Researchers who analyzed online chess games say strong engines keep the board in a state of high strategic tension deep into a game, while human players tend to simplify their positions as the game goes on. Scientific American reported on the findings, which appeared in a paper in APS Open Science.
The result may point to one reason chess engines play the way they do, though the study found a correlation and did not show that tension causes strong play.
Computers have beaten the best human players at chess since 1997, when a system called Deep Blue won, and engines have kept improving. Researchers are still trying to understand where that strategic edge comes from. Many engines can evaluate a huge number of moves by brute force, but how that shapes their style of play has been unclear.
Early chess programs, Deep Blue among them, played out many branches of a decision tree before choosing a move, a technique called tree search. Many modern engines, including Stockfish and Leela, use trained neural networks to guide that search.
The team collected human versus human games and Stockfish versus Leela games from online databases. At each point in a game, they measured what they call strategic tension: the number of mutual interactions possible between opposing pieces. Tension is often higher when more moves are available.
According to Scientific American, the average tension in human games was higher than in the engine games until each player had made 20 moves. After that point, human tension dropped below the engine level, while the engines' tension stayed high.
The researchers suggest an explanation for the early stretch: people lean on memorized tactics and hope that complex interactions between pieces will make an opponent slip. Once the position becomes unfamiliar, players may simplify their arrangement of pieces because they cannot hold complicated dynamics and several options in their heads.
Merim Bilalić, a psychologist at Northumbria University in England who was not involved in the research, told the magazine the result makes sense. People need to reduce cognitive load, he said, and engines do not really have that problem.
The data showed several other patterns. Higher-rated human players maintained greater tension throughout their games. Tension was also greater in games with more time on the clock: highest in classical chess, lower in rapid chess and lowest in blitz. When Stockfish searched deeper trees of possible play, tension increased.
Higher tension went along with greater ability in both humans and engines. The study did not show, however, that tension led to that ability or came from it.
Vito D. P. Servedio, the study's senior author and a physicist at the Complexity Science Hub in Vienna, said his team's paper is probably the closest yet to understanding what these machines are doing. Otherwise, he said, it looks as if they are playing magic.
Chess remains a useful testing ground for studying thinking in people and machines, and this work adds a measurable trait that separates the two. For most readers the result will not change how they play or how they use software. The practical significance, as the researchers see it, lies in where AI decision-making is heading.
Servedio said the findings may apply to higher-stakes situations such as diplomacy, where AI might reach better outcomes but at the risk of complicating matters. He warned that high tension makes systems very unstable, so a small error from one side could cause a disaster. That is his view of what the results could imply. The study itself looked only at chess, and it did not test AI in diplomacy or any other setting.
The open question is whether the link between tension and strength holds up in other areas, and whether it is a cause of strong play or merely a sign of it. The study does not settle either point.
Around Prague, based on reporting by Scientific American