That's a good question. I think that chess computers win by 'studying' thousands --possibly millions-- of matches, including many classic moves and beginning strategies. The one thing a computer does better than a human brain is calculate large and complex probabilities in nanoseconds. They are hyper-fast pattern-recognition engines.
I mentioned elsewhere an article about a Go master who beat a champion-level computer in 14 of 15 matches by using a strategy that was so simple and basic that no player worth his salt would ever use it. To a human, the strategy was both obvious and blatant, and very easy to counter once you recognized it. But the computer had no matches to study, and struggled to find a way to counter it.
Now, I think that this is partly lazy programming. And experiences like that can teach AI programmers how to better prepare an AI bot to learn and improve. And that might be the real random factor here: we don't know how well any AI has been programmed until we have enough experience working with it to see what it does. Which is probably a scary prospect once we start giving these AI bots important jobs to do.