AI models choose violence ?

[Maxmars]

I've been thinking about Kenzo's question: "AI models choose violence?"  and it occurred to me that given the "programmatic" nature of any presumed "A.I.'s" existence... is it really "possible" that A.I. can "choose" anything at all?  

A.I. is entirely mathematical... no aspect of it is not.

So, can A.I. "choose" or is it more like a little ball "choosing" where to land on a roulette wheel?