> I've never seen any indication that Stockfish can know what a human is thinking, novice or grandmaster.
But it's incredibly easy to train ML models to predict what a human thinks is a good move. Then you just send a signal when "most probable move" is different than "chess engine strongest move".
I have worked on this problem. I would not say incredibly easy. Maia chess (linked in other reply) is pretty good but only has an accuracy of around 55% iirc, and that’s with enough data to train on. Grandmaster level games are rarer and have less data.
But it's incredibly easy to train ML models to predict what a human thinks is a good move. Then you just send a signal when "most probable move" is different than "chess engine strongest move".