Two tools can look at the same losing move. One is an engine, which returns an evaluation, a line, and a depth. The other is an AI coach, which returns a sentence about what the player missed and a rule for next time. Both are right about the move. Only one of them changes how the player plays on Thursday.
Below are two common club blunders: the queen that grabs a pawn on b2 and the bishop that lands on h3. Run through both tools.
What an engine actually says
Stockfish and its relatives output three things: an evaluation, a principal variation, and a depth. The evaluation is a single number. The principal variation is the sequence of moves the engine expects if both sides play perfectly from here. The depth is how many half-moves ahead it looked.
All three are accurate to a degree no human can match. All three are close to useless as teaching on their own.
Consider a common club blunder: the queen grabs a pawn on b2 early in the opening. The engine’s response is a line beginning with a rook to b1, an evaluation somewhere between a pawn and a piece in White’s favour depending on the exact position, and a variation that ends with the queen lost or badly misplaced eight moves later.
A player who takes that pawn once will take it again, because the engine has explained nothing. It has only measured the damage.
What a coach adds
A coach, human or AI, starts from the same position and answers a different question. Not how bad is this, but what did the player fail to see, and what rule would have caught it.
On the b2 pawn, the coach says something like: the queen has gone to the one square where it can be attacked with gain of time by a rook and a bishop, and the squares it needs for retreat are all covered, so the pawn was poisoned. Then comes the reusable part. Before taking a pawn with the queen, count the retreat squares. Fewer than two, leave it.
The second thing a coach adds is a name. Overloaded defender, back rank weakness, removing the guard, the pin that was not really a pin. Names matter because memory works on patterns, not on numbers. A player who has heard the phrase overloaded defender three times will spot the fourth one unprompted. A player who has seen minus 3.4 three times has learned nothing except that mistakes are expensive.
A worked blunder
Here is a position that turns up in thousands of games a day around the 1,200 level.
White has castled kingside and pushed the h pawn one square to h3. Black has a bishop on c8, a queen on d8, and a clear diagonal from c8 to h3. Black takes the h3 pawn with the bishop, White recaptures with the g pawn, and Black brings the queen to d7, intending to follow it to h3. It has given a bishop for one pawn and a damaged enemy king position.
The engine’s opinion is roughly minus three, with a line in which White meets the queen’s arrival with a knight move and keeps everything. That is the engine’s entire contribution.
An AI coach reads the same position and talks through it the way a club coach would. Black wanted an attack on the king, which is a reasonable thing to want. But an attack needs several pieces near the enemy king before a sacrifice makes sense, and Black has exactly one, the queen, with the knight still on g8 and both rooks at home.
The tell was on f3. The knight there reaches g1 or h2 in one move, which means the defence arrives before the attack does, and with no second attacker coming, the sacrifice ends on the move it started.
What the coach leaves behind is a habit, not a line. Before any piece goes into the fire near the enemy king, count the attackers within two squares of that king, then count the defenders. In this game, the count was one against three. Give up material when the attackers win the count by two or more, and not before.
That is a paragraph rather than an eleven-move variation, and it is the paragraph the player will still have next week.
| Engine | AI coach | Human coach | |
|---|---|---|---|
| Verdict on the move | Exact, in centipawns | Approximate, in words | Approximate, in words |
| Why the move failed | Not given | Named pattern plus the missed idea | Named pattern plus the missed idea |
| Rule for next time | Not given | Usually given | Given, tailored to the student |
| Availability | Instant, free | Instant, low cost | Weekly, expensive |
| Blind spots | None in calculation | Can misstate a line if not engine-backed | Tires, misses tactics |
Photo: “Chess game Staunton No. 6” by Wilfredor, Wikimedia Commons (CC0)
Where AI coaches still fall short
The weakness of a language model on its own is that it will describe a variation with total confidence and get a move wrong. That is unacceptable in chess, where one wrong move in a line changes the verdict.
The tools worth using run an engine underneath and hand the language model the engine’s lines to explain, so the calculation stays exact and only the words are generated. This split is the same pattern that works everywhere models meet high-consequence output: sound AI agent architecture keeps a deterministic component in charge of anything that must be correct and lets the model handle interpretation.
Ask a coach that works this way for the refutation of a move, and the line is real. Ask one that works from language alone, and the line may be invented. Chess happens to be an unusually clean case of a broader rule about which tasks generative models suit, because every claim can be checked against a calculation.
The other limit is context. A human coach who has watched a student for a year knows that this particular player always overextends on the kingside and shapes every lesson around it. An AI tool sees one game at a time unless the platform keeps a history. Some do, and a system holding a player’s last fifty games can say the useful sentence: this is the fourth time this month the queen has gone hunting pawns on move nine.
Why suggestion strength decides what you learn
Here is the part most players get wrong when they set up an analysis tool.
Maximum engine strength feels like the obvious choice. It is the least useful one for improvement. A 3200-level suggestion is frequently a move a club player could not have found, would not have understood, and cannot generalise from. It has the same teaching problem as the raw evaluation number: technically perfect, pedagogically empty.
A suggestion calibrated near your own rating behaves differently. It shows a move sitting just past the edge of what you already see, which is close enough to your existing pattern library that you can absorb it. That is the same principle behind the rule in the worked example above. Counting the attackers is a rule a 1,200 player can apply. Finding the engine’s eleventh move is not.
This is why adjustable strength matters more than raw engine power. ChessHelper.ai runs on Chess.com and Lichess with strength settable anywhere from 800 to 3200, plus a dashboard that stores game history and move-by-move analytics, which is the material a review session actually needs.
One boundary belongs alongside that, and the company states it in its own FAQ: external assistance during rated games breaches the terms of service of most chess platforms. Analysis tools belong in post-game review, study positions, and unrated practice. Using one during a rated game is not a grey area, and a rating built that way measures the software rather than the player.
Using one well
The right moment for an AI coach is the review after the game, because that is where an explanation has time to land and a rule has a chance of sticking.
Three blunders per game is plenty. Each one gets read for the pattern and the rule rather than the number. The rule goes into a notebook as a single line, which after a month becomes an honest list of the mistakes one particular player actually makes.
Set the strength near your own level while reviewing rather than at maximum. You want the move you could have found, not the move a supercomputer found.
Treat confident-sounding prose about a variation with suspicion unless an engine produced the line. Fluency is not accuracy, and that gap is the single most misread thing about current AI tools across every field, not just chess.
The bottom line
The engine number still has a place. It is the ground truth, and any coach’s explanation should agree with it. But the number is the end of the engine’s job and the beginning of the coach’s.
A player who finishes a review able to say what went wrong in one sentence, and what to check for next time in another, has learned something. A player who finishes with minus 3.4 has only been graded.
