• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Friday, November 20, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

TikTok study finds LLMs match fine-tuned models for conspiracy detection

by
November 20, 2026
in Technology
Reading Time: 5 mins read
0
TikTok study finds LLMs match fine-tuned models for conspiracy detection

TikTok study finds LLMs match fine-tuned models for conspiracy detection

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A new study published in Discover Artificial Intelligence provides a large-scale computational analysis of conspiracy theories on TikTok, comparing the performance of recent open-weight Large Language Models (LLMs) against fine-tuned traditional embeddings. The research team, led by Francesco Corso and colleagues, collected a longitudinal dataset of approximately 1.5 million videos shared by creators in the United States over a three-year period from 2021 to 2023. By leveraging the official TikTok Research API, the study aims to quantify the prevalence of conspiratorial content and evaluate the efficacy of modern AI architectures in identifying such material within short, vertical-format videos.

The researchers defined conspiracy theories as false or unverified narratives alleging secret, often sinister plots involving groups, organizations, or governments. To address this, they developed a systematic approach to data collection and analysis, focusing on three specific research questions regarding prevalence, platform changes, and detection capabilities. This quantitative approach contrasts with previous qualitative studies that relied on manual reviews of specific hashtags or topics.

To estimate the prevalence of conspiratorial videos, the team utilized a corpus of 90,000 web documents related to conspiracy theories to identify relevant hashtags. They applied a hashtag enrichment process to a sample of 1.5 million long videos, filtering for terms with clear conspiratorial semantics while excluding ambiguous terms like “cancer” or “coronavirus” that often appear in non-conspiratorial contexts. This method allowed them to establish a conservative lower bound on the number of conspiracy videos, estimating that up to 1,000 new conspiratorial videos were shared per month by 2023. The study noted that while the percentage of such videos remained stable around 0.1%, the absolute number grew steadily as the platform expanded.

The researchers also examined the impact of TikTok’s Creativity Program, a monetization initiative introduced in May 2023 that incentivized the creation of longer videos. By comparing video duration distributions before and after the program’s launch, they observed a statistically significant increase in the average length of videos across the platform. However, this shift was not specific to conspiratorial content; the fraction of conspiracy-related videos remained stable, indicating that the program drove a broader change in user behavior rather than a targeted increase in harmful narratives.

For the detection component of the study, the team evaluated several open-weight LLMs, including Llama3, Mistral, Gemma, DeepSeek-r1, and Qwen3, alongside the multimodal VideoLLaMA 3. They used audio transcriptions generated by OpenAI’s Whisper model for the text-only models, ensuring that the input was faithful to the spoken content of the videos. The models were tested using zero-shot prompting strategies, meaning they were not given examples of conspiracy theories before classification. Three prompting variants were used: a simple instruction, a prompt including a formal definition of conspiracy theory, and a step-by-step chain-of-thought prompt designed to elicit reasoning from the model.

The results revealed that while these large language models achieved high precision in detecting harmful content, with some configurations reaching up to 98%, their overall performance was comparable to a fine-tuned RoBERTa baseline. In balanced datasets, VideoLLaMA 3 demonstrated the highest precision, but it often suffered from lower recall compared to the traditional model. In unbalanced settings that more closely mimic real-world content moderation scenarios, where conspiracy videos are a minority, the performance of LLMs varied significantly.

The study highlighted a persistent trade-off between precision and recall in LLM-based detection. While larger and multimodal models tended to achieve higher precision, their recall often remained stable or decreased, suggesting they were conservative in assigning positive labels. This behavior was particularly pronounced in VideoLLaMA 3, which yielded the highest precision overall but failed to detect a substantial number of conspiratorial videos.

An analysis of error patterns provided insight into why these models struggle with certain types of content. The team identified that models often failed to detect conspiracy theories when the content was framed as a rejection of one theory in favor of another, such as rejecting flat-earth in favor of a “dome-earth” narrative. Additionally, models frequently misclassified satirical or humorous content as conspiratorial, failing to capture the ironic or parody context of the speech. Conversely, explicit references to well-known theories like Bigfoot or UFOs were sometimes missed if the visual or audio cues were subtle or if the topic was considered more “mystery” than “conspiracy” by the model’s internal logic.

The researchers also investigated the effect of transcription length on classification performance. They found a positive association between the length of the transcript and model accuracy, with both precision and recall generally improving as the text became longer.

Several limitations were acknowledged in the study, including the reliance on the opaque TikTok Research API and the potential for biases in the hashtag-based labeling process. The authors noted that their analysis is correlational rather than causal, particularly regarding the impact of the Creativity Program, as hidden confounders could influence the observed trends. Furthermore, the evaluation did not isolate the independent effects of model capacity, architecture, and prompt formulation, making it difficult to attribute performance differences to a single factor. The study also emphasized that the results are specific to TikTok and should not be generalized to other platforms or to short videos, non-English content, or predominantly nonverbal material.

Despite these limitations, the study provides valuable evidence for the development of automated content moderation tools. It demonstrates that while LLMs have the potential to support moderation strategies, their deployment involves significant trade-offs relative to fine-tuned traditional NLP approaches. The high computational cost of inference for large models, particularly multimodal ones, must be weighed against the benefit of avoiding manual labeling.

The research contributes to the growing body of literature on computational social science and the detection of harmful online content. By providing a reproducible pipeline for evaluating LLMs on a real-world dataset, the study offers a benchmark for future work in this area. The authors suggest that future research should focus on disentangling the effects of model size and prompt design, as well as exploring more robust methods for handling multimodal data. Additionally, they recommend investigating the behavioral patterns of conspiracy content creators over time to better understand the dynamics of misinformation propagation.

In conclusion, the study highlights the complexity of detecting conspiracy theories on social media platforms. The balance between precision and recall remains a critical challenge, and the high cost of deploying large language models requires careful consideration.

Subject of Research: Comparing fine tuned embeddings and large language models for detecting multimodal online conspiracy theories

Article Title: Comparing fine tuned embeddings and large language models for detecting multimodal online conspiracy theories

Article References: Corso, F., Bono, C. A., Morales, G. D. F., & Pierri, F. (2026). Comparing fine tuned embeddings and large language models for detecting multimodal online conspiracy theories. Discover Artificial Intelligence, 6(1), Article 1917. https://doi.org/10.1007/s44163-026-02723-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02723-w

Keywords: Comparing, fine, tuned, embeddings, large, language, models, detecting, multimodal, online, conspiracy, theories

News Source: Denise Maddox. (November 20, 2026). TikTok study finds LLMs match fine-tuned models for conspiracy detection. Scienmag.

Tags: ComparingconspiracydetectingembeddingsFinelanguagelargemodelsmultimodalonlinetheoriestuned
Share12Tweet7Share2ShareShareShare1

Related Posts

Optimized Machine Learning Predicts Sugarcane Bagasse Ash Concrete Compressive Strength Accurately

Optimized Machine Learning Predicts Sugarcane Bagasse Ash Concrete Compressive Strength Accurately

November 20, 2026
Stowers Institute Appoints Brandon Mayer as Chief Technology Officer

Stowers Institute Appoints Brandon Mayer as Chief Technology Officer

November 20, 2026

Drainage Area Primarily Controls Chattahoochee River Conductance Under Urban Watershed Dynamics

November 20, 2026

Hybrid Retrofitting Framework Improves Energy Efficiency in Aging Egyptian University Buildings

November 20, 2026

POPULAR NEWS

  • Study Uncovers Stark Regional and Social Disparities in Mozambique HIV Testing

    Study Uncovers Stark Regional and Social Disparities in Mozambique HIV Testing

    29 shares
    Share 12 Tweet 7
  • Optimized Machine Learning Predicts Sugarcane Bagasse Ash Concrete Compressive Strength Accurately

    29 shares
    Share 12 Tweet 7
  • Stowers Institute Appoints Brandon Mayer as Chief Technology Officer

    29 shares
    Share 12 Tweet 7
  • Activating MAC-1 Receptors Alleviates Brain Inflammation Following Intracerebral Hemorrhage in Preclinical Models

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Qualitative Study Maps Eight Pillars of Humane ICU Nursing Care

Study Uncovers Stark Regional and Social Disparities in Mozambique HIV Testing

Optimized Machine Learning Predicts Sugarcane Bagasse Ash Concrete Compressive Strength Accurately

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.