Well, you don't know how it was targeted. Maybe they don't want to sing to the choir so such ads might be targeted at people who are interested in evolution. Still it seems unlikely that the scientific-ness of something can be determined by an AI model at all, let alone just word vectors. What is and isn't science can be hard to rigorously pin down, that's why pseudoscience is problematic in the first place.
These are all reasonable points, but since I have lots of experience building recommendation systems at places like google, I have a pretty good understanding of what the embeddings are capable of learning (even so, Google News still does the same thing occasionally).
