Sadly, the true mathematical opposite of a good, politically-slanted video in recommendation feature-space, probably isn’t one that is good and politically-slanted in the other direction; but rather, one that’s both apolitical and very bad.

To a recommender model, the clusters for “left-wing” and “right-wing” content aren’t on opposite ends of the feature cloud; they’re likely very close neighbours.

A recommender algorithm would have to be a lot smarter to realize that some particular neighbouring cluster has an adversarial (i.e. “watching X makes you less likely to watch Y, without that just meaning that X and Y are redundant”) relationship with the one it’s looking at.