I rather expect that neural networks have made software detection of cheating more difficult than before. In chess, anyway, one thing that distinguished computer play from human play was the deep calculation of variations. Chess engines, and the human cheaters that used them, would make plays that didn't make much sense to human players unless they were based upon deep calculation. The same cannot be said of neural network go bots. In fact, one way to tell a bot is that it misreads a ladder!
This is well known, so a cheater is unlikely to have the same tell. There are other things that bots do that you could program a computer to find. They like to invade on the 3-3 point, they play new joseki and avoid some of the old joseki, they don't pincer very much, they don't extend on the side or make splitting plays on the side very early in the game, etc. The problem with detecting cheaters is that all of these are learnable by humans. And a suspected cheater may be playing better than he or she did before, because of learning those things. It is not easy to detect cheating by looking at game records.
