"we estimated the time trends of decision quality only for the decisions that differed from the optimal AI decisions." which I take to mean that the graph above shows only the improved decision quality of moves that are not direct memorizations of moves done by AIs?
Graph of decision quality among professional Go players. A sudden increase after AlphaGo. It is not only because they are learning from the AI. Players are suddenly inventing new moves at a faster rate too!
Ok, so this is interesting: it seems maybe the quality improved not after AlphaGo, but after the release of open source versions. Because they allowed players to look under the hood and understand the strategy. Just seeing the AI moves was not enough.
There is a parallel here to cognitive apprenticeship theory, ie the idea that the reason cognitive skills are hard to learn is that we can't see them so you should externalize them for students. For humans to learn from AIs, perhaps making reasoning explicit is needed.
The data is from Games of Go on Disk which transcribes games from pro tournaments. I suspect that it is pretty hard to cheat on a live pro tournament, but not sure what precautions they take.