Abstract
To investigate ideological symmetries and asymmetries in the expression of online prejudice, we used machine-learning methods to estimate the prevalence of extreme hostility in a large dataset of Twitter messages harvested in 2016. We analyzed language contained in 730,000 tweets on the following dimensions of bias: (1) threat and intimidation, (2) obscenity and vulgarity, (3) name-calling and humiliation, (4) hatred and/or racial, ethnic, or religious slurs, (5) stereotypical generalizations, and (6) negative prejudice. Results revealed that conservative social media users were significantly more likely than liberals to use language that involved threat, intimidation, name-calling, humiliation, stereotyping, and negative prejudice. Conservatives were also slightly more likely than liberals to use hateful language, but liberals were slightly more likely than conservatives to use obscenities. These findings are broadly consistent with the view that liberal values of equality and democratic tolerance contribute to ideological asymmetries in the expression of online prejudice, and they are inconsistent with the view that liberals and conservatives are equally prejudiced.