It might be the case that the models internally predict the nationality of the user given other indicators, e.g. tells in the way they write.