How does Truth Evolve into Fake News? An Empirical Study of Fake News Evolution

Mingfei Guo, Xiuying Chen, Juntao Li, Dongyan Zhao, Rui Yan

Abstract

Traditional fake news research often treats misinformation as a static entity. However, in practice, news stories mutate as they propagate through social networks—a process termed "Fake News Evolution" (FNE). This study introduces the FNE dataset, tracking the transformation of objective truth into initial fake news and subsequently into "evolved" fake news. Findings indicate that as stories evolve, they become more objective in tone and increasingly difficult for standard machine learning models to detect.

Introduction: The "Telephone" Effect in Misinformation

Most research into fake news treats it as a static object: a piece of text is either "true" or "false." However, in the real world, news is a living thing. As a story travels through social media and various fringe news outlets, it undergoes a process similar to the children’s game "Telephone." Malicious actors or biased editors modify, distort, and add subjective layers to a story over time. If a model is trained only on the first "draft" of a lie, it might fail to catch the third or fourth iteration, which has been polished to look more credible and professional.

Methodology and the FNE Dataset

The researchers employed a multi-stage analytical approach to understand this mutation process, categorizing changes into four types: Fabrication (adding new lies), Denial (rejecting reality), Confusion (mixing up events), and Distortion (misinterpreting facts). They used Word Mover’s Distance (WMD) to calculate semantic shifts and tools like TextBlob and NLTK to measure changes in polarity, subjectivity, and parts of speech.

To support this analysis, the study introduces the Fake News Evolution (FNE) dataset, consisting of 950 "triples" (2,850 articles total). Each triple tracks a story from its Truth origin to its initial Fake News iteration and finally to its Evolved Fake News state. Data was sourced from Snopes and Archive.today, covering approximately 2 million tokens across politics, religion, health, and economics.

Evolutionary Patterns: The Fabrication Shift

A primary finding is the shift in disinformation tactics over time. While the initial transition from Truth to Fake News often involves distortion, the move to Evolved Fake News sees "Fabrication" become the dominant technique—the invention of entirely new, unrelated claims.

Figure 1: Distribution of disinformation types in the transition to initial fake news (left) versus evolved fake news (right).

Furthermore, evolved fake news tends to adopt a more "objective" and positive tone. Statistical testing (p < 0.01) confirmed that evolved fake news is significantly less subjective than its predecessors. By mimicking the formal, neutral language of legitimate journalism, these evolved lies become more persuasive and harder to distinguish from factual reporting.

Detection Challenges: A Moving Target

The evolution of misinformation poses a significant challenge for automated defense systems. When testing standard classifiers (Naive Bayes, Random Forest, AdaBoost) trained on popular datasets like LIAR, performance consistently degraded when identifying evolved news compared to initial fake news.

Figure 3: Detection performance across different machine learning models for Truth, Fake News, and Evolved Fake News.

For instance, using Adaptive Boosting, the detection accuracy dropped from 0.1874 for initial fake news to 0.1579 for evolved versions. This "moving target" problem suggests that current models are over-optimized for specific linguistic markers that disappear as a lie matures and is refined for broader consumption.

Conclusion and Future Directions

The FNE study highlights that fake news is a dynamic threat. While the core topics (e.g., "Trump," "Clinton," "law") remain consistent, the narratives are completely inverted and professionalized over time. Future work should focus on "teaching" models the underlying features of the evolution process itself, enabling them to predict how a story might mutate before it reaches its most deceptive form.

Summary updated

Abstract

Automatically identifying fake news from the Internet is a challenging problem in deception detection tasks. Online news is modified constantly during its propagation, e.g., malicious users distort the original truth and make up fake news. However, the continuous evolution process would generate unprecedented fake news and cheat the original model. We present the Fake News Evolution (FNE) dataset: a new dataset tracking the fake news evolution process. Our dataset is composed of 950 paired data, each of which consists of articles representing the three significant phases of the evolution process, which are the truth, the fake news, and the evolved fake news. We observe the features during the evolution and they are the disinformation techniques, text similarity, top 10 keywords, classification accuracy, parts of speech, and sentiment properties.
BibTeX
@inproceedings{guo2021how,
  author = {Mingfei Guo and Xiuying Chen and Juntao Li and Dongyan Zhao and Rui Yan},
  title = {{How does Truth Evolve into Fake News? An Empirical Study of Fake News Evolution}},
  booktitle = {The Web Conference},
  year = {2021},
  doi = {10.1145/3442442.3452328},
}
Figure