I can say quite a lot about players and students shouldn’t use AI to learn, like you shouldn’t use AI to judge every 0.1 or 0.5% winrate loss or even just mistakes differ less than 2 or 3 points. Or try to research branches in joseki that have similar winrate, and just pick whatever without understanding the surrounding stones associated with each choice, or just try to justify overplay in your own games, despite it creates early fighting or variations that you obviously aren’t prepared for (like don’t jump out and simplify the situation and insist on fighting it out, without the reading skill). If you cannot follow the logic, even for pro’s moves/tesuji, that would confuse you and even make you play worse, since if you cannot internalize them, it would simply put doubt in your own judgements.

You have to be able to play them out much further down the line to see where the punishment lies in these recommendations, or they won’t mean a thing or be helpful. A number is just a number and during the game, you won’t magically get a sense of these numbers in your head. You will still need to see the shapes and reading them out. Help to find local weak spots are probably the best use. But be sure to settled positions elsewhere, or use localized suggestion filters, to help only finding those local position analysis. For the whole board judgement, players usually need to be high dan or above to really appreciate the reason for early tenuki, and leaving aji.

A bit late, but might as well reply anyway. It’s been a while since I played much, but when I used AI to train, that “main problem” came up a lot for me too. For me, one of the key things in using AI to review is to “play” against the AI, like when you want to explore a tesuji it comes up with, try to read to think of possible follow-ups before you even click the board, so it can’t show you the continuation. It’s the same idea as a tsumego, if you just click without thinking it’s much harder to come away with more understanding.

But a lot of the “ideas” the AIs have, especially in complicated fights, are just past my understanding too but that’s okay! It also might be that the move the AI prefers the most isn’t the best move to play at your/my level. To exaggerate a bit, if a specific attack gains 1/2 of a point but you have to read a 20-move-deep sequence in order to play it or your group could die, you obviously can’t play it if you don’t understand.

I’ve used the AI to refute or confirm ideas I’ve had myself, to train myself on fuseki/joseki and game direction, and for moves the AI thinks are great that I’ve missed, I try to explore why that might be. But I don’t worry about the times I just don’t understand why a move is better if I can’t figure it out, as long as I come away from the review with some greater understanding, it’s a success! (and maybe it’ll be time to understand that idea later, when I’ve improved)

One thing AI has taught me is that it’s not really about the specific move (unless it’s life and death), but about direction. Almost always playing roughly in the correct direction results in +/- 2 point changes at best, whereas wrong direction, even if it’s locally good can easily cost you 5-10 points. It really helped me stop fussing over the “correct sequence”/joseki/fuseki and focus on mistakes that were actually costing me the games.

For training my intuition I find replaying/memorizing pro games is still far more effective, since the moves follow human reasoning and shapes. AI seems to work best as a review tool for finding/exploring mistakes.