The bot isn’t really cognizant of its own replies.

…Well, it could be with a postfilter, ranging from sanity checking its own reply, to a tiny separate safety model (of which many exist), to a basic keyword check. But apparently Twitter has no resources for such things.

But the point is this isn’t really the intended design.