When AI hands experts stronger moves in Go, do they come to understand the reasoning, or just play better?

This explores whether people actually grasp the reasoning behind an AI's stronger moves, or whether they just get better results by following or absorbing what the AI does. The short answer from this collection is that it tells us more about whether humans *improve* than about whether they *understand*. Those are different questions, and the space between them is where the interesting material sits.

The strongest evidence comes from Go. An analysis of 5.8 million professional moves from 1950 to 2021 found that decision quality rose sharply after AlphaGo in 2016 Did superhuman AI actually improve Go players' decision quality?. The telling detail is that players also began making more *novel* moves, and that novelty explains part of the gain even after excluding moves copied directly from AI. That rules out pure imitation. Players weren't just replaying the engine's lines. They were producing new play of their own. That suggests some real absorption of what makes the AI's choices good. But the study measures outcomes, not explanations. Whether players could say *why* a move works, or had only built better intuitions, remains open.

The gap matters because getting better without understanding has a known downside. One framework describes how people mistake fluent AI output for sound reasoning and then reinforce their own biases, with each trap amplifying the others Why do people trust AI outputs they shouldn't?. In games, that looks like a player trusting a move because it came from the engine, not because they see its logic. It's also worth knowing that language models aren't strategic oracles. They often miss game-theoretically optimal play as games get more complex, and they recover only when guided through structured reasoning steps Do language models make rational strategic decisions in games?. So 'AI-suggested' doesn't automatically mean 'superior'. Go engines trained through self-play are a different kind of system from a chatbot giving advice.

The more surprising thread is that understanding may depend on *how* the AI hands over its advice. Work on 'learning to guide' finds that machines pointing out which parts of a situation matter, without issuing a verdict, reduce people's tendency to anchor on the AI's answer Can AI guidance reduce anchoring bias better than AI decisions?. A lab study found that assistants pairing advice with reflection questions beat those that only advised Do reflection questions help people make better decisions with AI?. Formal argumentation structures go further. They lay out a decision as a map of claims and counterclaims, so a person can point to the exact premise they doubt Can formal argumentation make AI decisions truly contestable?. Read together, these suggest understanding isn't a fixed property of the human. It's a design choice, and a bare 'play here' recommendation is the format least likely to produce it.

There is also a reverse direction. Instead of humans learning to read AI strategy, AI can be steered toward strategies humans recognize. Just thirty minutes of human demonstration, used as a light constraint during self-play training, pulled an agent toward human-compatible behavior Can human data steer self-play RL toward human-compatible behavior?. That raises a real tradeoff: AI strategy that humans can follow may sometimes mean giving up some of the alien edge that made it superior in the first place. The collection doesn't yet have a direct study of whether experts can articulate the reasoning behind superhuman moves. That remains a gap.