AI has dominated chess for 25 years, but now it wants to lose

Maia Chess is a project looking to make AI more human-like in its actions, competing on a more level playing field.

Try 3 issues for £5 when you subscribe to BBC Science Focus Magazine!
Published: February 14, 2023 at 7:00 am

Way back in 1985, a team of researchers at Carnegie Mellon University developed a computer purely to play games of chess. After moving to IBM, the computer was further developed, culminating in the obvious test – a match against then-world champion Garry Kasparov.

However, the computer known as Deep Blue at this point wasn’t enough for Kasparov; it lost four games to two. But like any good underdog, the computer was down but not out. It came back a year later to beat Kasparov in a narrow victory, winning by a single game.

Garry Kasparov playing his match against Deep Blue © STAN HONDA

Since this monumental moment, artificial intelligence (AI) has had an iron grip on chess, and even Go – one of the oldest games in history. As artificial intelligence improves, it continues to outperform human ability, even at the highest level of chess, beating even those players unbeatable by humans.

But as companies pour money into making AI the greatest chess player it can be, one company is taking a different route, making a chess-playing AI that is as bad as us mere mortals.

Maia Chess is a project created by a team of researchers from the University of Toronto, Cornell University and Microsoft Research back in 2020. They aim to make an AI that ‘plays the human move – not necessarily the best move’.

As a result, this would be an AI that is more human-like in its decisions. It can make human mistakes, won’t constantly be accessing the millions of possible outcomes, and overall, won’t be quite so daunting to play against.

The model learns from online human games instead of from self-play. It has been trained on millions of games, attempting to predict what the human move would be in every position.

However, teaching a model on millions of games, all at vastly different levels would create a human-like AI with a somewhat confused psyche on how to play chess.

To avoid this, the team found the best performance when personalising Maia to certain levels. The team trained nine versions of Maia, one for each of the Elo milestones (professional chess rankings) between 1100 and 1900.

Each of these versions learned from 12 million human games, understanding how to play like a human would at that level. By measuring how often Maia’s predicted move was the same as a human, they can accurately understand the kind of mistakes humans make at each level, and at what point humans stop making these mistakes.

This isn’t just a model that can play like a human would, it is also an insight into the way that humans play chess. Maia can accurately predict upcoming mistakes and understand why a player is losing their games.

So what’s the point? Why would we want an AI that plays like us? The team behind Maia hopes that this can be used as a tool for teaching. By understanding the common mistakes of players at your level, and being able to predict your mistakes, Maia could be used to give feedback and aid your ability.

However, this wouldn't be the first time AI had been utilised as a teacher and the results aren't always perfect. "AI has been used for the purposes of education for many years with mixed results. It’s always been difficult to measure exactly how effective an AI is at educating a person," says Prof Peter Bentley, a computer scientist and author based at University College London.

"Many AI researchers get a little distracted by their usual measures of 'accuracy' or 'prediction' by the AI and do not properly assess the benefits (if any) to the student. However, with the advent of Large Language Models (ChatGPT, for example) there’s huge excitement about the potential."

The question is less about whether Maia can improve your chess ability, but more about whether it can actually be an effective teacher.

"It’s always going to be hard to beat a human teacher, who can work with us, judge our moods, our abilities, our likes and dislikes, and be motivating and inspiring such that we love to learn with them," says Bentley.

"Will an AI ever be that inspirational teacher you remember from school? I’m not sure."

Maia has also been set up to allow players to have games against an AI that is more realistic of the human games that they will have. In the future, this technology could be used to predict the moves of an individual player, learning their weaknesses and common mistakes.

Of course, not everyone is looking to improve their chess game, but this style of AI could be adapted. Instead of using AI as a method of beating humans or doing their menial tasks for them, Maia Chess shows how AI can be used as a teacher.

Want to try your hand against Maia? You can play against it at any level between 1100 and 1900 for free. Simply open Maia's Lichess page and click the two crossed-over swords in the top right-hand side.

About our expert, Peter Bentley

Peter is a computer scientist and author who is based at University College London. He is the author of books including 10 Short Lessons in Artificial Intelligence and Robotics and Digital Biology.

Read more: