New research raises concerns about the effectiveness of AI moderation tools in detecting hate speech in Ethiopian languages, particularly in the context of the ongoing violence in the Tigray region. Platforms like Facebook have been accused of inadequate moderation of hate speech targeting Tigrayans, and a lawsuit has been filed against Meta (formerly Facebook) for failing to take action against death threats posted on its platform. The research highlights the limitations of multilingual language models, which rely on cross-lingual transfer and can result in errors and false alarms when dealing with low-resource languages. The study emphasizes the need for language-specific and culturally-specific content moderation in marginalized languages. The report does not draw conclusions about Meta or the Tigray conflict specifically but highlights challenges in handling languages like Tigrinya, including missing characters in language models. The research suggests that social media companies should work with language communities and researchers to develop more effective moderation approaches tailored to specific languages. While some startups are working on solutions for low-resource languages, major tech companies still prioritize multilingual approaches despite their limitations.

Read more on Rest Of World