Algorithmic blind spot: (In)visible racism, multimodality and the limits of AI hate speech detection

Automated hate-speech detection has advanced rapidly, yet significant blind spots remain in how artificial intelligence recognises online abuse and processes its contextual and culturally situated meanings. This article examines why anti-Roma hate often evade moderation on social media. Drawing on insights from multimodal discourse analysis, pragmatics, and critical race theory, this article shows two mechanisms through which anti-Roma hate escapes detection on YouTube. Some expressions are semiotically invisible to commercial AI systems, which cannot recognise humour, irony, euphemism, metaphor, or multimodal pairings of text and image. Others, which are socially normalised , are excluded from algorithmic or policy definitions of hate. This article conceptualises this tension as the algorithmic blind spot , a structural limitation of current AI hate speech detection models rooted in Western-centric datasets, narrow operational definitions of harm, and decontextualised text models; it argues that understanding the (in)visibility of online hate requires treating racism (or any other form of hate) as both semiotic and technological phenomena.