AI·News & analysis
Why AI chess engines keep the board tense longer than grandmasters
A new study of 2,400 games finds that top engines hold on to complicated positions long after human grandmasters start simplifying.

Tide
Ripple
Sci-fi
5/10
Reality
Lab
Machines that enjoy a messy board more than we do.How we rate
Researchers at the Complexity Science Hub in Vienna and the University of Parma measured "strategic tension" in 1,200 grandmaster games and 1,200 Stockfish vs Leela Chess Zero games.
Top engines keep tension high for longer, while people simplify the board once they leave memorized opening lines. Tension rose with skill in humans and with search depth in Stockfish. The team says the idea may reach other fields, but it only shows a link, not a cause.
What to know
- Researchers compared 1,200 human grandmaster games with 1,200 Stockfish vs Leela Chess Zero games.
- They measured "strategic tension", a score for how many ways pieces attack, defend and control squares.
- Human tension peaks early, then drops. Engine tension stays high for longer and engines keep more pieces on the board.
- Higher tension went with higher skill in people and with deeper search in Stockfish, but the study does not show it causes the skill.
Researchers in Vienna and Parma measured how tense a chess board is, and found that top engines keep it tense for much longer than human grandmasters do. The engines seem to like the mess.
So what does "tense" mean here, and why would a computer enjoy it?
What is "strategic tension"?
The team needed a number for how complicated a position is. So they treated the board like a web.
Every piece and every empty square is a dot. A line joins two dots when they interact. The more lines, the more tension.
The study counts three kinds of lines:
- Attack: one piece could capture an opponent's piece on the next move.
- Defense: a piece could recapture if the piece it protects gets taken.
- Control: a piece watches an empty square where an opponent might want to go.
The score is built from the whole web, using a standard tool from network math called the spectral radius. You do not need the math. Think of it as a "how knotted is this?" meter.
In real life Picture a crowded kitchen at dinner time. Lots of people reaching for the same pan means high tension. An empty kitchen with one person making toast means almost none.
Which games did they look at?
The authors are Adamo Cerioli of the University of Parma, and Edward D. Lee and Vito D. P. Servedio of the Complexity Science Hub in Vienna.
They compared two sets of games, with 1,200 games in each:
- Human games: classical games by grandmasters, the top rank in chess.
- Engine games: Stockfish against Leela Chess Zero, two of the strongest chess programs, from the Top Chess Engine Championship between 2019 and 2024.
A chess engine is a program that plays chess. Those games had 120 minutes per player plus 12 seconds added after each move.
They also added human games from an online chess site, rated from 1,000 to 2,700, in rapid and blitz formats. Rapid means about 10 to 25 minutes per game. Blitz means about 3 to 5 minutes.
Finally, they had Stockfish play extra games at fixed "depths", up to 50. Depth is how many moves ahead the engine looks.
What did they find?
By the numbers: once each side had made about 20 moves, the average tension in human games fell below the tension in engine games, and it stayed lower.
Before that point, people lean on memorized opening lines. Those lines often lead to positions full of attacks and defenses.
Then the book runs out. Players reach positions they have not seen. Humans tend to trade pieces and simplify.
Engines do not do that. They keep more pieces on the board and keep more links alive. In the paper's words, the engines have a weaker tendency to simplify.
The paper also says human games reach a higher single peak of tension. So the story is not that engines are always more tense. It is that engines stay tense for longer.
Why do people simplify?
The senior author, Servedio, says the work gets closer than earlier studies to what these machines are doing. "Otherwise it looks like they are playing magic," he said.
For the human side, the likely reason is limited memory. Holding many threats and options in your head is hard.
Merim Bilalic, a psychologist at Northumbria University in England who was not involved, said the result makes sense. People need to reduce the mental load, but "engines don't really have that problem."
Why it matters: this fits an old idea in chess. A player who keeps options open is not forced to decide early. The engines can afford to wait.
How do modern engines pick a move?
Chess programs have beaten the best human since 1997, when a computer system called Deep Blue won. Researchers are still trying to explain why today's engines are so strong.
Early engines, Deep Blue among them, played out many branches of possible moves before choosing one. This is called tree search.
Many modern engines, such as Stockfish and Leela, also use trained neural networks to guide that search. A neural network is a program that learns patterns from examples. Brute force alone does not explain how the games unfold, and that is the gap this study tries to fill.
What does a game look like on the tension meter?
The paper describes a typical game in four stages:
- Opening: tension builds as pieces come into contact, in roughly the first 30 half-moves.
- Middlegame: tension stays near its peak.
- Simplification: pieces get traded and tension slowly falls, from about half-move 60.
- Endgame: few pieces remain, and tension settles low and steady.
A half-move is one turn by one player. The authors use it as their unit.
The authors read the different shapes this way. Engines build "distributed positional advantages that unfold gradually over multiple moves", while people favor attacking setups. Engine games also show more links and loops overall, which means a richer web of threats, defenses and controls.
To show the idea, the paper walks through a single game from an online blitz event, played in November 2024 between Magnus Carlsen and Alexey Sarana. A single game bounces around a lot, so the authors average over many games to see the pattern.
Does it make anyone stronger?
Tension grew with skill in several places:
- Human ratings: higher-rated players kept more tension through their games.
- Time limits: classical games showed more tension than rapid games, and rapid games more than blitz.
- Search depth: when Stockfish looked deeper, tension rose.
That looks like a clean story, but the authors are careful. Tension rises with ability, yet the study does not show that tension creates the ability, or that ability creates the tension.
It also found that tension reflects how a game goes, but does not predict it well. The paths of winning and drawn games split around the middle of the opening and middlegame, before tension peaks.
In engine games that ended in draws, many pieces stayed on the board, and tension did not separate draws from wins as sharply as it did for people.
What are the limits?
This is one team's measure applied to chess only.
- Engines: the work centers on Stockfish, with Leela Chess Zero included. Other engines might behave differently.
- Depth: the Stockfish tests stop at depth 50.
- Human data: the human games mix formats and players of different strengths.
- Cause: the study shows a link, not a cause.
Those are the authors' own cautions, and they are a good reason to treat the result as a first look.
Could this reach beyond chess?
The authors suggest trying the same measure on other games, such as Go, and on other kinds of programs.
Servedio also points to harder settings, like diplomacy. If AI systems keep tension high, they might find better outcomes. They might also keep a situation tangled.
"This is very dangerous because when you have high tension, systems are very unstable," he said. "So just a small error from one of the sides will cause a disaster."
The paper makes a similar point about international relations. It is a warning about a possible future, not a finding. Nobody tested it outside chess.
What it means for you
- If you play chess: keeping options open for longer is a habit of the strongest players, human or machine. Do not rush to trade.
- If you follow AI: engines do not just calculate faster. They also play a different style, one that avoids simplifying.
- If you read headlines about "AI in diplomacy": this is a chess study with a hint of a bigger idea. It is not evidence about real politics.
The bottom line
A new measure of chess tension shows that top engines hold on to complicated positions while people start to simplify. Higher tension goes with higher skill, but the study does not say it causes it. It is a neat peek at how machines play, and a reminder that chess is still a handy lab for studying minds.
Key facts
- Games studied
- 1,200 human grandmaster games and 1,200 engine games
- Engines
- Stockfish and Leela Chess Zero (2019 to 2024 tournament games)
- Researchers
- Complexity Science Hub Vienna and University of Parma
- Measure
- Strategic tension, from a network of piece interactions
- Stockfish depth tested
- Up to 50 moves ahead
Got questions?
Quick answers, plain wordsWhat is strategic tension in chess?
It is a score the researchers built to show how many ways the pieces on a board interact. Every attack, defense and square control adds links, and a board with many links has high tension.
Who did the study?
Adamo Cerioli of the University of Parma, Edward D. Lee and Vito D. P. Servedio of the Complexity Science Hub in Vienna. Servedio is the senior author.
Which chess engines were studied?
Stockfish and Leela Chess Zero. The engine games came from the Top Chess Engine Championship, 2019 to 2024, and Stockfish also played extra games at fixed search depths.
Do engines have higher tension than humans?
Not at every moment. Human games hit a higher peak early, when players follow memorized lines. After that, human tension falls below the engines, which keep it high for longer.
Why do humans simplify the board?
Researchers suggest people reduce the mental load. Merim Bilalic, a psychologist who was not part of the study, said the result makes sense because people need to cut the load and engines do not have that problem.
Does more tension make a player stronger?
The two go together, but the study does not prove one causes the other. Higher-rated people and deeper Stockfish searches both showed more tension.
Does the time limit change the tension?
Yes. Classical games, with the most thinking time, showed more tension than rapid games, and rapid games showed more than blitz.
Can this apply to things other than chess?
The authors say it may. They suggest other games and fields like diplomacy, but this is a suggestion, and the study only looked at chess.
Is the code available?
Yes. The authors say the Python code for the tension measure and the Stockfish simulations is on GitHub.
SourcesarXiv
Topics and tagschess, stockfish, leela chess zero, ai research
Go deeper
Related stories

Microsoft's AI Frontiers research group is over, says co-founder John Langford
John Langford says the roughly 50-person group ended after it was folded into Microsoft AI in September, and that its founding leaders all left.

OpenAI says its AI disproved a famous computer science rule of thumb
OpenAI and Anthropic each say an AI model found a faster way to crack a famous puzzle that experts thought had no shortcut. Outside experts are still checking.

OpenAI publishes 722 math papers written by an unreleased AI model
OpenAI put 722 manuscripts from an internal model on GitHub, with computer-checked proofs for many of them, after mathematicians asked AI labs to show their work.
More in brief
- Blue J raises $176 million to put AI agents to work on tax returns and IRS lettersNov 16
- FBI warns hackers are using the ARTEX AI tool to attack organizations worldwideNov 16
- Apple opens a 200,000-person study on how people use AI for healthNov 16
- Anthropic's IPO is reportedly oversubscribed as it aims to raise up to $90.4 billionNov 16
- Pentagon tests robot vehicles that hunt targets as a pack, with a human to approve fireNov 15
- Judge lets DRAM price-fixing lawsuit against Micron, Samsung and SK hynix go onNov 15