Ethiopia: Research Finds Amharic Language Used to Evade Tiktok Moderation, Bypass Hate Speech Detection

Addis Abeba– A study by the Institute for Strategic Dialogue (ISD) found that the Amharic language is being used on TikTok to “bypass moderation” and “evade hate speech detection” by exploiting gaps in the platform’s content moderation systems.

The research identified 134 comments where Amharic was used in two ways: the first method, referred to as “coded,” involved “directly translating hate speech into Amharic using the Ge’ez script.” The second method, called “bypass,” involved “placing Amharic text alongside hate speech written in a European language.”

The report states that “Of 16 comments reported to TikTok, the platform removed or limited the visibility of only 5.” It added, “This suggests that TikTok’s systems are failing to recognise hate speech, even when written in English and manually reported, if it appears alongside Amharic language text.”

ISD noted that “previous studies on computing and natural language processing have pointed to the challenge of policing hate speech by native Amharic speakers, reflecting technical challenges and inequalities in the resources platforms allocate to different languages.” It explained that Amharic is a “low-resource language,” meaning there is less online data available to train natural language processing systems, which platforms rely on for automated moderation.

The study also found that similar tactics were being used in other “low-resource” languages, including Punjabi, Nepali, Konkani, and Tigrinya, suggesting that “bad actors are innovating and exploiting this vulnerability.”

ISD urged TikTok to “first reevaluate its moderation systems” so that “‘bypass’-type uses of non-Roman alphabets do not allow hate speech in high-resource languages to escape moderation at both the initial time of posting and after being reported.”

It further called on TikTok to “focus its computational and natural language processing capabilities” on non-Roman alphabets and low-resource languages, warning that current gaps “allow bad actors to rebuild networks and account followings” even after suspensions.