Examining the Limitations of Existing Content Moderation Systems in Detecting Sheng-Based Cyberbullying on Kenyan TikTok

Abstract

TikTok has become the leading social media platform among young Kenyans, but its growth has come with a rise in cyberbullying, much of it written in Sheng or in comments that mix English, Swahili, and Sheng within a single sentence. The moderation tools used to filter harmful content are mostly trained on English data drawn from Western platforms, and earlier research suggests they miss as much as 68 percent of cyberbullying written in this style. This paper examines the scale of that failure directly, by testing four widely used content moderation tools, Perspective API, TextBlob, a generic BERT classifier, and RoBERTa, on a set of 5,000 Kenyan TikTok comments labelled by trained Kenyan linguists. The results show that all four tools fall well short of acceptable performance, with accuracy ranging from 44.7 to 64.8 percent and false negative rates between 43 and 67 percent. A culturally aware alternative built on the AfriBERTa architecture, included in the comparison for reference, reached 93 percent accuracy and an 8 percent false negative rate, underlining the scale of the gap. The findings confirm that current commercial moderation systems are not equipped to protect Kenyan TikTok users, particularly those communicating in Sheng, and that locally calibrated alternatives are both necessary and achievable.

Keywords: content moderation, cyberbullying, Sheng, code-switching, TikTok, natural language processing

DOI: 10.7176/CEIS/17-1-07

Publication date: August 28^th, 2026