Probably expected just due taking a model trained on all sorts of content and fine-tuning specifically on non-fiction data.

Wouldn't conclude anything about /pol/ in particular without at least comparing the same done for HackerNews/Reddit/etc.