Do you suppose they are running the messages through any LLM? I don't know. I would guess it's too much volume to run all mail through a "good" model, but no idea whether it would be feasible to run mail through the kind of dumb model that generates "AI Overviews."

Oh for sure. I have even been solicited by companies that do such a thing, and brag about how they reword every email so it doesn't get flagged as spam.