Study of 322 Papers Finds Computer Science and Social Sciences Split on Defining Hate Speech
A new open access paper from Dr Stefanie Ullmann of the Minderoo Centre for Technology & Democracy (MCTD) at Cambridge finds that the way researchers define hate speech technically has direct consequences for what automated systems can detect. Her analysis of 322 research papers shows two discourses that are entangled but incompatible.
- Dr Stefanie Ullmann of MCTD Cambridge published an open access paper on how hate speech is conceptualised in AI detection research.
- The study analyses 322 papers from Computer Science and from the Social Sciences & Humanities using a corpus-assisted discourse studies (CADS) approach.
- Computer Science treats hate speech as discrete labels on snippets of language suited to benchmarking, while the Social Sciences and Humanities see it as a social practice.
- Ullmann argues these framing choices determine which harms become optimisable tasks and which stay outside the frame.
- She does not argue that computer science should look like sociology or discourse analysis.
Two ways of seeing hate speech
The study, titled "Language as Data, Hate as Task: The epistemic normalisation of hate speech in AI", uses a corpus-assisted discourse studies (CADS) approach. It examines how language is conceptualised in hate speech detection research, drawing on 322 papers from Computer Science and from the Social Sciences & Humanities.
According to the MCTD, the work identifies two distinct discourses. In Computer Science, hate speech is operationalised as discrete labels attached to individual snippets of language, a format suited to benchmarking. In the Social Sciences and Humanities, it is treated as a social practice embedded in histories of racism, sexism and platform governance.
Why the framing matters
Dr Ullmann argues that these are not simply neutral side effects of different disciplinary priorities. According to the paper, choices about how to conceptualise language, bias and racism determine which harms are turned into optimisable tasks and which remain outside the frame when hate speech is detected online.
In her words: "Under those conditions, choices about how to conceptualise 'language', 'bias' and 'racism' are not neutral side-effects of disciplinary focus; they determine which harms are rendered as optimisable tasks and which remain outside the frame."
In practice, this means a system built around labelled snippets can only act on what that framing makes visible. Harms understood as part of wider social histories or platform governance may not fit into a labelling task at all.
Not a call to turn computer science into sociology
Dr Ullmann is explicit that the paper does not ask one field to imitate the other. She says she is not claiming computer science should "look like sociology or discourse analysis". Her argument is instead that "the way 'technical problems' are defined has direct consequences for what can be seen and acted on as hate speech."
The paper is available open access, and the MCTD links to it as Stefanie Ullmann, 2026.
What it means
The findings point to a question that sits upstream of model performance: how a problem is defined before any system is built. For researchers and developers working on automated hate speech detection, the study suggests that benchmark results reflect the assumptions built into the task, and that those assumptions decide which harms get addressed. The source does not provide specific recommendations or performance figures, so the practical implications remain a matter for further work by the field.
Minderoo Centre for Technology & Democracy: Language as Data, Hate as Task: The epistemic normalisation of hate speech in AI
