Humans have improved at Go since AIs became best in the world

AIs can beat the world’s best players at the board game Go but humans are starting to improve too. An analysis of millions of Go moves has found that professional players have been making better and more original game choices since Go-playing AIs overtook humans.

Before 2016, AIs couldn’t beat the world’s best Go players. But this changed with an AI called AlphaGo developed by London-based research firm DeepMind. AlphaGo defeated multiple Go champions, including the then number one ranked human player.

Since then, other AIs have also been developed that are considered “superhuman”. Though they can be used simply as opposition players, they can also help analyse the quality of any given move and so act as a Go coach too.

Minkyu Shin at the City University of Hong Kong and colleagues decided to investigate whether the introduction of these superhuman Go-playing AIs has led to a marked improvement in human play.

The researchers gathered a dataset consisting of 5.8 million move decisions by professional players between 1950 and 2021. They then used a Go-playing AI to help calculate a measure called a “decision quality index”, or DQI, which assesses the quality of a move. They deemed a move “novel” if it had not been previously attempted in combination with the preceding moves.

The analysis found that human players had made significantly better and more novel moves in response to the 2016 advent of superhuman AI. Between 1950 and 2015, the improvement in quality of play was comparatively small, with a median annual DQI oscillating between roughly -0.2 and 0.2.  Whereas after superhuman AI, the DQI leapt upward, with median values above 0.7 from 2018 to 2021.  In 2015, 63 per cent of games showed novel strategies, whereas by 2018, that figure had risen to 88 per cent.

Stuart Russell at the University of California, Berkeley says that the improved human Go playing resembles a phenomenon in the 1990s, when backgammon players began changing opening moves in response to the advent of highly skilled computer players. The fact that it’s an AI making the assessment also plays a role, he says. “It’s not surprising that players who train against machines will tend to make more moves that machines approve of.”

The paper shows cultural transmission from Go-playing AIs back to humans, says Noah Goodman at Stanford University in California.  “The thing [the paper] makes me think really hard about is that right now we’re seeing another abrupt change in AI which is chatbots.  What abrupt changes are we going to see in different cultures as a result of interacting with and learning from chatbots?”

Topics: