Nearly three decades after Deep Blue stunned the chess world, researchers are still trying to answer a deceptively simple question: what exactly makes AI so good at chess? A recent paper in APS Open Science, highlighted by Scientific American on November 15, offers a measurable clue — and it comes down to a property the authors call "strategic tension."
Chess engines have long been able to out-compute humans, brute-forcing huge numbers of candidate moves through tree search. Modern neural-network engines such as Stockfish and Leela Layers use trained networks to guide that search, and their play has only grown more formidable. What the new study adds is a statistical portrait of how their play differs from ours — not just in who wins, but in what the board itself looks like as the game unfolds. For readers following AI research news, the work is a reminder that even "solved-feeling" domains like chess still hold open questions about how machines actually think.
Measuring Tension on the Board
The study was led by physicists at the Complexity Science Hub in Vienna, with senior author Vito D. P. Servedio. The team collected both human-versus-human games and Stockfish-versus-Leela games from online databases, then measured a quantity they call strategic tension at every position: essentially, the number of mutual interactions possible between opposing pieces. High-tension positions are those where many pieces square off against each other — rich with latent captures, threats and counter-threats that could be unleashed by a single move.
Tension is a natural quantity for physicists to reach for. It captures, in a single number, how much complexity is latent in a position — how much the game state would change if either player pushed the board over the edge into a cascade of exchanges.
Humans Simplify; Engines Keep the Board on Edge
The central finding is a divergence in how tension evolves. In human games, average tension was actually higher than in AI games during the first 20 moves — the phase where people lean on memorized openings and well-drilled tactics, hoping the complexity will eventually force an opponent to slip.
But after that point, the lines cross. Human tension drops below that of AI games, which remains high deep into the middle game. The researchers' interpretation: as positions become unfamiliar, human players simplify, trading down pieces and steering toward calmer structures because they cannot hold the full complexity of many interacting pieces and multiple options in mind at once. The engines feel no such constraint. They sustain positions dense with mutual threats that most human players would find intolerable to navigate — and precisely those positions, the study suggests, are where human accuracy degrades.
Skill, Time Controls and Search Depth
The data also showed that tension tracks ability and deliberation in both humans and machines. Higher-rated human players maintained greater tension throughout their games than lower-rated ones. Games with more thinking time showed more tension — highest in classical chess, lower in rapid, and lowest in blitz — consistent with the idea that sustaining complexity is expensive in cognitive effort. And on the machine side, the team observed that when Stockfish searched deeper trees of possible continuations, tension increased.
The authors are careful about causality. Higher tension correlated with greater ability in both humans and AI, but the study did not show that tension produces that ability — or that it flows from it. It is a fingerprint of strong play, not yet a demonstrated mechanism.
"It Looks Like They Are Playing Magic"
The researchers' ambition goes beyond characterizing chess aesthetics. "Probably our paper is the closest to understanding what these machines are doing," Servedio told Scientific American. "Otherwise it looks like they are playing magic."
That framing — treating engine play as a physical system to be measured rather than a black box to be admired — is what distinguishes the work from earlier analyses of machine chess. Rather than asking whether an engine's move is best, the team asks what the aggregate statistics of engine play reveal about the kind of search and evaluation the machine is performing.
A Warning Beyond the Chessboard
Servedio also drew an unsettling implication for domains far from 64 squares. If AI systems reach better outcomes in complex, adversarial settings by sustaining high tension — keeping many volatile interactions alive simultaneously — then deploying them in something like diplomacy could mean machine-optimized strategies that humans find unmanageably fragile. "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 analogy is speculative, and the authors do not claim to have measured diplomatic dynamics. But it points at a genuine tension (of the organizational kind) in AI deployment: the same property that lets an engine outplay a grandmaster — comfort with sustained complexity — may be the property that makes its strategies risky to adopt in human institutions that cannot reliably maintain them.
What the Study Does Not Show
The limitations are worth stating plainly. The study establishes correlation between tension and playing strength, not a causal mechanism connecting the two. It compares a particular pair of engines against a particular population of human games from online databases, and chess — with its discrete moves and perfect information — remains a best-case setting for this kind of analysis. Extending "strategic tension" to messier domains is a hypothesis, not a result.
Still, the paper adds something the chess-AI literature has lacked: a simple, computable statistic that cleanly separates how humans and machines manage complexity over the course of a game. Chess has long served as a model system for studying cognition in both humans and machines, and Deep Blue's 1997 victory over Garry Kasparov marked the moment machines first proved they could out-think us at our own game of calculation. Nearly 30 years on, the question has shifted from whether machines can win to why they win — and, increasingly, to what their way of winning might teach us about the systems we are now handing real-world decisions to.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
