I Tested ChatGPT as a Chess Coach. Does It Actually Work?
Amateur chess has always had one problem: improving at a basic level is relatively easy. Solving tactics, playing lots of games, and knowing a few strategic principles is usually enough to make quick progress.
But there comes a point — perhaps around 1900–2000 Elo — where that is no longer enough. Serious tactical mistakes become less frequent, and increasingly difficult strategic concepts start to matter: pawn structures, good and bad pieces, initiative, compensation, prophylaxis, coordination… These are hard to discover on your own. You need to watch many games annotated by experts, analyze your own games… and having a coach also helps a lot. The problem is that coaches are very expensive for the average amateur.
Stockfish — and other engines — are incredibly strong as players, but… they do not talk to you. They tell you that a move loses 0.15 points… but they do not explain why. And many small inaccuracies of that kind, accumulated over the course of a game, can turn a comfortable position into a lost one without you really understanding when it happened. A coach needs the ability to communicate: both to talk to the student and to listen to them.
And this is where artificial intelligence starts to become interesting. Just a couple of years ago, ChatGPT — and other artificial intelligence tools, or AI — was unable to play a legal game of chess. It made grotesque mistakes constantly. But progress has been extremely fast.
So the question I ask myself is… can it act as a coach? Can it explain plans, mistakes, strategic ideas, suggest good practical moves, and identify the critical moments of a game in a useful way for a club player? And if that is not possible yet… how long will it take before it becomes a good coach for club players?
That is why I decided to run this experiment.
Table of Contents
Introduction
Prompt
First Game
Second Game
Conclusions
First of all, I want to make it clear that I will be talking about ChatGPT, but this applies to AI in general 😉
Today — as of May 12, 2026 — ChatGPT is already capable of playing chess at a decent level. The version tested is 5.5. As an example, here is a game I just played:
It seems clear that it plays better now, although it is still weak, especially because it almost never makes illegal moves. On every move, it stops to check whether the moves are legal. I tried making illegal moves, and it always warned me that I could not play them. Unfortunately, it still makes elementary tactical mistakes. First it left a bishop hanging, and later it lost its queen in one move. Most of the time it spent thinking was used to check whether my move was legal. The final position is key to understanding what still goes wrong. It thinks there are pawns controlling b3 and b5! There are no such pawns, and even if there were, those pawns would not control the king’s escape squares.
However, during the game it did make very accurate comments about what I was threatening, or what it wanted to achieve with its own move. So… could it give advice on positional themes, planning, and so on? Let’s find out 😉
My goal is to see whether ChatGPT 5.5 is capable of analyzing games and giving good advice, at least to beginners, and to write down the conclusions so that I can repeat the experiment in the future and see whether there has been any improvement. Could ChatGPT replace chess coaches in the coming years, at least as a way for amateurs to save the cost of a private coach?
I tested it with two of my own games, which I had already analyzed on the blog years ago. I basically told it to imagine that it was my coach and to analyze the games so that I could better understand what had happened in them. One of the games is more tactical, while the other is more positional. In principle, I expected it to make more mistakes in the tactical one. I ran the experiment in a temporary chat, so that ChatGPT would not “remember” that it was part of an experiment. Let’s see the result.
The prompt I used was the following, suggested by ChatGPT itself:
Act as my chess coach. I want you to analyze this game in a way that is useful for a club player.
Do not simply tell me whether a move is good or bad. I want to understand what happened in the game: the plans for both sides, the critical moments, the important mistakes, and the ideas I should learn in order to play similar positions better.
Please organize the analysis as follows:
- General summary of the game.
- Opening: whether the setup was reasonable and what plans each side had.
- Middlegame: critical moments, changes in evaluation, and important mistakes.
- Endgame or decisive phase: why the game was won/lost.
- Three practical lessons I should remember.
- The key points of the game.
If you see a clearly bad move, suggest an alternative and explain the idea, not just the variation.
You do not need to analyze every move. I prefer a clear, educational, and honest analysis. If a position is too tactical or you are not sure about a specific line, say so instead of making things up.
Here is the game in PGN:
[PASTE THE PGN HERE]
I thought it was a very good prompt, so I used it without changing anything.
This game was already analyzed by me on this blog: My Most Beautiful Game (sorry, only in spanish). I recommend reading it first, so you can better understand where ChatGPT got things right and where it went wrong.
Since the answer is very long, I am going to summarize what it says. My comments are in red. The analysis took about five minutes.
General summary of the game
The game was a very sharp Nimzo-Indian where Black won material early with ...Bxc3+ and ...Qxc3+, but in return left the queen far away from the king and began to seriously weaken the castled position with ...g5-g4.
White played coherently: he accepted having a somewhat worse structure/material balance, but obtained activity, the bishop pair, open files, and a direct attack against the black king. The game is decided because Black tries to defend with active tactics, but at the critical moment 22...Nd4? allows a sequence that trades queens and loses a piece.
The central idea: Black won pawns, but White won tempi and squares around the king.
So far, everything is correct.
After: 8.Bd3 Bxc3+ 9.bxc3 Qxc3+ 10.Kf1 Black wins the c3-pawn, but also accepts several risks:
- the black queen is advanced and may lose tempi;
- the black king has already castled and can be attacked;
- White keeps the bishop pair;
- Kf1 is not pretty, but the white king is not as exposed as it may seem.
The first important strategic moment is:
10...g5?!
A healthier alternative for Black would have been to play more soberly, for example with ideas like ...Nc6, ...d6, or ...d5, finishing development before launching pawns in front of the king.
Still correct — it is surprising me! However, here I miss a comment about the sacrifice Nxg5.
Very good idea by White:
12.Bd6!
This bishop on d6 is a very annoying piece. It controls f8, hits e7, makes Black’s coordination difficult, and will be key in the final mating net. It is not just an “active” move: it is a piece installed on a square that greatly restricts the opponent.
This is true. Bd6 is not objectively the best move, but it makes coordination difficult, is very annoying, and limits the opponent. It even controls f8, preventing the king from escaping.
The critical sequence begins with:
13.h4! g4 14.Ng5
White understands the position well. If Black has advanced the g-pawn, it must be attacked as a hook. h4 is not only trying to exchange a pawn: it aims to open lines against the black king.
I could not have said it better 😮
After:
14...cxd4 15.Rc1 Qa5 16.Ne4
White combines attack with gaining tempi against the black queen. Rc1 forces the queen to retreat, and Ne4 centralizes the knight, increasing the pressure on f6, g5, and the black king.
Then comes a very instructive tactical sequence:
16...Nxe4 17.Qxg4+ Ng5 18.hxg5 Qxg5
Black manages to eliminate the white knight, but in return White opens the h-file and removes an important defender. The practical point is that, although Black is not doing badly materially, the king begins to run out of defenders.
Here its comments start to become a bit weaker, although it is true that Black is not doing badly materially — being pawns up — and that, in exchange, the king has few defenders left.
Very good white maneuver:
21.Rh3!
This is a strong-player attacking move. It does not look for an obvious immediate threat, but prepares Rh5, attacking the queen and increasing the pressure on the king.
This is half true. Rh3 is a good move, but its idea is not to play Rh5 — that makes no sense… the idea is Rg3.
The game continued:
22...Nd4?
Black’s idea is tactical: to attack the white queen on f3. If White automatically plays 23.Rxg5?, then 23...Nxf3 and Black compensates by capturing the white queen.
But White finds the correct move:
23.Qh3!
This is probably the most important move of the game.
Here it gets the tactical theme right, and it is true that Qh3 is the only winning move, but there were not many options to choose from anyway.
After: 25.exd4 Rh8 26.Ke2 a5 27.Rf1+ Ke8 28.Rxh6, Black’s position collapses.
The move 28.Rxh6 does not only win a pawn. It also creates mating threats. A very nice detail: if Black plays 28...Rxh6?, then 29.Rf8#.
This comment surprised me: 100% correct.
Three practical lessons you should remember
- If you win material in the opening with your queen, your priority should be consolidation.
Black won pawns, but kept playing ambitiously with ...g5, ...g4, ...f5. When you are ahead in material, the correct thing is often to give back some initiative, finish development, and avoid opening your own king. - An advanced pawn in front of the king is a hook.
After ...g5, White responded very well with h4. This is a universal idea: when your opponent advances pawns in front of their castled king, you do not always need to attack with pieces first; often it is enough to push the flank pawn and open lines. - It is not enough to create a tactical threat; you have to check whether your opponent can escape with tempo.
22...Nd4? looks active because it attacks the white queen, but it fails to 23.Qh3!. The white queen moves away and keeps the threat against the black queen. This pattern is very important: when you play an intermediate move, check whether your opponent has a move that defends and attacks at the same time.
100% agreed. Maybe they are not the three best lessons, but all three are good lessons.
The key points of the game
The strategic key was that White turned a materially dubious position into an attacking position thanks to activity: active bishops, the h-file, the f-file, and the rook lift Rh3-h5.
The tactical key was 23.Qh3!, punishing 22...Nd4?
The final key was the dominance of the bishop on d6, especially over f8. That bishop, placed there from the opening, ended up being decisive in the mating net.
Overall, it was a very instructive game: White did not win because of one isolated combination, but because all his decisions moved in the same direction: opening lines, activating pieces, and punishing the weaknesses created by Black’s kingside pawns.
It repeats the mistake of saying Rh3-Rh5, but the rest is good.
This game was already analyzed by me on this blog: Analyzing My First Two Games (again, only in spanish). I recommend reading it first, so you can better understand where ChatGPT got things right and where it went wrong.
The analysis took about three minutes.
At first, both sides played reasonably, but the game very quickly turned into a position with no minor pieces, only queens and rooks. That changed the nature of the struggle a lot. What mattered was no longer so much the squares for knights or bishops, but:
who activated the queen better;
who controlled the open files;
who could open a second front;
who calculated the threats against the king better.
Black won two pawns on the kingside with 36...Qxh4 and 37...Qxf4, but did so at the cost of allowing the white queen to enter on a7 and then allowing the b5 break. In the end, White coordinated queen + rook on the seventh rank very well, and Black did not react in time to the mating danger.
The game was not decided by a single isolated mistake early on, but by an accumulation: Black won material, but lost control of the activity.
Everything is correct here: major-piece endgame strategy, and a good explanation of what happened in that ending.
White chose an ambitious plan: gaining space with e5 and building a strong center with d4-e5. This is a reasonable setup. White’s idea is to play with more space, prepare natural development, and later look for breaks such as f4-f5 or queenside expansion with b4-b5.
Black’s plan was typical against an advanced white center: exchange pieces, put pressure on d4, and avoid becoming too passive. The decision 8...Bxc3+ is coherent: Black eliminates a knight that could support d5/e4 and simplifies.
15...Nxe3 16.Qxe3 Bxf3 17.Bxf3
More minor pieces are exchanged. This partly favors Black, because when you have less space, exchanging pieces usually eases your position. White is not worse, but loses dynamic potential: without knights or bishops, it will be harder to attack Black’s king directly. In return, White keeps space and will later be able to play with rooks and queen.
Everything is perfect up to this point. Strategy in a Caro-Kann type structure, and strategy when you have less space.
One white alternative to consider at that moment was not rushing h3, and instead playing something like 15.Qd2 or 15.Rfd1, keeping the tension. The idea is not an immediate tactic, but to avoid allowing Black to force so many comfortable exchanges.
Here it gets a bit confused. Qd2 and Rfd1 are not good moves. h3 is a good move, and something like Bd3 is also possible. In reality, I cannot really do anything to avoid the piece exchanges.
White’s move 21.f4 is thematic. It reinforces e5 and, in some positions, threatens to advance f5. Black later responds with 24...f5. This move is very important. It blocks the kingside and prevents White’s f5, but it also has a strategic drawback: by closing the position further, Black reduces his own breaks. In structures of this type, Black often wants to play ...f6, not ...f5, in order to attack the base of White’s center. A more dynamic alternative for Black was 24...f6! The idea would be to challenge the e5-pawn directly.
Very good comment. f5 is good and bad at the same time.
Then comes a critical phase:
31.Rg5 g6 32.h4 h6 33.Rg3 Kh7 34.Rf3 h5 35.Re3 Qe7 36.Qc2
Here Black plays:
36...Qxh4 37.Qc5 Qxf4
Black wins two pawns, but this is a double-edged decision. Materially, it looks very attractive, but the black queen ends up far away from the queenside defense, and the white queen enters with force.
Here it is wrong. 32.h4 was bad, 36.Qc2 was also bad, and taking the pawns is good for Black. Although it is true that White has compensation because the queen becomes active on the other flank.
42.b5!
This is probably one of White’s best moves in the game. Black was playing on the kingside, so White opens the queenside. In closed positions, when the opponent becomes too optimistic on one wing, the correct response is often to open the other side.
Here I think it exaggerates by saying it is one of the best moves of the game, but it is true that this is the move White has to play.
The serious mistake comes with:
48...Rh6?! 49.Rc7 Qe6?
The move 49...Qe6? is the decisive error. The black queen abandons the g8-square and allows the final tactical motif. A much tougher defense was to try to exchange queens with something like:
49...Qg6!
This is already wrong. The losing move is 45...Qg6, allowing Qf8. What is true is that Qe6 speeds up the defeat, because it is already mate in 3.
Three practical lessons you should remember
- In closed structures, activity is worth more than a couple of pawns. Black won h4 and f4, but meanwhile allowed Qxa7, b5, and Rb3-b7/c7. In positions with somewhat exposed kings, winning pawns by moving the queen away from the defense can be dangerous.
- When your opponent attacks on one flank, look to open the other. White’s move 42.b5! is very instructive. Black was advancing pawns on the kingside: ...g5-g4, ...h5-h4. If White only defends, he may end up getting crushed. But by playing b5, White forces Black to deal with another front. That break opened the way for the white rook to reach b7/c7, which ended up being decisive.
- Queen + rook on the seventh rank is an extremely dangerous combination. The final mate does not appear out of nowhere. It is built like this:
I mostly agree with these three lessons.
The key points of the game
The game can be understood through five ideas:
- Black simplified the opening well, exchanging minor pieces and reducing White’s space advantage.
- 24...f5 closed the position too much. It was playable, but ...f6 was probably a more active way to fight against White’s center.
- White defended well against the attack with Rg3, exchanging an attacking rook instead of remaining passive.
- Black was too materialistic with 36...Qxh4 and 37...Qxf4. He won pawns, but allowed the white queen to enter via a7 and the queenside to open.
- 42.b5! was the key strategic break. Without that move, White would probably just have been suffering while two pawns down. With it, the white rook reached the seventh rank and the attack changed hands.
I do not think Black lost because of materialism, but because he stopped paying attention to White’s threats and allowed White to enter through the back door. The rest looks correct to me.
- At the moment, ChatGPT is not very reliable tactically, although it can still make simple observations about threats and small combinations.
- However, it does make good conceptual observations. It talks about initiative, king weaknesses, compensation, annoying pieces, attack, and so on.
- If Stockfish evaluations were added to the PGN, its observations would probably be better, since it would be able to explain the game from the “truth” of the position.
- Many comments are not completely precise for the specific position, but they still contain useful principles for club players.
- Given ChatGPT’s rapid progress, my conclusion is that within a few years, any amateur will be able to have a personal chess coach in their pocket. Soon, it should be able to help players around 2000 level. For professionals, it will still take a few more years — or at least a good combination of Stockfish and ChatGPT.
- The future will probably be one of two things: either ChatGPT calculates better thanks to information obtained from Stockfish, or Stockfish learns to talk through the incorporation of AI.
Coming soon...
The next part of the experiment would be to show it Stockfish’s evaluation and see how it explains that evaluation — or the changes in the evaluation.
Because the main current problem does not seem to be concepts, but calculation and tactical precision.
It is clear that the future is AI-assisted game analysis. How long will it take websites like Chess.com to offer this service? Right now, Chess.com’s analysis tool makes rather strange comments that do not really help you understand your games, such as “the best option was to kick the knight away”, “the best option was to take the knight”, or “it was the only good move.” But it does not explain why.
This is already starting to happen. In its May update, @MattyDPerrine explained that Matty, its AI chess coach, is becoming more specific: fewer generic comments, fewer board-visualization mistakes, and more explanations based on actual variations. That seems like the right direction. AI should not replace engines like Stockfish, but use their tactical accuracy to explain positions in a more useful, human way.
https://www.chess.com/blog/MattyDPerrine/chessalyz-may-update-matty-got-specific
Blog translated from my blog, in spanish, with ChatGPT:
https://www.chess.com/es/blog/damafe/probe-chatgpt-como-entrenador-de-ajedrez-funciona
