AI moderation misses most African languages ^{CEFR B1}

Moderators and researchers describe a gap between the languages people use and the languages AI moderation tools can process. Bereket Tsegay, who worked at TikTok’s Kenya hub, said he often watched videos he did not understand because they were in Luo, Dholuo, Kikuyu and Dinka.

A 2025 study found that only 42 African languages appear meaningfully in major language models, and just four are handled with any consistency: Amharic, Swahili, Afrikaans and Malagasy. That leaves more than 98 percent of Africa’s languages largely invisible to moderation systems. The platforms’ dependence on English-language data causes both false positives and false negatives.

Concrete cases underline the problem. A Kenyan creator had an account removed in February 2025 and later restored. Between January and March 2025 TikTok removed more than 450,000 videos from Kenya and banned over 43,000 accounts; by Q2 removals had climbed to 592,000. In Ethiopia false claims about troops seizing Eritrea’s Red Sea port spread on Facebook before fact-checkers debunked them.

Experts warn the burden falls most heavily on creators and journalists who use local languages. Research groups such as AfricaNLP and academic teams in Pretoria, Nairobi and Addis Ababa are building datasets, and partnerships like Cohere with HausaNLP add data to models. Governments and regulators are also creating strategies to address the gap.