Altering Words to Evade Perceived Moderation: Decoding Algospeak in Chinese Social Media Video Captions

Abstract

Video content creators on social media often modify their caption text to bypass algorithmic moderation as they perceive it -- a practice known as algospeak. This behavior of adjusting expressions online could become especially complicated in Chinese online space where both moderation and censorship overlap, and where certain types of expression are explicitly restricted. In this study, we conduct a large-scale analysis of algospeak by examining over 200,000 movie-recap videos from 96 YouTube channels, most of whom are also prominent on Douyin (the domestic implementation of TikTok in China). We extract the video's hard‑coded captions and employed AI to detect altered terms. We then examine the prevalence and characteristics of these alterations, and leverage two commercial text‑sensitivity detectors to infer the potential motivations and policies driving these algospeak behaviors. We also explore whether factors such as content reach and monetization strategies correlate with creators' propensity to modify captions. We find that the presence of embedded ads in videos as a form of brand deals was negatively associated with overall alteration rates: content creators increased alterations for a few explicitly warned topics while reducing them for other topics. Overall, we argue that algospeak in Chinese videos represents a collectively adopted strategy that both evades perceived moderation and reinforces restrictive norms within online communities, and is partly attributable to a trust gap between users and the platform on the matter of moderation.