But also, the percentage of the userbase responding poorly also responded almost an order of magnitude more when the person on the other end was a woman.
That's not what their data is showing. That would be a valuable break-out to see.
What their data is showing is number of messages. If you look at the chats, it appears they're counting multiple abusive messages (one minute to the next and many within the same minute stamp) from single actors.
If one jerk sends eight abusive messages in one minute, they're counting all of those in the order of a magnitude you're referring to. It makes sense the abusive actors will send as many abuse messages as they can as quickly as they can, as one would expect that to get shut-down.
Simply put, the abuse they're showing, which is about ~14 rude + harrassment messages per week to Julia, could easily consist of just two or three individual jerks sending a burst of messages (which again appears to be the case based on the screenshots). By their numbers, they're getting ~4,500 visitors per week. Out of those, it appears that about 0.1% or less of those visitors are likely a problem.
Ok, then I'm confused. In the post it shows, for example this:
https://i.imgur.com/P4QCWO4.png
That says ~14 messages per week. So you're saying the labeling is wrong, that it should say chat sessions per week?
The overall math in that scenario is also confusing me, if that refers to sessions. If the average abusive session contained even just six or seven messages, times 14 abuse sessions per week, upwards of 100 abuse messages contained within the abusive sessions in a week, ~400 per month. Roughly half of your customer interaction messages, or more, at that time, were occuring in abuse threads? Or does "2,100 messages" actually refer to 2,100 individual chat sessions?
The average "abusive session" contained only about three or four messages, not 6-7. Otherwise, your math is right.