diversity, the usage of moral language appears to be one of the main determinants of misinformation communication strategies. This finding is important because existing research tends to overemphasize the role of sentiment while neglecting the prominent role of appeal to morality.  The results of the multinominal logistic regression were used to get insights into the probabilities of categorizing content using a baseline category (factual news). Next, I complement the results with predicted probabilities based on simulations for each category in four scenarios corresponding to the four variables using the previous multinominal logistic regression model. In other words, one can predict the probabilities for each category in the scenario (variable) with the following ranges: Readability Regarding the first differences, I use the expected values. The difference between predicted and expected values is subtle but important. Even though both result in almost identical averages (r = 0.999, p < 0.001), predicted values have a larger variance  because they content both fundamental and estimation uncertainty (King, Tomz, & Wittenberg, 2000). Therefore, to calculate the expected value I apply the same procedure for the predicted value and average over the fundamental uncertainty of the m simulations (in this case, m = 100). Specifically, the procedure is to simulate each variable setting all the other variables at their means and the variable I am interested in at its starting point (lower range). Then, I change the value of the variable to its ending point (high range), keeping all the other variables at their means and repeat the simulation. I repeat this process 100 times and average the results for the starting and ending points. The result is a full probability distribution that I use to compute the average expected value and the confidence intervals. From there, I calculate first differences, which are the difference between the two expected, rather than predicted, values (King et al., 2000).
These results show the predicted probabilities for all choices of the multinominal logit model I employed. For each variable I performed 100 simulations with confidence intervals settled at 0.025 and 0.975. The results are the following (Figs. 5-8): For lower levels of complexity (readability = 6.78), the probabilities for a given text being classified in each category is:  . These results show that the junk science and fake news categories are both the most sensible to changes in the readability score in opposite directions: increasing the readability score significantly increases the probability of being categorized as junk science while decreasing it significantly increases the probability of a given content being categorized as fake news.
Next, for the cognitive effort measured through perplexity, one can see that lower levels of perplexity (perplexity = 91), the probabilities for a given text being classified in each category is: factual ( These results indicate that increasing the perplexity of content increases the probabilities of a content being classified as junk science, clickbait or factual news while decreasing perplexity augments the probability of a piece of text being classified as conspiracy theory, fake news, hate speech or rumors. As for sentiment analysis, highly negative values (sentiment = −32) give the following probabilities for each category: factual  These results indicate that highly positive sentiments increase the probability of a content being classified as junk science or rumors, while an increase in the negative sentiment leads to an increased probability of being classified as hate speech, fake news, conspiracy theories or clickbait.
Finally, looking at low levels of appeal to morality (morality = 0.3), the probabilities for a given text being classified in each category is: factual ( . Morality plays a role in identifying categories by augmenting their probabilities of being classified as factual news, rumors, hate speech or junk science as the appeal to morality decreases; and on the contrary, it leads to higher probabilities of being classified as fake news clickbait or conspiracy theories as the appeal to morality in a given text increases.

Discussion
The spread of misinformation content is an important and complex problem because not all misleading content is created equal. These results indicate that there are significant differences between content characteristics from factual sources and misinformation content in the form of clickbait, conspiracy theories, fake news, hate speech, junk science and rumors.
As for the size of these differences, I showed that fake news are, on average, 18 times more negative than factual news (having hate speech the highest difference with 30 times more negativity than factual news); that factual information is 15% more lexically diverse (with rumors being 19% less diverse than factual news); that fake news appeal to moral values 37% more than factual content (although hate speech appeals 50% more to morality than factual sources); and that fake news are 3% easier to process than factual sources (while fake news are 13% easier to process). From here, as seen in the hierarchical clustering, I showed how factual content is different from all the other categories and how fake news and conspiracy theories are closely related; that hate speech and rumors are also similar to conspiracy theories and hate speech or how clickbait and junk science are more similar to factual sources than to the other misinformation categories. These results are important because the proposed fingerprints of misinformation provide grounding for the refinement of previous results and better interventions to mitigate their spread in the realm of technological companies, fact-checkers and media outlets or public policy in the light of the volume, breadth, speed and explainability challenges.
We know that our capacity for discerning between factual and false content is highly influenced by social norms (Frenda et al., 2011;Zhu et al., 2010) and moral foundations (Dehghani et al., 2016;Graham et al., 2009), defined as the "interlocking sets of values, practices, institutions, and evolved psychological mechanisms that function to suppress selfishness" (Graham et al., 2009(Graham et al., , p. 1031. This is especially relevant for people with low levels of media literacy (Lazer et al., 2018) because the results show that, in general, misinformation requires less cognitive effort for being processed and is more reliant on emotions than factual information: it is easier to read, and falsified content creators use a less diverse vocabulary, which may explain why misinformation is more prevalent among low-literacy individuals (Lazer et al., 2018), and more appealing to negative emotions and moral values, which may influence our ability to discern between misinformation and factual content.
This has implications for previous and future research. It may be the case that the results moderate the motivated reasoning mechanism (Bago et al., 2020;Kraft et al., 2015) by augmenting or diminishing its effect depending on the level of emotional evocation and morality and the cognitive effort needed to process it. For example, emotions can be used as a strategy to convey a particular point of view when the appropriate methods are less likely to yield the desired conclusion (Kunda, 1990). Likewise, the results open the door to refining previous research. Preceding evidence shows that misinformation is more viral than factual news (Vosoughi et al., 2018). However, prior studies only focus on one or just a few specific categories, usually fake news. The obtained results offer the opportunity to explore to what extent different misinformation categories are more viral than others and how their structural features in terms of cognitive effort and appeal to emotions drive this virality.
In addition, this method can help technology companies, media outlets and fact-checking organizations to prioritize the content to be checked. This is aligned with the call for "public interest algorithms" to identify misinformation and protect consumers (Wheeler, 2017). With the model I provided and the resulting outcomes, one can inform more complex models to boost people's abilities to identify misinformation. Regarding technology companies, there is evidence that correcting misinformation through "related stories" features in their platforms can reduce the prominence of misleading content (Bode & Vraga, 2015), although just around 3.7% (Pennycook & Rand, 2017). Beyond the small effect, these systems rely on content or sources that have already been flagged as potential misinformation. In contrast, using the results provided in this paper, these platforms may use the same system without the need to explicitly identify the source as misleading and recommending "related stories" based on the assessment of the readability, perplexity, emotionality, and morality levels of the content being shared. The feasibility of real-time deployment should not be a big problem since this method uses manageable calculations in a reasonable amount of time (less than one second for a 2000 words article). Therefore, the proposed method does not add friction to the user experience of a given platform. This will mitigate the volume, breadth, and speed challenges they face to mitigate the spread of misinformation, since the presented findings can be used to build machinelearning classification algorithms that could categorize content at the exact moment it is posted.
The findings also suggest promising interventions for early detection and the identification of check-worthy content, which is important for fact-checking organizations because the flood of misinformation can easily overwhelm them (Pennycook & Rand, 2017). The fact that factual content significantly differs in content characteristics from non-factual content opens the door for improving in the fact-checking process. Since it is impossible to check everything that is being shared on social media, factcheckers focus on content that has already become viral (usually, with information provided from social media platforms in the form of rankings). However, the fingerprints can be used to automatically analyze every URL being shared in social media. If so, platforms and fact-checkers can complement their rankings based on virality with rankings based on the probability of a content being non-factual. More importantly, providing the differences of misinformation categories allows to focus on specific typologies, for example focusing on content targeting people's beliefs or influence major events (like fake news) and not so much on content to attract network traffic (like clickbait) (Vosoughi et al., 2018). Resembling the case of technology companies and social media platforms, the results contribute to mitigate the volume, breadth, and speed challenges that fact-checkers face when they manually assess content for veracity.
The results also highlight the importance of the efforts of public and private institutions to enhance media literacy (European Commission, 2018;Maksl et al., 2016). The notion of media literacy is based on the skills and competencies needed to successfully navigate a complex ecosystem (Eshet & Eshet, 2004;Guess et al., 2020). If people are aware that a specific content is aiming to trigger a reaction through less cognitive load and higher appealing to emotions, they will be less likely to share it (Martel et al., 2019). This is capital as our societies become more and more digital, especially in the light of worrying levels of media literacy. For example, the Pew Research Center found that only 17% of US adults have the skills to safely navigate the information ecosystem (Horrigan, 2016). However, investing in media literacy is particularly important in developing countries, where there are millions of Internet users with low-literacy rates (Mustaffa et al., 2007;Sun & Nekmat, 2008). In these situations, misinformation often escalates into violence, like in the case of ethnic brutality in India (McLaughlin, 2018) or violence against doctors in Latin America (Taylor, 2020). In similar contexts, improving media literacy to process more complex ideas will likely reduce the cognitive load associated with factual information (Ellis & Yuan, 2004;Swabey et al., 2016) and, therefore, the prevalence of misinformation. Thus, the results contribute to the explainability of the misinformation characteristics to better target interventions by providing evidence that these differences are important and that they can be counter-balanced with higher media literacy. For example, since my results indicate that misinformation is more emotional, media literacy efforts should include specific strategies to identify and raise awareness of highly emotional content.
However, this study also has some limitations. First, although I use a big sample of news articles, these are coming from a selection of websites. Even though I tried my best to select different sources for each category, I am aware of the potential limitation that the results could be based on domain-specific features. Secondly, while one of my research objectives was the explainability of the measures and models, I also acknowledge that-potentially-there may be more complex measures and models that may be more useful to discern between misinformation content and factual sources. The objective was not to provide the most detailed model to describe misinformation, but a parsimonious description of key features. Moreover, the results must be interpreted in the light of some methodological limitations that I expose in the corresponding section. Additionally, although the objective of this paper was not discussing the evidence at the source level-but at the misinformation category level, it is also worth mentioning that sources belonging to the same category may behave differently. For example, some fake news sources may use a more neutral tone, while others may show a higher appeal to negative emotions. In other words, the results should be interpreted on the aggregated or average effect, and not in the specific source effect. Finally, although we employed widely used dictionaries in the literature, the results are also contingent to the specificities of each dictionary I employed.
Despite its limitations, the obtained results also suggest interesting paths for future research. For example, how is the fingerprints of misinformation evolving over time? Since misinformation creators are aware of counter-measure interventions, they may be adapting to them through adversarial machine learning (Huang et al., 2011), that is, changing their structural characteristics to avoid detection in automated systems. Similarly, it is interesting to seek if misinformation is becoming more disguised as factual information? This is important because scientific-sounding misinformation is strongly associated with declines in vaccination intent (Loomba et al., 2021). To do so, researchers could look at common characteristics and relationships between sources for each category. On the other hand, the transition from printed and televised news towards the digital world has caused a huge disruption in the traditional media industry, challenging their business models (Shearer, 2021). Focusing on the evolution of traditional media outlets could shed light on the question whether reliable sources are trying to draw on the mechanisms of virality to help increase their penetration (e.g., reliable news becoming more appealing to emotions, titles becoming more exciting, content easier to read, etc.)?
To the best of my knowledge, the findings presented in this paper are a compelling and impactful demonstration that not all misinformation is created equal. Of course, these differences have huge implications that also highlight the contributions. First, they contribute to diminish the problem of scalability in misinformation detection (the volume challenge), by providing a set of features that can differentiate between misinformation and factual content without human intervention. Secondly, they increase the evidence about structural differences among misinformation types (the breadth challenge), which allows to focus on or differentiate specific categories and create tailored countermeasures for each one. Thirdly, I propose a method that can be useful to identify misinformation before it becomes viral (the speed challenge), since the categorization of an URL can be done at the exact moment it is posted, allowing for on-spot detection. Finally, the model that is easily understandable in daily language (the explainability challenge), facilitating its adoption by technical and, specially, non-technical audiences.

Data availability
The datasets generated during and/or analyzed during the current study are available in the following GitHub repository: https:// github.com/several27/FakeNewsCorpus.