Résultats du processus de synthèse de l'article arXiv : 2407.12545v1

Résumé : TikTok has skyrocketed in popularity over recent years, especially among younger audiences, thanks to its viral trends and social challenges. However, concerns have been raised about the potential of this platform to promote and amplify online harmful and dangerous content. Leveraging the official TikTok Research API and collecting a longitudinal dataset of 1.5M videos shared in the US over a period of 3 years, our study analyzes the presence of videos promoting conspiracy theories, providing a lower-bound estimate of their prevalence (approximately 0.1% of all videos) and assessing the effects of the new Creator Program, which provides new ways for creators to monetize, on the supply of conspiratorial content. We evaluate the capabilities of state-of-the-art open Large Language Models to identify conspiracy theories after extracting audio transcriptions of videos, finding that they can detect harmful content with high precision but with overall performance comparable to fine-tuned traditional language models such as RoBERTa. Our findings are instrumental for content moderation strategies that aim to understand and mitigate the spread of harmful content on rapidly evolving social media platforms like TikTok.