If the majority of online text is written through the lens of political American subculture, then it stands to reason that models trained on it would reflect those values.