From Detection to Counterspeech: Auditing AI Moderation and Fact-Checking Practices in Ethiopia’s Multilingual Online Sphere

AI-assisted content moderation promises to manage online incivility at scale while sparing human moderators traumatic content. Yet most evidence comes from Western languages and contexts. This study examines how AI hate-speech detection interacts with counterspeech and fact-checking in Ethiopia’s polarized Amharic and Afan Oromo spaces. Drawing on infrastructural invisibility and civic labor, it triangulates a computational audit of three classifiers against 838 hand-annotated posts (2020–2025), document analysis of platform self-descriptions and transparency reports, and 20 interviews with Ethiopian fact-checkers, volunteer flaggers, and counter-speakers. The most widely used generic classifier cannot read either language natively and, on English translations, recovers only about a tenth of hate speech; locally-oriented classifiers perform far better on Amharic but collapse on Afan Oromo, leaving it effectively unserved. Across every tool, the dominant error is under-detection, not over-removal, and the same silence recurs in platforms’ self-descriptions. Rather than sparing anyone the work, AI’s failure displaces it onto an unpaid volunteer ecology that absorbs political and psychological costs platforms benefit from yet refuse to name.