💬 NLP

Language Mutations Sustain the Persistences of Conspiracy Theories on Social Media

This study analyzes a three-year dataset of X posts to demonstrate that linguistic and structural mutations in conspiracy theories, particularly through simplification and assimilation, significantly extend their online lifespan, suggesting that content moderation strategies must target core claims to effectively counter their persistent variations.

Original authors: Calvin Yixiang Cheng, Dorian Quelle, Scott A. Hale

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Calvin Yixiang Cheng, Dorian Quelle, Scott A. Hale

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Idea: Why Conspiracy Theories Won't Die

Imagine a conspiracy theory as a living organism, like a virus or a weed. You might think that if you debunk it (show it's false), it will die. But on social media, these theories often keep coming back, even after being proven wrong.

This study asks: Why do they survive so long?

The answer isn't just that people are stubborn. It's that the theories change their clothes. Just like a chameleon changes its color to blend in, conspiracy theories change their words, characters, and details to stay fresh and avoid being ignored. The authors call this "language mutation."

The Experiment: Tracking the "Shape-Shifting"

The researchers looked at nearly 450,000 posts about COVID-19 conspiracies on X (formerly Twitter) over three years. They didn't just count how many times a theory was posted; they tracked how the words changed as the theory spread from person to person.

They treated a "conspiracy claim" like a family tree. Even if the posts look different, if they share the same core idea (e.g., "The virus was made in a lab"), they belong to the same family. The researchers watched how these families evolved over time.

The Three Ways Theories Mutate

The study looked at three specific ways these theories change to survive:

1. The "Drift" (Semantic Mutation)

The Analogy: Imagine a game of "Telephone." The first person says, "The government is hiding the truth." By the time it gets to the tenth person, it might sound like, "The authorities are keeping secrets." The meaning is slightly different, but the core is the same.
The Finding: Theories that changed their wording slightly (drifted) lived 27% longer than those that stayed exactly the same. It seems that a little bit of novelty keeps people interested, preventing them from getting bored (information fatigue).

2. The "Emotional Shift" (Psycholinguistic Mutation)

The Analogy: Think of a conspiracy theory as a movie script. Sometimes the script changes from a scary horror movie to a sad drama, or from a political thriller to a health warning.
The Finding: Theories that changed their emotional tone survived longer. Specifically, mutations in these areas helped them stick around:

Social Words: Changing "they" to "us" or "our friends."
Time: Shifting focus from the past to "right now."
Health & Risk: Swapping general talk for specific words about sickness, danger, or death.
Thinking: Changing how they describe "knowing" or "believing."
The Result: If a theory started talking more about "health risks" or "family," it lasted significantly longer.

3. The "Cast Change" (Actor-Action-Target Mutation)

The Analogy: Imagine a play where the plot is "The Villain is poisoning the town."

Actor (Who): The villain changes from "The Government" to "Big Pharma" to "A specific Politician."
Action (What): The crime changes from "Hiding data" to "Killing people" to "Stealing money."
Target (Who gets hurt): The victim changes from "Everyone" to "Children" to "The Poor."
The Finding: When the theory swapped out the villain or the victim, it lived longer. Interestingly, changing the Action (what the villain is doing) was the most powerful way to extend the theory's life. Also, theories that started targeting vulnerable groups (like children or the poor) tended to survive longer, likely because it triggered stronger emotional reactions.

How They Change: Two Main Patterns

Through a close look at the text, the researchers found two main ways these mutations happen:

Simplification (The "Lazy Reader"):
People often make complex conspiracy theories easier to digest. They strip away the details and turn them into simple slogans like "Wake up!" or "Vote Red!" or just labeling something a "hoax." It's like turning a complex novel into a one-sentence tweet. This makes the theory easier to share.
Assimilation (The "Echo Chamber"):
People twist the story to fit what they already believe. If someone is already angry at a specific group, the conspiracy theory mutates to blame that group more directly. It's like a mirror that reflects the viewer's own biases back at them, making the theory feel more personal and true.

The Big Takeaway

The paper concludes that conspiracy theories survive because they are flexible. They don't just repeat the same lie; they adapt. They change their words, their emotional tone, and their characters to stay relevant and engaging.

What this means for stopping them:
The authors suggest that if we want to stop these theories, we can't just debunk one specific version of the story. Because the story changes so easily, we need to target the core, unchanging idea (the "atomic claim") rather than every single variation. If you only fight the "Government" version, the theory will just mutate to blame "Big Pharma" and keep living. You have to fight the underlying logic that allows it to survive in the first place.

Technical Summary: Language Mutations Sustain the Persistences of Conspiracy Theories on Social Media

Problem Statement
Conspiracy theories exhibit surprising persistence on social media, often re-emerging despite repeated debunking. While prior research has focused on individual psychological characteristics or the viral diffusion of static content, the role of language mutations—linguistic changes occurring during transmission—in sustaining this persistence remains under-explored. The authors posit that as conspiracy claims circulate, they undergo semantic and structural modifications (mutations) that may reduce information fatigue, renew audience attention, and allow the claims to adapt to new contexts, thereby extending their lifespan. The study seeks to empirically quantify the extent to which these mutations predict the persistence of conspiracy claims online.

Methodology
The study utilizes a three-year dataset (January 2020 – November 2022) comprising 446,829 COVID-19 related conspiracy posts collected from X (formerly Twitter). The methodology involves a multi-stage pipeline:

Data Collection & Validation: Posts were retrieved using keyword matching derived from 562 fact-checking articles, followed by filtering with a fine-tuned GPT-4o-mini classifier ( $F_1 = 0.90$ ) and reverse validation by human annotators.
Claim Matching: To analyze persistence at the "claim" level rather than the individual post level, semantically similar posts were clustered using sentence embeddings (all-mpnet-base-v2) and a threshold-based similarity graph (cosine similarity threshold = 0.88). A claim is defined as a cluster of posts sharing a core assertion that can be addressed by a single fact-check.
Mutation Quantification: Mutations were measured across three analytical scales:
- Semantic Drift: Measured as the cumulative cosine distance between early posts (within 24 hours of the seed) and the seed post.
- Psycholinguistic Properties: Analyzed using LIWC22. Mutations were defined as a >50% change in the frequency of specific categories (e.g., pronouns, health, cognitive processes, risk, social references) between consecutive posts.
- Actor-Action-Target (AAT): Extracted using LLMs (GPT-4o-mini) and clustered via K-Means. Mutations were identified as changes in the cluster assignments of actors, actions, or targets within a claim.
Survival Analysis: The persistence of claims was modeled using Kaplan-Meier estimators and Accelerated Failure Time (AFT) Weibull models. The "event" was defined as the absence of new posts in a claim cluster for 30 consecutive days. The analysis controlled for engagement metrics (likes, retweets) and user characteristics.
Qualitative Analysis: A subset of long-lasting claims was manually coded to identify specific mutation patterns (simplification, assimilation) in psycholinguistic and AAT elements.

Key Results
The study finds that language mutations are significantly associated with extended claim lifespans:

Semantic Mutations: Claims exhibiting early semantic drift (within the first 24 hours) have a 27% higher likelihood of persisting compared to non-mutating claims ($TR = 1.27, p < 0.005$).
Psycholinguistic Mutations: Claims with mutations in specific categories survive significantly longer. Notably, mutations in social reference words (e.g., "we" vs. "they") increased survival time by 122% ($TR = 2.22$), and cognitive process mutations increased it by 67% ($TR = 1.67$). The magnitude of fluctuation in health-related language was particularly predictive, with claims showing higher variance surviving nearly three times longer.
AAT Mutations: Changes in actor, action, or target categories all correlate with longer lifespans. In multivariate models, action mutations emerged as the strongest independent predictor, increasing lifespan by 68% ($TR = 1.68$).
Qualitative Patterns: Two predominant mutation patterns were identified:
- Simplification: Users reduce cognitive complexity, often through labeling (e.g., "hoax"), call-to-action appeals, or reducing reasoning words.
- Assimilation: Users distort messages to align with pre-existing attitudes, often amplifying threats or shifting targets to vulnerable populations (e.g., children, the poor).

Significance and Claims
The authors argue that this study provides the first empirical evidence quantifying the effect of semantic drift and specific linguistic mutations on the lifespan of conspiracy theories. By demonstrating that even small semantic tweaks or structural changes in AATs can sustain a claim, the paper extends information mutation theory to contemporary online communication contexts.

The findings suggest that content moderation strategies should evolve from targeting static narratives to addressing the mutability of claims. Specifically, the authors propose that debunking efforts should focus on "atomic claims" (core, simplified assertions) rather than comprehensive narratives, as mutations often follow simplification patterns. Furthermore, understanding assimilation patterns could inform personalized debunking strategies that acknowledge users' existing cognitive frameworks. The study concludes that the persistence of conspiracy theories is not merely a function of belief but is actively sustained by the linguistic adaptability of the claims themselves.