Flyspell marks for example the second "AND" in "an AND split and AND join"
as a duplicate.

This seems to be deliberate: in line 1153 of flyspell.el (on current
emacs-26 branch) flyspell-word-search-backward is called with IGNORE-CASE
set to t.

I don't understand this: if it's not the same case, it's probably
deliberate, no?

I checked that simply removing this `t' argument makes doublon detection
work as I'd expect: "AND AND", "and and" and "And And" are all marked as
doublons, but anything that's not the same case isn't.

I'm using Emacs 26.1.

-- 
https://rrt.sc3d.org