*Result*: NLP as language ideology: discursive and algorithmic constructions of 'toxic' language in machine learning research.

*This article considers natural language processing research as a language-ideological practice, looking specifically at the task of toxic language detection, which impacts nearly everybody online through automated content moderation. Industry discourse constructs the category of toxicity through a series of oppositions between civil/healthy/referential/rational and unhealthy/toxic/indexical/emotional. Examples from a toxicity correction dataset demonstrate how this ideology can become encoded algorithmically: a focus on preserving referential content in text "detoxification" results in neglect of important poetic, expressive, and social-indexical functions. Overall, the discursive framings of toxicity construct the ideal speaker in terms of what I call a "referentialist" language ideology, which values rational debate in the (regulated) liberal-democratic public sphere. Ultimately, toxicity detection and other metapragmatic tasks do not merely model the existing pragmatic categories but actively construct them. Toxicity in particular potentially reinforces exclusionary norms of white maleness and promotes online subjectivities that are useful (profitable) to the commercial platforms that shaped the task. Since there is no avoiding NLP as language-ideological practice, independent NLP researchers must acknowledge the political potency of their work by continually reflecting on the categories they work with in relation to models of social-political formation. [ABSTRACT FROM AUTHOR]*