Hi, reading more about Gmail Smart Reply in https://www.blog.google/products/gmail/save-time-with-smart-reply-in-gmail/ i got a new theory:
The AI learns from the user's mail habits to be able to propose three quick replies in the personal writing style of the user. If you annoy the AI from outside (see also "Goozim" :)) then it will lure the user into such a quick reply. So the initiator learns a brief psychological profile of that user in respect to obvious vulgarity and spam. By fake mail headers in the original poking mails, the reply then appears here. The annoyer may be subscribed or may simply harvest the replies from our web archives. Have a nice day :) Thomas