Introduction

What is the purpose of misinformation? It may be confusing why anybody would spread blatantly incorrect information, but misinformation is actually designed to gain as much traction as possible. By relying on baiting emotions, appealing to morality, and creating dramatic conspiracy theories, misinformation spreads six times faster than factual information does. The tactics used to make misinformation so appealing are reviewed in a recent study by Carlos Carrasco-Farré. He identified several “fingerprints of misinformation,” or hallmarks of misinformation online, including grammar, writing style, vocabulary diversity, and emotional and moral appeals. By learning how to identify these characteristics, people can better avoid misinformation and ensure that sensationalist content will not affect their critical thinking, which is crucial when misinformation relies on leveraging our fundamental morals and biases. 

Methods

The study used a large text database called the Fake News Corpus, which has millions of news articles from 194 websites labeled into seven categories. Articles are labeled as factual (reliable news) or as one six misinformation types: clickbait, conspiracy theories, fake news, hate speech, junk science, and rumors. The domains were labeled through crowdsourcing checking for suspicious words in titles or URLs, author credibility, inconsistent style and grammar, and manipulated images. 

Articles shorter than 200 words or longer than 2000 words were removed to avoid text noise bias. This left 147,550 articles. Next, researchers computed 4 fingerprint variables:
Readability was measured through the Flesch-Kincaid index using the number of words, sentences, and syllables to estimate required reading effort. 
Perplexity measured lexical cognitive load using entropy or “surprise per next word.” Lower surprise = easier prediction = less thinking. Higher surprise = more randomness = more thinking.

Sentiment polarity used AFINN scoring from positive to neutral to negative based on emotion word frequency. 
Moral appeal used the Moral Foundations Dictionary, standardized per 500 words to avoid longer text moral overcount bias (biased scoring just for being longer). 
Hierarchical clustering was done using Euclidean distances (straight line distance) to measure the similarity of shapes among categories (how similar or different misinformation is).

The final analyzed sample size was 92,112 articles containing public-interest digital content. 

Analysis

Several patterns emerged from the data. The study found that misinformation takes less thinking on average to read than factual journalism. Fake news texts were 3% easier to read than factual, reliable sources. They were also 15% less lexically diverse. Reliable texts used more unique vocabulary for complex ideas and precise word meaning, while misinformation follows more predictable patterns. 

Reliable news stayed mostly neutral in sentiment polarity (around −0.19) while hate speech was the most negative fingerprint category, relying 10–18 times more on negative sentiment words than factual reliable baseline. Misinformation overall relied 10x more on negative sentiment than factual baseline. This strategy pushes readers to react before thinking. Moral appeal was one of the strongest fingerprints. Misinformation appealed 37% more to morality than factual journalism, hate speech appealed 50% more, conspiracy appealed 45% more, and fake news appealed 37% more. Moral identity language encourages fast us-versus-them thinking, like group signaling, loyalty, or moral policing, which encourages sharing before checking for accuracy.

Clustering similarity showed two main clusters. Cluster 1: conspiracy, fake news, hate speech, and rumors are highly similar. Cluster 2: clickbait and junk science are closer to factual readability but still have distinct fingerprints of emotionality and morality. Factual information diverges most strongly from all misinformation when fingerprints are seen together. Essentially, factual information is very obviously different from misinformation when directly compared.

Readability showed opposite sensitivity: when readability rose, classification probability dropped for fake news (−35%), and rose for junk science (+40.19%).This means different misinformation categories leave fingerprints that make it harder to sort them correctly. 

Due to this, there is no “one size fits all” approach to combating misinformation, and this will need to be acknowledged when building tools to counter the spread of misinformation. AI (machine learning algorithms) is already being used to great effect, but LLMs (large language models) also heavily contribute to the creation of misinformation by oftentimes making up facts and statistics.

Conclusion

Though factual information offers truth and real knowledge, accessing it comes at the cost of readability. This contrasts with misinformation, which is designed to make us feel instead of think by using easy to read, emotionally charged wording. So, the next time you see a headline that catches your attention, take a moment to think about its intention. If a claim seems too far-fetched or absurd to be true, it may be exactly that—misinformation.

Full article: https://www.nature.com/articles/s41599-022-01174-9#Abs1