The irony is that these machine learning algorithms are effectively coincidence machines. There's no rhyme or reason as to why they work once you have even just a handful of neural net layers, and people invest quite a lot into these magic recommendation machines.

As for toxicity, I don't think there's one magical root cause for all of it. And I'd argue that most features and designs to increase engagement _are_ one of the many causes for people to create lucrative polarizing digital content.

edit: Also I'm not sure how my original post was confusing correlation with causation. You can replace "correlation" with "coincidence" there without issue.