This paper asks, what if we look into less analysed forms of hate speech instead of well-studied types? We explore the issues encountered when trying to identify ‘fringe hate speech’ with AI-enabled tools on social media. With fringe, we mean online hate speech targeting a specific identity group that has not gotten much attention, is understudied, and, therefore, is less detected in systems for online hate speech detection. The study is based on a case study of developing a multilingual and multiplatform hate speech detection tool. We present two examples of fringe hate speech: 1) misogynoir (racism against Black women) and 2) anti-Palestinian racism. Online hate speech detection relies on a sociotechnical ecosystem, often emphasising scrutinising the technologies applied. However, the social aspects of including humans in the loop for annotation and content analysis also influence identifying forms of hate speech and lead to biased detection rates. We propose a list of fallacies for setting an agenda for discourse and further research.