I've intentionally been running it now just to see how wrong its summaries of messages are. So often, it relays the exact opposite of the intent and I want to understand why.
My impressions are the exact opposite: it perfectly summarizes everything to the point like 98% of the time at least.
Yep, that's my experience as well. I take screenshots (and report them) when I see a wonky one, which has happened twice.