Breazu, P., & Katsos, N. (2026). Algorithmic blind spot: (In)visible racism, multimodality and the limits of AI hate speech detection. Discourse Studies.

Link: https://doi.org/10.1177/14614456261487859

Open access: Yes

Notes: AI is deeply intertwined with current models of platform governance, particularly in content moderation. But AI is far from infallible. IN this article, Breazu and Katsos discuss the “algoritmic blind spot,” centring on the “structural inability of current AI moderation systems to recognise racism that is either semiotically complex or socially normalised” (p. 2). To explore this, they analyzed anti-Roma gatekeeping in YouTube comments. IN this way, they identified two mechanisms to conceptualize the algorithmic blind spot: semiotic invisibility (where “explicit hate is recontextualised into a coded, affectively charged form that conceals aggression under the guise of humour” (p. 12)), and social normalization (where, through explicitely harmful language, “it reinforces long-standing racial stereotypes” (p. 12). In this way, the deployment of AI to moderate which content belongs on YouTube fails to properly identify harmful discourse that perpetuates and extends racist ideologies—a blind spot that is “reinforced by the dominance of text-centric models trained on large-scale, decontextualised datasets” (p. 14)