Learning from online hate speech and digital racism: From automated to diffractive methods in social media analysis

There has been a dramatic surge in uses of big data analytics and automated methods to detect and remove hate speech from social media, with these methods deployed both by platforms themselves and within academic research. At the same time, recent social scientific scholarship has accused social media data analytics of decontextualizing complex sociological issues and reducing them to linguistic problems that can be straightforwardly mapped and removed. Intervening in these debates, this article draws on findings from two interdisciplinary projects, spanning five years in total, which generated comparative datasets from Twitter (X). Focusing on three issues that we identified and negotiated in our own analysis - which we characterize as problems of context, classification and reproducibility - we build on existing critiques of automated methods, while also charting methodological pathways forward. Informed by theoretical debates in feminist science studies and STS, we set out a diffractive approach to engaging with large datasets from social media, which centralizes tensions rather than correlations between computational, quantitative and qualitative data.