Predicting Misinformation Propensity From User Language

22 concepts12 papers

Traditional approaches to combating online misinformation operate like a game of whack-a-mole: an article is published, fact-checkers or algorithms evaluate its veracity, and if it is deemed false, the platform flags or removes it. However, this post-hoc approach is inherently reactive. By the time a piece of fake news is identified, it may have already been shared thousands of times, cementing its influence on public discourse.

To build a more proactive defense, a growing body of research shifts the focus from what is being said to who is saying it. The core insight is that a user's historical, everyday language on social media serves as a proxy for their underlying psychological traits, cognitive styles, and behavioral predispositions. By analyzing these pre-existing linguistic patterns, machine learning models can predict a user's propensity to propagate or refute misinformation before they ever encounter a specific rumor.

Language as a Psychological Proxy

The fundamental premise of this research direction rests on psycholinguistics: the words we choose reflect how we think, feel, and relate to society. For instance, Mohammadzaman Zamani et al. demonstrated in Predicting Human Trustfulness from Facebook Language that social media language can effectively predict stable personality traits like trustfulness, yielding insights comparable to traditional psychometric assessments.

When applied to the domain of online information disorder, these linguistic footprints reveal distinct psychological and motivational profiles. As Mansooreh Karami et al. show in Profiling Fake News Spreaders on Social Media through Psychological and Motivational Factors, individuals who spread fake news exhibit statistically significant differences in their psychological drivers compared to those who share reliable news. Rather than treating misinformation sharing as an isolated, accidental action, this paradigm views it as an extension of a user's cognitive and emotional makeup.

Consequently, Anastasia Giachanou et al. argue in Online information disorder: fake news, bots and trolls that automated systems can achieve high precision by capturing these traits through psycholinguistic patterns, emotional signals, and writing styles. This behavioral approach focuses entirely on the characteristics of the source rather than the truth value of individual claims, offering a robust alternative to content-based filtering.

The Proactive Advantage: Predicting Before the Spread

The primary operational advantage of user profiling over content verification is its temporality. Traditional rumor detection frameworks often rely on post-hoc analysis, which Fan Zhou et al. note in Catch me if you can: A participant-level rumor detection framework via fine-grained user representation learning fails to prevent early-stage propagation. In contrast, profiling users through language allows for early intervention.

Remarkably, this linguistic predisposition exists prior to exposure. In a study tracking user behavior, Qin Lv et al. discovered in Analyzing Behavioral Changes of Twitter Users After Exposure to Misinformation that individuals who interact with misinformation are linguistically distinct from the general population before any specific exposure event occurs. This implies that users are not suddenly transformed by fake news; rather, certain profiles are inherently predisposed to engage with it.

Capitalizing on this, Yida Mu et al. formulated a binary classification task in Identifying Twitter users who repost unreliable news sources with linguistic information to predict whether a user will repost content from unreliable sources based solely on the linguistic patterns of their prior tweets. By training models on pre-existing language, the authors demonstrated that it is entirely feasible to identify potential disinformation agents before they act, opening the door for timely interventions to mitigate the spread of disinformation.

Building Language-Based User Representations

To translate these insights into predictive models, researchers construct dense computational representations of users based entirely on their text output. Marco Del Tredici et al. showed in Words are the Window to the Soul: Language-based User Representations for Fake News Detection that language-based user representations, derived by aggregating a user's social media posts, significantly improve fake news detection accuracy. In fact, these user-centric models frequently outperform traditional models that rely exclusively on analyzing the text of the news articles themselves.

However, language does not exist in a vacuum. The most robust architectures combine these linguistic representations with structural and behavioral data. For example, Yi-Ju Lu et al. developed a graph-aware co-attention network in GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media that models the joint patterns between sensationalist language and user demographics (such as account age and verification status).

Similarly, Flora Sakketou et al. introduced the FACTOID dataset in FACTOID: A New Dataset for Identifying Misinformation Spreaders and Political Bias to integrate user language with social interaction graphs via graph attention networks. Their work confirms that modeling the propagation network alongside user traits provides a highly accurate, proactive signal for curbing misinformation, sometimes even outperforming content-based features.

Mapping the Spectrum: From Amplifiers to Refuters

This line of research moves beyond a simplistic binary classification of "good" versus "bad" users, mapping out a complex ecosystem of online behavior. By clustering behavioral and linguistic data, Val Alvern Cueco Ligo et al. identified distinct "user archetypes" within online discourse in User Archetypes and Information Dynamics on Telegram: COVID-19 and Climate Change Discourse in Singapore. These archetypes range from "Inquisitive Moderates" to "Conspiratorial Amplifiers," allowing platforms to implement targeted interventions, such as providing preemptive fact-checks to moderate users before they are influenced by extreme content amplifiers.

Crucially, this predictive capability also extends to identifying positive actors within the network. Zongmin Li et al. applied NLP features in Machine Learning Methods to Predict Social Media Disaster Rumor Refuters to identify "anti-rumor spreaders"—individuals who voluntarily step forward to debunk and refute disaster-related rumors. By analyzing the sentiment and text similarity of recent microblogs, their framework can predict who will actively fight misinformation, allowing platforms to amplify these organic refuters during crises.

Nuances and Open Challenges

While language-based profiling offers a powerful tool for proactive defense, it is not without tension. Extracting clean, predictive signals from massive, noisy streams of user text remains a significant engineering challenge. Furthermore, the boundaries of this approach must be carefully calibrated.

Dimosthenis Antypas et al. caution in A Multi-Faceted NLP Analysis of Misinformation Spreaders in Twitter that while subtle linguistic differences exist between spreaders and non-spreaders, relying solely on user profiling can be insufficient. The authors suggest that the most effective strategies must continue to balance both user-centric behavioral profiling and content-centric verification to fully safeguard the integrity of public discourse.

Key Foundational Work

Machine Learning Methods to Predict Social Media Disaster Rumor Refuters (2019) — contributed 2 concept elements to this direction.
Predicting Human Trustfulness from Facebook Language (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_language_as_a_proxy_for_user_traits_b203793b,
  title = {language as a proxy for user traits},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/language-as-a-proxy-for-user-traits/rdr_b203793b012b4b22a442c06cbf3af04b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_language_as_a_proxy_for_user_traits_b203793b_v1,
  title = {language as a proxy for user traits},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/language-as-a-proxy-for-user-traits/rdr_b203793b012b4b22a442c06cbf3af04b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake