Edit: Apparently single-threaded stockfish is deterministic. Maybe my experience with non-deterministic chess engines just has to do with the handicapping related to providing easier levels of play.
- Opening Book learning. - Position Learning.
Most decent chess engines have these features.
Humans wouldn't want their opponent to play the exact same way every time. For a simple example, playing white, one wouldn't want the computer to always open 1. h4.
So, either the computer has to pick an opening, or the human opponent has to pick one.
Of course, that extends past the first move, so, assuming the computer to play deterministically, one would end up with the human opponent having to specify "let's play opening X, variant Y, as typically played by Z up to black's 15th move". It probably is easier to use a "set up a custom board" feature than to specify that.
And it wouldn't even stop there. The human specifying that particular opening likely wants to study that opening, not just the single continuation that the computer thinks to be the best one.
Also, as another poster said, a chess engine may use Monte Carlo simulation to pick a move, and that means using a random number generator.
My experience with chess engines being non-deterministic is that if I pick the same opening I don't always play the same game with the same engine. I really might just be off here, and that variance might be related to random artificial handicapping.
e.g. 1.e4, d4, c4 and Nf3 will get picked more often than 1. b3, c3, nc3 and f4, and 1. g3, e3 even less (or not at all).
If you always win with the same game, then instead of AI, we are rather talking about AS (Artificial Stupidity).