What would happen if two chess AIs faced each other?


If two chess AIs faced each other, what would happen? What if they always played the most optimal move and never made a mistake?

As of today, according to the ChessGames.com database of over 800,000 games, White wins 37.67% of games compared to 27.92% for Black. 34% of games end in a draw. In other words, White’s winning percentage is 54.67% compared to Black’s 44.92%. These odds are comparable for both human players and computer players. It would appear that due to this slight advantage, a game between two perfect AI players would result in White winning more often than Black. However, another thing to consider are drawn games.

As noted above, games in general end in a draw about 34% of the time, but this rate skyrockets the more skilled the players. According to Chess News, chess games end in a draw around 20% of the time when the highest FIDE rating between the two players is 1500 and increases to nearly 60% for a FIDE rating of 2800, which represents the highest rated players in the world. When considering all World Chess Championships since 2008, games have resulted in a draw 69.7% of the time over 76 games. There are a few theories for this, but the main theory is that White’s advantage is not enough to alone allow White to gain a winning position, and since Black is also a very skilled opponent who makes few mistakes, the position remains neutral for both players until they agree to a draw.

However, I don’t believe this will apply to a perfect computer AI because of how well they can calculate future moves. In many cases, even the best players are unable to tell that a position allows a player to force a win rather than a draw because they require calculating hundreds of moves in the future, compared to the average of around 50 moves per game. In an interview, Garry Kasparov, which many consider to be the greatest player of all time, commented on a game like this:

It said mate in 490 moves, first mate. Now, I can tell you that — even being a very decent player — for the first 400 moves, I could hardly understand why these pieces moved around like a dance.

The problem being faced is that the sheer number of possible positions in chess may be impossible to calculate. In 2007, Jonathan Schaeffer and his colleagues at the University of Alberta published a proof that solved checkers. After 16 years of computation on as many as 200 consumer desktop computers, all possible endgames involving fewer than 10 pieces were evaluated into a database of 39 trillion positions. For chess however, it is currently estimated that there are more unique chess positions than there are atoms in the observable universe, making it impossible to calculate with today’s technology. Only endgames with no more than 7 pieces (including the two kings) have been solved, and it is not expected than 8 pieces will be solved any time soon, let alone all 32 pieces.

So to answer the question asked in the title, an AI that can perfectly play can not exist with current technology because it cannot calculate the sheer number of possibilities in chess. However, the best AI today perform similarly to the best players in the world, forcing a draw on their opponent if a winning solution cannot be found.