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

Iranian Journal of Science and Technology, Transactions of Electrical Engineering · Springer Nature · 2026-11-19 · DOI: 10.1007/s44163-026-02723-w

Аннотация

TikTok has skyrocketed in popularity over recent years, especially among younger audiences. However, there are public concerns about the potential of this platform to promote and amplify harmful content. This study presents a computational analysis of conspiracy theories on TikTok. By leveraging the official TikTok Research API we collect a longitudinal dataset of 1.5M videos shared by creators in the U.S. over three years. We estimate a lower bound on the prevalence of conspiratorial videos (up to 1000 new videos per month) and evaluate platform changes in association with TikTok’s Creativity Program for monetization, observing an overall increase in video duration regardless of content. In addition, we evaluate the capabilities of recent open-weight Large Language Models, both textual and multimodal, to identify conspiracy theories in short vertical format videos. While these models achieve high precision in detecting harmful content (up to $$98\%$$), their overall performance remains comparable to a fine-tuned traditional model such as RoBERTa. Our findings indicate that Large Language Models have the potential to support content moderation strategies on TikTok; however, their deployment involves trade-offs relative to fine-tuned traditional NLP approaches, particularly in terms of performance consistency.

Читать у издателя →

Похожие статьи