My profile

How ChessGate's AI Works

A complete guide to the technology that makes each chess personality feel like the real player.

← Back to ChessGate

How ChessGate's Personalities Work

The oldest fascination in computer chess, a laboratory's way of measuring, a counting machine from 1890 and a modern engine. How fifty players from the 1830s to today came to play on ChessGate, each recreated from their own games.

A machine that plays like a human

In 1770 Wolfgang von Kempelen unveiled the Turk at the court in Vienna: a figure in Ottoman dress seated at a cabinet, moving the pieces with a mechanical arm. It toured Europe and America for decades and played, among many others, Napoleon and Benjamin Franklin. Its secret, a chess master hidden inside the cabinet, was kept for a remarkably long time. What drew the crowds was never that the machine was strong. It was that it seemed to think the way one of us thinks.

That fascination outlived the Turk by a long way. In 1948 Alan Turing wrote a chess program before there was a computer that could run it, and later played it out by hand, move by move, against a colleague. In 1997 Deep Blue beat the world champion. In 2017 a machine taught itself chess from the rules alone and overtook everything that had come before it. Today the strongest engines are far beyond any human, and they have made the old question sharper rather than softer. The machines play better than anyone. They also play like no one. Ask an engine to play like Capablanca and it does not know what you mean.

Style: easy to recognise, hard to define

Show a strong player ten moves of a game and they will often tell you whose game it is: a Tal game, a Petrosian game. Ask what exactly gave it away and the answer trails off. Something about the initiative, something about the pawns, a feeling. Style is the thing chess players recognise most readily and define least well.

Earlier attempts to put style into a machine worked from the outside in. Give an engine a famous player's opening repertoire and it plays the first dozen moves in character, then turns back into an engine. Turn a dial towards attack and you get an attacker, not a particular one. Add random mistakes to make it weaker and you get a player who exists nowhere. The question is genuinely hard, and nobody has closed it. We do not claim to have closed it either. We claim to have found a different way in, and we spent years on it.

A different way in: the laboratory

ChessGate is made by DCS Analytics, a company that writes software for laboratories. It occurred to us that we had met this problem before, in another trade.

In a laboratory, an instrument is never trusted on its own. It is calibrated against reference standards: samples whose values are known. You measure the standards, record how the instrument's readings relate to the known values, and from then on every reading of an unknown sample is interpreted through that calibration. The instrument supplies precision. The standards supply truth.

A modern chess engine is a near-perfect instrument: precise, tireless, impartial, and without any opinion about style. A grandmaster's games are reference material: thousands of positions in which the known value is the move the master actually played. Put the two together and you can do what a laboratory does every morning. Measure the instrument against the standard, position after position, and record how the two relate. That relationship is the style. Something nobody could define turned out to be something you could measure, not in any single position but across thousands of them, with the engine holding the ruler.

This is why ChessGate is, before it is anything else, an applied data analysis project. The chess knowledge comes from the engine and from the masters. The method for relating the two comes from the laboratory.

A card for every grandmaster

In 1890 the American census was counted on machines designed by Herman Hollerith. Every person became a card; every attribute, from age to occupation, became a hole punched in a fixed place. Sorting and counting machines then answered questions nobody had thought to ask when the cards were made, simply by reading the holes in new combinations. The count was finished in a fraction of the time the previous census had taken, and the company Hollerith founded became part of IBM.

We borrowed the idea, which is older than every computer and still one of the best ideas computing has. Chess has more positions than there are atoms in the universe, so asking whether Karpov ever saw this exact position is pointless; he almost never did. Instead we ask which characteristics the position has: the pawn structure, which pieces are active, where the kings stand, what each side is threatening, and the phase of the game. Every player on ChessGate has a card in Hollerith's spirit, with a mark for what they preferred at each combination of characteristics and phase. A position the player never saw still falls through holes that were punched by thousands of positions they did see.

Hollerith's machines answered questions nobody had asked when the cards were punched. Ours answer a question Karpov was never asked: what would you do here?

The formula

We are often asked what the formula is. The ingredients are no secret.

  1. The instrument. Our own engine reads the position and names the good moves.
  2. The calibration. A measurement, made for each player from their own games, of how the master's choices relate to the engine's.
  3. The card. What the player preferred at each combination of position characteristics and game phase.
  4. The plan. Masters follow through. The recreation remembers what it set out to do a few moves ago and prefers the move that carries it on.
  5. The margin and the floor. Style may cost a little, by a measured amount. It may never cost the game.

How the five are weighed against each other, and where the margins lie, is the part that took the years. It is also the part we keep.

A strong engine for the chess, a personality for the choice

Think of an actor who has studied a role for years. The script supplies the lines; the actor supplies the character. The engine supplies the good moves; the personality chooses, among them, the one this particular player would have played. The result is a move that is sound and, at the same time, recognisably theirs.

A simple example

Karpov to move. The engine offers several good moves:

Karpov's recreation plays c4. The move lies within the margin we allow, and it is the kind of move Karpov played in thousands of games: quiet, restricting, preparing the next step. Tal's recreation, in the same position, looks hard at Bxf7.

If Bxf7 were the only move that holds the position, every personality plays it. Character never overrules the board.

Strength with a safety net

The levels from club player to grandmaster are calibrated by measurement, not by guesswork. At every level the personality chooses only among moves close to the engine's best, and a final check makes sure the chosen move does not walk into a forced mate. Within that margin, the most characteristic move wins. A personality plays its style with conviction; it will not throw the game to do so.

A laboratory does not calibrate an instrument once and forget it. It runs control samples and watches for drift. We do the same with the personalities: we measure how often a recreation chooses as the real player did in positions from their own games, and recalibrate when it drifts.

What you see while you play

The Engine Output panel on the game page shows each decision in plain words: the phase of the game, the candidate moves, which of them were in the running, and the move that was chosen together with the plan it serves. When a move comes straight from the player's opening repertoire, the panel says so.

Three players, one method

The method is the same for every player. The evidence differs, and so does the character. Karpov and Tal run on identical code and play nothing alike.

Karpov, the boa constrictor

Quiet moves early, exchanges when they help, a taste for space and sound pawns. His plans squeeze: a weakness is found, fixed, and converted in the endgame.

Tal, the magician

Initiative above material. Checks, sacrifices and open lines in every phase, and plans that build, move by move, towards the opponent's king.

Fischer, the machine

Balanced across every aspect of the position and precise in all phases. Central control first, then piece activity, then the tactics that follow from both.

Judge for yourself

We do not claim that the question the Turk raised in 1770 is settled. We claim that a laboratory's habits, a counting machine's idea and a modern engine, put together with a great deal of patience, produce players you can recognise.

The surest test is your own. Play a master whose games you know, then play another at the same level, and ask yourself whether you could have told them apart with the names hidden. The Mystery GM on this site is exactly that test: a grandmaster who plays without a name.

What we keep to ourselves

How the calibration is computed, how the cards are read and weighed, and where the margins lie: that took years to develop and is the heart of ChessGate. It stays with us. What we will say is this: every move a personality plays traces back to evidence from the games it was built from, and the same evidence is what makes Petrosian defend, Nezhmetdinov attack and Capablanca simplify.

In short