Auditing TikTok’s recommendations: Understanding the prevalence of misinformation on the platform

Short video platforms have rapidly increased in popularity, bringing forth a concerning trend; widespread dissemination of misinformation partially owing to the persuasive nature of the rich media these platforms support. However, little research has systematically investigated the role of algorithmic recommendations in this phenomenon. Our study addresses this gap by auditing the recommendation algorithm of TikTok, the leading short video platform, through four experiments. We explore the prevalence of misinformation in TikTok's auto-complete suggestions, search results, and For You Page (FYP) recommendations. Our audits yielded a novel dataset comprising 8535 auto-complete suggestions, 1606 search queries, and 2893 videos labelled by stance (promoting misinformation, neutral in nature, or debunking misinformation). Our analysis reveals that TikTok's algorithm exhibits sensitivity to user behaviours, often in unexpected ways. We find the algorithm recommends substantially more misinformation to newly created accounts lacking search histories, potentially shaping the tone of the content these accounts engage with. We also find that watching and engaging with videos (e.g., liking and favouriting) dynamically influence misinformation prevalent on the platform, potentially reinforcing existing beliefs and contributing to echo-chambers. Further, we emphasise how engaging with misinformation on one topic serves as a gateway, prompting the algorithm to recommend misinformation spanning both related and unrelated topics. Our work stands as the first large-scale, systematic, and empirical assessment of misinformation in TikTok's recommendations, contributing valuable insights on the dynamics of misinformation dissemination on TikTok and also underscoring the pressing need of responsible auditing practices across digital platforms.