Go and Chess After AI: What Has Changed

Two moments when computers passed humans

In both chess and Go, the first time a computer beat a top human was a turning point. In 1997 the chess computer Deep Blue won a six-game rematch against Garry Kasparov, then the world champion. Go has far more possible moves than chess, and many expected it would take a long time before a computer could beat a top professional. Yet in March 2016, in a five-game match in Seoul, AlphaGo beat Lee Sedol 9 dan four games to one. Game four, which Lee won, is still talked about today. Since then, the way both games are studied and researched has changed a great deal.

What AlphaGo left behind

AlphaGo first learned from human game records and then improved by playing against itself. In October 2015 it beat the European champion Fan Hui 2 dan in all five games, and after Lee Sedol in 2016 it beat Ke Jie 9 dan in all three games in May 2017. That same year AlphaGo Zero was announced, which knew only the rules and learned entirely from self-play, followed by AlphaZero, which used the same approach to learn chess, shogi and Go. A program that started without human knowledge, played moves that contradicted long-held theory and still won, came as a shock to players and researchers alike.

Changes in Go opening theory

The most visible change is in the opening. It used to be widely thought that invading the 3-3 point under a star-point stone early was a loss, because it handed the opponent outside strength. When AI programs began playing early 3-3 invasions freely, the move became common even in professional games. AlphaGo's fifth-line shoulder hit in game two against Lee Sedol also became famous as a move that broke with human common sense. Some long-used joseki have fallen out of favour, and new sequences that AI prefers are studied like joseki. But copying the shape without understanding why it works makes it easy to get lost in what follows.

Changes in chess research

In chess, computers became stronger than humans much earlier than in Go, so the changes have built up over a longer time. Top players prepare openings by checking huge numbers of lines with a computer, and after the game they use the analysis to find their mistakes. As a result, opening preparation has become extremely deep, and moves once abandoned as too risky have come back after computer checking. In the endgame, positions with very few pieces have been calculated all the way to the end and stored, so it is known for certain whether such an ending is a win or a draw.

Reading the evaluation graph

Go and chess broadcasts now often show an evaluation graph calculated by an AI. It helps viewers follow the game, but a few things are worth keeping in mind. Rather than the number itself, thinking about what happened at the moments where it jumped is far more useful for study.

• In Go, win rate is an estimate of the chance of winning, not a point margin
• The same position can get different numbers depending on the program and thinking time
• Humans cannot easily keep finding the best moves the AI picks
• A big jump in the graph does not mean the move was an easy mistake for a human

How study has changed

Once, lessons from strong players and commentary in books were almost the only source of right answers; today anyone can have their own games analysed by an AI. You can see right away where the game tipped, but looking only at the suggested moves from the start skips the process of thinking for yourself. A good order is to review alone first and find the decisive moments, then compare with the AI's judgement. When a suggested move differs from yours, play out the following moves yourself and keep at it until you can explain why it is better. A move you cannot explain is not yet yours.

Tournaments and fair play

As AI became stronger than humans, secretly getting help from a computer during a game became a new problem. In-person tournaments often restrict electronic devices in the playing hall, and online playing sites run measures to prevent cheating. Specific rules vary by event and organisation, so check the official guidance before taking part. Even in a friendly game, keeping an analysis program open while you play is disrespectful to your opponent. Analysis belongs in the review after the game.

Why human games still matter

Computers being stronger does not make human games meaningless. World championships and many other chess events have continued since 1997, and in Go a new generation of players who grew up using AI for research is now active. Human games have stories that come from decisions under time pressure, mistakes and comebacks, and contrasting styles. AI is less a replacement for those stories than a tool for understanding them more deeply. If you are just starting out, learn the rules and enjoy playing against people first, and bring in AI when you reach the point where you need to study.