After game analysis. When someone picks the computer best most from a position you can ask how likely it is they came up with that on their own. Expert chess players recognize computer moves fairly accurately in a lot of situations (this is easy to test, just take a bunch of positions, show the last move and ask if it was human or computer). From there human judges look at the game and decide how likely it is someone cheated.

Of course if you only cheat once you get away with it as nobody will be sure. However cheating once isn't enough to change a game - in fact it will probably change it for the worse on your side (without knowing what the computer intended you follow a bad line after the otherwise good move). Thus you can look at a series of games and concluded if the player cheated or not. From there you take after the fact adjustments which mostly means recalculate everyone's score as if those games didn't happen.

Interestingly enough, chess masters generally will score alpha zero moves as human not computers. It will be interesting to see how online cheating progresses once better AI based chess computers become available.