Submitted by Richard Arlett on Mon, 19/01/2026 - 14:04
Report from the symposium “Critical Frontiers in Hate Speech and AI Research"
By Petre Breazu and Napoleon Katsos (University of Cambridge)
Why is hate speech so widespread on social media, yet so rarely flagged or removed by platform moderation systems? This question framed the one-day symposium Critical Frontiers in Hate Speech and AI Research, held on 6 November 2025 at Trinity College, University of Cambridge. Supported by Cambridge Language Sciences, the event brought together researchers from linguistics, media studies, and artificial intelligence (AI) to examine the limits of current approaches to AI-driven moderation and platform governance.
A central theme of the symposium was the recognition that, despite rapid technological advances, automated hate speech detection continues to suffer from significant algorithmic blind spots. Participants highlighted that these blind spots are not only technical but also conceptual, as they stem from narrow definitions of harm, decontextualised text-based models, and limited attention to how meaning is produced in real-world digital environments. As online communication becomes increasingly multimodal and shaped by platform-specific norms, existing moderation systems often struggle to identify indirect, coded, or socially normalised forms of hate.
The keynote programme addressed these challenges from complementary interdisciplinary perspectives. Ariadna Matamoros-Fernández (University College Dublin) introduced the concept of platformed racism, showing how racism is co-produced through platform design, governance practices, and user cultures, rather than simply expressed by individual users. Her talk highlighted why AI moderation systems cannot be understood as neutral tools but must be situated within broader socio-technical infrastructures.
Kay O’Halloran (University of Liverpool) followed with a keynote on a Multimodal AI Approach to Platformed Hate, arguing that effective detection requires embedding social semiotic theory into AI systems. Drawing on examples from multimodal analysis, she demonstrated how hate emerges across language, images, and other modes, and why reducing harmful content to isolated linguistic features is inadequate.
Further insights into the intersection of hate speech, disinformation, and migration were offered by Encarnación Hidalgo-Tenorio (University of Granada), Aritz Gorostiza (University of Málaga), and Juan-Luis Castro-Peña (University of Granada). Using the semi-supervised algorithm Nutcracker, they showed how fake news and conspiracy narratives portray migrants as economic burdens, security threats, or culturally incompatible “others”, narratives that reinforce exclusionary ideologies but frequently escape standard hate speech detection systems.
These perspectives were developed further in two thematic panels on Discourse, Normalisation, and Platformed Hate and Algorithms and the (In)visibility of Hate, which examined annotation practices, platform standards, and the societal consequences of what AI systems fail to detect.
The symposium concluded with a roundtable discussion reflecting on freedom of expression, platform accountability, and interdisciplinary collaboration. A shared conclusion emerged: addressing online hate requires treating racism and discrimination as both semiotic and technological problems. We are grateful to Cambridge Language Sciences for supporting this event, and a series of publications is currently in preparation.
Funded by AI@Cam through the Cambridge Language Sciences (CLS) project “Improving Language Equity and Inclusion through AI”.