Did superhuman AI actually improve Go players' decision quality?

Decision quality among professional Go players rose after superhuman AI arrived, and the rise came with more novel play. The authors date the "advent of superhuman AI" to a series of events between 2016 and 2017, anchored by AlphaGo's March 15, 2016 defeat of a human world champion. Across more than 5.8 million move decisions from 1950 to 2021, they score each actual move with KataGo, a superhuman program. It simulates 10,000 game patterns per decision and compares the win rate of the human move with that of the counterfactual AI move, producing a Decision Quality Index. The excerpt reports that "humans began to make significantly better decisions following the advent of superhuman AI." Novelty, measured as each game's first historically novel move, also rose: novel decisions "occurred more frequently and became associated with higher decision quality after the advent of superhuman AI."

The mechanism the authors propose is that superhuman AI "can ultimately unearth superior solutions previously neglected by human decision-makers who may be focused on familiar solutions." Because AI can identify optimal decisions "free of human biases (especially when it is trained via self-play)," its moves widen the options players consider, and the paper suggests they "induced them to explore novel moves." The main alternative explanation is memorization of AI moves. The authors answer it by excluding every human move that matched the optimal AI decision, and a "sharp increase in decision quality" remains. They stop short of making novelty the whole story: the improvement "may be partly explained by increased novelty."

The library's nearest notes frame human-AI interaction differently. In the Atria Dawn analysis, agents propose methods while humans keep most final decisions and steer exploration, so exploratory judgment stays with people. The Go paper shows a different locus: the human side of exploration itself widened after a machine showed unfamiliar moves. The hybrid-society study (Do humans learn to prefer AI partners over time?) also shows humans adapting to AI through repeated interaction, but the adaptation there is partner choice, where here it is move choice and its measured quality. The long-horizon agent study (Do frontier AI agents actually conduct novel research or just optimize?) finds genuine novelty rare in agent output. Read together, the cases suggest novelty has shown up in human play after AI exposure but not in agents' own work. That is a contrast of scope, since the Go excerpt does not test agent novelty.

The excerpt establishes an association, not the full causal chain. The authors list as open whether AI raised novelty and thereby raised quality, and through which other mechanism it might have acted. The quality measure is also a machine estimate: moves are judged better by KataGo's win rates, not by a proof assistant, expert review or players' own explanations. The excerpt does not say that players understood why the new moves were better. It can show that Go decisions got better and that novelty tracked the change, but not that the people making those decisions grasped the reasons. At the strength this evidence allows, "better" here means better by a superhuman evaluator, and whether the improvement was understood is a separate question the excerpt leaves open.